Author: Pete

You are on a job site, in a client meeting, or deep in a spreadsheet when the phone rings. Again. By Friday you have a string of missed calls, a few voicemails you dread opening, and that quiet worry that a lead went somewhere else.

Here is the short answer. An AI voice agent for small business is a phone assistant that answers (or places) calls, follows a script you wrote, captures what the caller said, and hands the result to you, your calendar, or your CRM. You do not need a developer. You pick one repetitive call type, test it on a free trial, and keep humans on everything that needs judgment.

If you want the platform build guide next, we already wrote a full walkthrough of how to pick an AI voice agent platform and build your first call agent. This page is the decision layer for owners and operators: is it worth it at your size, what does it cost, which calls fit, and how to start without breaking trust with customers.

What an AI voice agent for small business actually is

Strip the demos and you get four parts working together:

• Ears: speech-to-text turns the caller’s words into text.
• Brain: a language model decides what to say next from your instructions and knowledge.
• Mouth: text-to-speech (or a speech-to-speech model) answers in a natural voice.
• Phone line: a real number that can ring, transfer, or hang up like a person would.

That stack is what vendors sell as a “voice agent” or “AI phone agent.” For a small team it usually replaces the worst parts of reception: after-hours intake, “are you open?”, booking a slot, or taking a short intake before a human calls back. It is not a full employee, and it is not a chatbot on your website. It lives on the phone.

If you already automate desk work with text agents, the idea will feel familiar. The difference is latency and manners. Callers interrupt. They change topics mid-sentence. A good platform handles turn-taking so the agent does not talk over people. A bad one sounds like a hold menu with better grammar.

When a small business should (and should not) use one

Use a voice agent when three things are true at once:

1. You miss the same kind of call every week (intake, booking, status check, after-hours).
2. The answers are knowable from a short script plus a calendar or FAQ, not from deep diagnosis.
3. A warm transfer to a human is available when the caller needs a person.

Skip it (for now) when the call is the product. Therapy, complex sales discovery, legal advice, medical triage, and any conversation where tone and trust are the whole job still need a human on the line. An agent can greet and route. It should not pretend to be your senior advisor.

Good first jobs for a five to twenty person shop:

• After-hours: capture name, need, and best callback time.
• Appointment booking for a single service with clear slots (see our guide to an AI agent for appointment booking automation if the booking logic already lives in n8n or Make).
• Lead intake: ask three qualifying questions, then notify the owner (pair this with a desk workflow from AI agent for lead qualification).
• Simple FAQ: hours, location, pricing bands you are happy to publish out loud.

Bad first jobs: “handle all customer support,” “replace our receptionist tomorrow,” or “cold-call a bought list.” Those fail for process reasons, trust reasons, or legal reasons. Start narrow.

What it costs for a small business (real numbers)

Most platforms bill by the minute. Retell AI publishes pay-as-you-go voice agent pricing at $0.07 to $0.31 per minute, depending on the model and voice you choose, with no separate platform fee on that plan, and 20 concurrent calls included before you buy more capacity (Retell AI pricing).

That headline range is not one flat sticker. Retell itemizes the minute as voice infrastructure plus LLM plus text-to-speech plus telephony. Their own worked example lands around $0.125 per minute with GPT 5.1, a platform voice, and a US number, before add-ons (Retell on usage-based pricing). Their cost calculator guide walks the same formula so you can price the exact model you will run (how to calculate monthly voice AI cost).

Translate that into a small-business month. Say you route 500 minutes of missed and after-hours calls to an agent:

• At about $0.12 per minute, usage is roughly $60.
• At about $0.20 per minute (heavier model or premium voice), usage is roughly $100.
• Add a phone number (often a few dollars a month) and any CRM connection time you pay a contractor for once.

Compare that to the cost of a missed booking or a part-time receptionist hour. For many service businesses, covering the boring calls at under $150 a month is cheaper than hiring for the gap, and cheaper than losing the lead. It is not “free AI.” It is a metered tool. Track minutes in week one so the bill never surprises you.

Other platforms (Vapi, Bland, ElevenLabs Agents, and similar) use the same idea with different bundling: some quote a hosting fee plus pass-through model costs, some bundle speech and LLM into one rate. Always read whether telephony is included. The minute you hear on a sales call is rarely the full bill.

Best AI voice agent options for a small business (how to choose)

You are not picking “the best AI in the world.” You are picking the lightest tool that can do one call type well.

Score each option on six practical checks:

1. Browser demo in under an hour. If you need a sales call to try it, skip it for now.
2. Clear per-minute math. You should be able to estimate a 500-minute month before you buy.
3. Warm transfer. The agent must hand a live caller to a human without hanging up.
4. Calendar or webhook. Booking and CRM updates should not require a custom app.
5. Transcript export. You need the text after every call for quality and follow-up.
6. Your language and accents. Test with a real customer voice, not only the demo voice.

For a first agent, browser-first platforms with a free credit or trial (Retell, Vapi, ElevenLabs Agents, and peers) beat custom builds. Building from raw APIs only makes sense when voice is your product, not when voice is your phone line. If you want a wider agent-builder comparison for desk work (not only phone), see our no code AI agent builder guide and the plain walkthrough on how to build an AI agent without code.

The three call types that work for small teams

1. After-hours intake

Script: greet, say you are the after-hours assistant, ask name, reason for calling, urgency, and best callback window. Confirm you will text or email the team. Offer to transfer if someone is on call.

Why it works: callers get acknowledged. You get structured notes instead of a mumbled voicemail.

2. Booking a simple appointment

Script: confirm the service, offer two time windows, book into your calendar tool, send a confirmation SMS or email if the platform supports it.

Why it works: the decision tree is short. Failures are obvious in the transcript.

3. Lead qualification before a callback

Script: ask company size, timeline, and budget band (only if you already ask those on forms), then create a task for sales.

Why it works: your morning starts with ranked callbacks, not a pile of “call me back” messages. For chat and desk versions of the same idea, our automate customer support with AI agents guide covers the text side.

Write the script like you train a new hire on day one. Short sentences. One question at a time. A clear exit: “I can transfer you to a person, or take a message.”

You are the after-hours phone assistant for Bright Lane Plumbing.
Greet the caller, say you are an AI assistant, and offer to take a message or book a callback.
Ask one question at a time: name, address or area, what broke, is water still running, best callback number and time.
If the issue is an active leak or gas smell, say you will escalate and offer a warm transfer to the on-call number.
Never invent prices. Never claim a technician is on the way unless the calendar confirms a booking.
End by repeating the callback window and thanking them.

How to set one up without coding (this week)

1. Pick one call type from the list above. Write the happy path on one page.
2. Create an account on a platform that lets you test in the browser. Paste the script. Add your FAQ as a small knowledge file if the tool supports it.
3. Connect a test number or use the web call tester. Call it yourself three times: calm caller, rushed caller, angry caller.
4. Add a warm transfer to your mobile. Fail closed: if the agent is unsure, transfer or take a message.
5. Wire the outcome to email, Slack, or a sheet. You want the transcript where you already work.
6. Run it on nights and weekends only for seven days. Review every transcript. Fix the top three confusion points.
7. Only then move it to missed-call overflow during the day.

Expect the first week to feel fiddly. That is normal. You are teaching tone and boundaries, not installing a toaster. Keep a human in the loop until the transcripts look boring in a good way.

Rules you cannot skip (legal and trust)

Inbound reception agents that answer your own number are the lower-risk starting point. Outbound sales or marketing calls with an AI voice sit under stricter US rules.

In February 2024 the FCC confirmed that the Telephone Consumer Protection Act (TCPA) restrictions on artificial or prerecorded voices cover AI systems that generate human-sounding voices. Calls that use those technologies generally need the prior express consent of the called party (FCC Declaratory Ruling FCC-24-17). The FTC has also affirmed that Telemarketing Sales Rule protections against robocalls extend to voice-cloning style abuses in telemarketing (FTC March 2024 announcement).

Practical habits for a small business:

• Say early that the caller is speaking with an AI assistant for your company.
• Offer a path to a human on every call.
• Do not clone a real person’s voice for outreach without clear rights and consent.
• Keep recordings and transcripts only as long as your policy needs, and tell people if you record.
• Get local legal advice before any outbound campaign. This article is not legal counsel.

Trust is the product. If a caller feels tricked, you lose the lead and the review. Honesty costs one sentence at the start of the call.

Mistakes that make callers hang up

• Too wide a job. “Be our receptionist” is not a script. One call type is a script.
• No escape hatch. If the agent cannot transfer, frustrated callers bang the handset.
• Invented prices or promises. Lock the agent to your published numbers or say it will confirm by text.
• Long monologues. Ask, listen, confirm. Callers interrupt. Let them.
• Ignoring transcripts. The gold is in the failure cases from day two and three.
• Launching on the main line on Monday morning. Start after hours. Earn the promotion.

AI chatbot vs voice agent for small business customer service

People searching this topic often mix chatbots and phone agents. They solve related problems on different channels.

• Chatbot / text agent: website widget, WhatsApp, email draft. Cheap per conversation, easy to edit, great for FAQs and ticket triage. See prompts for desk work in our ChatGPT prompts for customer support agent kit.
• Voice agent: live phone. Higher per-minute cost, higher emotion, higher payoff when the customer will not open a chat.

Many small businesses need both later. Start with whichever channel already loses you money. If your missed-call list is longer than your unread chat queue, start with voice. If your site chat is full of the same three questions, start with text.

Pricing for chat tools is usually per seat or per resolution. Pricing for voice is usually per minute. Do not compare them on a single “AI customer support” line item without separating channels.

FAQ: AI voice agent for small business

What is the best AI voice agent for small business?

The best one is the one you can test this week, price for 500 minutes, and warm-transfer to a human. Retell, Vapi, Bland, and ElevenLabs Agents are common starting points because they let non-engineers ship a first agent. Re-score after you have real transcripts, not after a vendor demo.

How much does an AI voice agent cost per month?

Usage often lands near the low hundreds of dollars for a few hundred to a thousand minutes, based on published per-minute rates such as Retell’s $0.07 to $0.31 range and their ~$0.125 example configuration. Add number fees and any one-time setup help. Measure your own minutes before you forecast a year.

Can a small business owner set this up without coding?

Yes, for intake, FAQ, and simple booking. You write instructions in plain language, connect a calendar or webhook, and test. Custom CRM deep-links or multi-location routing may need a freelancer later. Your first agent should not need that.

Is an AI receptionist legal?

Answering your own inbound line with a disclosed AI assistant is widely done. Outbound AI voice calls sit under TCPA and telemarketing rules in the US, including the FCC’s confirmation that AI-generated human voices are treated as artificial or prerecorded voice. Disclose, get consent where required, and talk to a lawyer before dialing lists.

Should I replace my receptionist?

Usually no. Cover the gaps: nights, overflow, repetitive intake. Keep humans for walk-ins, upset customers, and anything that needs judgment. The goal is fewer missed leads, not a ghost office.

Your next step

Write one page: the single call you miss most. Pick a platform with a free trial. Ship an after-hours agent that takes a message and emails you the transcript. Review seven days of calls. Then decide whether to expand.

When you are ready to go deeper on platforms and build steps, use the AI voice agent platform guide. For wider automation around the desk, keep building small agents that save real hours, not science projects.

Want practical AI you can use at work this week? Join Launchpad free below.

If you are weighing chat and voice for the front desk, read our guide to conversational AI agents for businesses — it shows how to pick one narrow job and a clear human handoff before you buy tools.

Tidying the books is another task AI can take on. Read our guide to AI bookkeeping for small business.

The calendar invite says “Q4 kickoff, 5 minutes from you.” Or a colleague is leaving on Friday and somehow you are the one giving the farewell. You open a blank document, type “Thank you all for coming,” and stall.

Here is how to write a speech with AI without it sounding like a greeting card: do not ask ChatGPT, Claude or Gemini to “write me a speech.” First dump your own facts, names and one or two real stories into the chat. Tell it who is listening, how long you have, and the one thing you want people to remember. Ask for an outline before a draft, then a draft written for the ear: short sentences, plain words, room to pause. Read it out loud, cut every line you would never say, and rehearse once with a timer or an AI coach. A five-minute work speech usually takes about 30 to 45 minutes this way, and most of that time is you talking, not typing.

Below: the six steps, copy-paste prompts, a worked example, prompts for the most common work speeches, and the mistakes that make an AI speech fall flat.

Why “write me a speech” gives you a speech nobody believes

Ask any chat assistant for “a speech for my team’s year-end meeting” and you get the same thing everyone gets: “journey,” “incredible milestones,” “none of this would have been possible without each and every one of you.” It is not wrong. It is just nobody’s speech.

The AI has no idea what your team actually did this year, who stayed late before the launch, or what joke only your department will get. A speech works when it is specific and when it sounds like the person saying it. Those two things have to come from you. What the AI is good at is the slow, fiddly part: putting your scattered points in a clear order, finding a strong opening line, trimming a ten-minute ramble down to five, and pointing out where a sentence is hard to say out loud.

So the method below keeps a simple split. You bring the material and the voice. The AI brings structure, speed and a second pair of eyes.

What you need

• A chat assistant: ChatGPT, Claude or Gemini. The free plans are enough for a speech. New to them? Our ChatGPT tutorial for beginners, Claude tutorial for beginners and Gemini tutorial for beginners get you set up.
• Ten minutes of your own notes: names, numbers, dates, the moments you want to mention. Rough is fine.
• Your phone or a microphone for the rehearsal at the end.

Step 1: Dump the facts, not the speech (5 minutes)

Before you open the AI, write down everything the speech could include. Do not try to make it sound good. Just list it:

• The occasion and the audience (“all-hands, about 40 people, half of them remote”).
• What happened: projects, results, the hard bits, the people involved.
• One or two small stories. “Sam fixed the checkout bug at 11pm and then brought croissants the next morning” beats any adjective the AI can come up with.
• Anything you must say (a thank-you, an announcement, a next step) and anything you must not say (numbers that are not public yet, a sensitive reorg).

Typing feels slow? Talk instead. Record yourself rambling for three minutes on a walk, then turn the recording into text. Our guide on how to transcribe audio with AI shows the quickest ways. Speech that starts as speech usually sounds more natural at the end.

One rule for work: check your company’s AI policy before you paste in confidential numbers, client names or HR details. If in doubt, use placeholders like “Client A” and fill them in afterwards.

Step 2: Pick the one thing and the length (2 minutes)

Toastmasters, the public speaking organisation, suggests asking yourself: if listeners only remember one thing from my speech, what do I want that one thing to be? Your audience should be able to sum up your speech in one sentence (Toastmasters International, How to Write and Deliver a Speech). Write that sentence down. It goes into every prompt.

Then turn your time slot into a word budget. People speak at very different speeds, but a comfortable pace for slide presentations is often given as roughly 100 to 125 words a minute, and audiobooks are recommended at around 150 to 160 (Wikipedia, Words per minute). Toastmasters gives a similar rule of thumb: about three double-spaced pages take around five minutes to say out loud.

In practice: aim for about 120 words per minute of speaking time. Three minutes is roughly 360 words. Five minutes is roughly 600. If there is no fixed slot, go shorter. Nobody has ever complained that a kickoff speech ended early.

Step 3: Ask for an outline first, not a full draft (5 minutes)

A full draft in one go locks you into the AI’s structure before you have agreed with it. An outline is quick to check and quick to fix. Paste your notes from Step 1 below this prompt:

I need to give a [length]-minute speech at [occasion] to [audience].
The one thing I want them to remember: [your sentence].
Below are my raw notes. Use only facts from my notes. Do not invent stories, names, numbers or quotes. If something is missing, ask me.

Give me 3 different outlines (opening, 2 or 3 main points, close). For each, suggest an opening line that starts with a moment or a question, not with "thank you for coming".

My notes:
[paste notes]

Pick the outline that feels closest to how you would tell it over coffee. Change the order if you like. Then tell the AI which one you chose and what you changed.

Why three openings? Toastmasters points out that you have about 30 seconds to catch the audience’s attention, and recommends starting with a story or a small conflict instead of a formal greeting (Toastmasters). Seeing three options side by side makes the strongest one obvious.

Step 4: Get a draft written for the ear (5 minutes)

A speech is heard once, at the speed you talk. Nobody can scroll back. That changes how it should be written, and you have to tell the AI, because by default it writes for reading.

Write the speech from outline [B], about [number] words.
Write it to be spoken, not read:
- short sentences, most under 15 words
- plain words I would use in a meeting, no "journey", "testament", "synergy" or "each and every one of you"
- use "we" and "you" more than "I"
- repeat the main message once in the middle and once at the end
- mark a pause with [pause] where a line needs a moment to land
- end with one clear thing I want people to do or feel
Use only the facts and stories from my notes.

Two of those lines come straight from speaker advice. Toastmasters recommends using “we” and “you” more than “I” to pull the audience in, and using repetition to reinforce your main point (Toastmasters). The word ban list is ours: those are the words that make listeners tune out, because they have heard them in every corporate video.

Step 5: Make it sound like you (10 minutes)

This is the step most people skip, and it is the one that matters most. Read the draft out loud once, at full volume, as if the room were already there. You will notice three kinds of lines:

• Lines you would never say. Too formal, too clever, a word you would not use. Mark them.
• Lines you trip over. Too long, or two hard sounds in a row. Mark them.
• Lines that say nothing. “This has been a truly remarkable year.” Remarkable how? Mark them.

Then give the AI a sample of how you actually talk. An email you wrote quickly, a Slack message, or a transcript of you explaining something works better than a description like “casual but professional.”

Here is a sample of how I normally talk: [paste a message or transcript].
Rewrite the lines I marked with ** so they sound like the sample. Keep everything else as it is.
Then list any remaining line that sounds like a greeting card or a press release, and suggest a plainer version.

Finally, check every fact. Names spelled right, numbers correct, the project called what your team calls it. AI assistants can get details wrong, and a wrong name in a farewell speech is the one thing people will remember.

Step 6: Rehearse with AI and a timer (10 minutes)

Practising out loud is the foundation of every other speaking tip, according to Toastmasters, which also advises timing yourself and recording yourself to catch filler words and speed (Toastmasters). Two AI tools make this easier than talking to a mirror:

• Speaker Coach in PowerPoint. Put your speech title on a single slide, choose Rehearse with Coach on the Slide Show tab, and give the speech. Microsoft says it looks at your pacing, pitch, filler words and wordiness, and gives you a report afterwards. It works in English, and Microsoft describes it as available in free preview on the web and mobile (Microsoft Support, Rehearse your slide show with Speaker Coach).
• ChatGPT Voice. Start a voice conversation and say “Wait until I ask you to respond.” OpenAI’s help page lists that as a way to keep Live from answering while you talk. Give the speech, then ask: “What was my main point? Where did it drag? Which line should I cut?” OpenAI also notes that voice transcripts may not exactly match what was said, so treat its feedback as a rough check (OpenAI Help Center, ChatGPT Voice).

If you run over time, do not speed up. Ask the AI: “Cut this to [number] words without losing the two stories.” Then rehearse once more. After two run-throughs, switch from the full text to five or six bullet points on a card, so you look at people instead of paper.

An example: a 3-minute thank-you at the team meeting

Made-up example to show the flow. Lena leads a support team of twelve. After a rough quarter with a new ticketing system, she has three minutes at the monthly meeting to say thank you without it sounding like an HR email.

Her notes (Step 1): new system went live in July, first two weeks chaos, queue doubled, Priya built a cheat sheet everyone used, Marco answered tickets on his day off (she wants to thank him but not encourage that), by September the queue was back to normal, next goal: shorter first replies.

Her one thing (Step 2): “We got through the hard part because people helped each other, and that is how we will hit the next goal too.” Three minutes, so about 360 words.

What the AI did: gave three outlines; she picked the one that opens with “Week two of the new system, the queue counter hit a number I had never seen before.” The ear-friendly draft came back at 410 words. Reading it out loud, she cut “a testament to our resilience,” replaced “leverage” with “use,” and asked the AI to rewrite the Marco line so it thanked him and also said clearly that days off are days off.

Result: a 3 minute 10 second rehearsal in Speaker Coach, which flagged a lot of “so” at the start of sentences. She took out four of them. Total time spent: about 35 minutes, of which ten were the walk where she recorded her notes.

Prompts for the most common work speeches

Use these as the first line of the Step 3 prompt, then add your notes.

KICKOFF / ALL-HANDS
A [5]-minute kickoff for [team] at the start of [quarter/project]. Goal: everyone leaves knowing the top 3 priorities and why they matter. Open with one concrete moment from last quarter.

FAREWELL FOR A COLLEAGUE
A [3]-minute farewell for [name], who is leaving after [years]. Warm, a little funny, never embarrassing. Use only the stories in my notes. End with a toast or a clear send-off line.

ACCEPTING AN AWARD OR THANKS
A [90-second] thank-you after receiving [award]. Short, humble, names the people who helped, ends quickly.

WELCOME AT A CLIENT OR COMPANY EVENT
A [2]-minute welcome for [event]. Say who we are, why we are all here, and what happens next. One light opener, no jokes about the audience.

OPENING OF A CONFERENCE TALK
The first 60 seconds of a [20]-minute talk on [topic] for [audience]. Start with a problem they recognise, then say what they will be able to do after the talk.

Giving a talk with slides? Write the speech first with this method, then build the deck around it. Our guide on how to use AI to make a presentation covers the slides and speaker notes side.

Which AI is best for writing a speech?

For a work speech, the honest answer is: the one you already use. ChatGPT, Claude and Gemini can all outline, draft and tighten a speech well, and the difference between them matters much less than the quality of your notes and the sample of your own voice. A few practical differences:

• ChatGPT has the voice mode described above, which is handy for rehearsing on your phone.
• Claude handles long pasted notes and transcripts comfortably and tends to follow “do not invent anything” instructions closely, in our experience.
• Gemini is convenient if your notes already live in Google Docs or Gmail.

Dedicated “AI speech writer” websites exist too. Many of them work much like the prompts above, with a form in front. If you use one, the same rule applies: feed it your real material, or you get a generic speech.

Mistakes that make an AI speech fall flat

• Letting it invent stories. An AI will happily write “I remember the day Sarah walked into my office…” about a Sarah who does not exist. Always include “use only facts from my notes” and check the draft.
• Going over time. Drafts often come back longer than you asked. Count the words and cut before you rehearse, not on the day.
• Keeping the written style. Long sentences with three commas read fine and sound terrible. If you run out of breath mid-sentence, split it.
• Reading it word for word. Use the full text to rehearse, bullet points to deliver.
• Pasting in confidential details. Follow your company’s AI rules and use placeholders when unsure.
• Skipping the read-aloud. Ten minutes of reading out loud fixes more than an hour of prompting.

Is it OK to use AI to write a speech?

For most work occasions, yes, as long as the speech is honest. The stories are yours, the thanks are real, and you are the one standing up and saying it. Think of the AI as the colleague who helps you tidy your notes, not the person who talks for you. Where it gets uncomfortable is when the content is not yours: invented anecdotes, borrowed jokes passed off as your own experience, or a eulogy for someone the AI never knew. Keep the material real and the help is no different from asking a friend to read your draft.

Quick answers

Can AI write a speech for free?

Yes. The free plans of ChatGPT, Claude and Gemini are enough for outlining, drafting and editing a speech of a few minutes. Speaker Coach is described by Microsoft as available in free preview on the web, even without an Office subscription.

How many words is a 5-minute speech?

Roughly 500 to 750 words, depending on how fast you speak. About 600 words is a safe starting point. The only reliable check is reading it out loud with a timer, and remember you will probably read faster alone than in front of an audience.

How do I make ChatGPT write a speech in my own voice?

Give it a sample of how you really talk (a quick email, a Slack message, a transcript of you explaining something) and ask it to match that. Then read the result aloud and rewrite any line you would not say. The sample does most of the work.

Can AI help me with nerves before a speech?

It can help with the part you control, which is preparation: a clear structure, a speech that fits the time, and a few rehearsals with feedback. It cannot stand up for you. If public speaking is a big hurdle, practising in front of real people, for example at a Toastmasters club, gives you something an AI cannot.

Can I use the same method for a wedding toast?

Yes. The steps are the same: your stories, one main message, a word budget, a draft for the ear, and a read-aloud. Just be extra strict about “use only my notes,” because invented memories at a wedding are worse than at a team meeting.

Your next step

Got a speech coming up? Do Step 1 today: five minutes of notes, or a three-minute voice memo on your next walk. That is the part only you can do, and with it in hand the rest of this guide takes about half an hour.

Want more short, practical ways to use AI at work, one real task at a time? Join the free Launchpad.

Prefer to look around first? Start on our home page.

Writing the draft in ChatGPT, Claude, or Gemini? See ChatGPT vs Claude vs Gemini for which one fits speech and presentation work.

Sources

• Toastmasters International: How to Write and Deliver a Speech
• Wikipedia: Words per minute (Speech and listening)
• Microsoft Support: Rehearse your slide show with Speaker Coach
• OpenAI Help Center: ChatGPT Voice

You know what your business does. You can explain it to a client over coffee in two minutes. But the moment you open a blank homepage, the words go flat: “We deliver tailored solutions.” You stare at it, delete it, and the website waits another month.

Here is how to write website copy with AI: do not ask ChatGPT or Claude to “write my homepage”. Give it the raw material first: who your customer is, the problem they have, what you offer, the proof you can actually show, and three sentences you wrote yourself so it can copy your voice. Then plan the page section by section, let the AI draft one section at a time, and edit every line for two things: can a busy visitor scan it in ten seconds, and is every claim true. Done this way, a first homepage draft fits in one focused afternoon, and the result sounds like you, not like every other site.

Below: the five steps, prompts you can copy for the homepage, about page, services page and landing page, a worked example, and the mistakes that make AI website copy sound like AI.

For Claude-heavy desk days beyond website copy, keep the best Claude prompts for work kit handy — email, meeting notes, and research drafts in the same paste-and-tweak format.

Why most AI website copy sounds the same

Ask any chat assistant for a homepage with no context and you get the average of the internet: “Welcome to [Company], where innovation meets excellence.” It is not the tool being lazy. It simply knows nothing about you, so it fills the gaps with the safest, most common phrases it has seen.

The fix is not a magic prompt. It is input. OpenAI and Anthropic both say the same thing in their own prompting guides: give the model clear instructions, the context it needs, and examples of what good looks like (OpenAI, prompt engineering guide, Anthropic, prompt engineering overview). For website copy, the context is your customer and your offer. The examples are your own sentences.

There is a second reason it matters. Website visitors mostly do not read. In Nielsen Norman Group’s research, 79 percent of test users scanned any new page they came across, and only 16 percent read word by word (Nielsen Norman Group, How Users Read on the Web). Their later analysis found the first 10 seconds of a visit are when people decide to stay or leave (Nielsen Norman Group, How Long Do Users Stay on Web Pages?). Generic copy loses that ten seconds. Specific copy wins it.

What you need

• A chat assistant: ChatGPT, Claude or Gemini. The free plans are fine for a single page. New to them? Our ChatGPT tutorial for beginners and Claude tutorial for beginners get you set up in half an hour.
• Your raw facts: what you sell, the price or price range if you show it, who it is for, how it works, and any real proof (client names you may use, real reviews, years in business, certifications).
• Three to five sentences in your own voice: an email to a client, a LinkedIn post, a voice memo you transcribed. This is what stops the draft from sounding like a brochure.
• Your customers’ own words, if you have them: review text, survey answers, the questions people ask on sales calls. These are the best headlines you will ever get, because they are how buyers actually describe the problem.

Step 1: Build a one-page copy brief (15 minutes)

Everything else depends on this step. Instead of filling in a form, let the AI interview you. It asks better follow-up questions than a blank document does, and you answer in plain speech.

Answer like you would answer a friend. Short and honest beats polished. When the brief comes back, read it once and fix anything that is not quite right. This page is now the single source for every page you write next.

If you have review text or survey answers, paste them in and add: “List the exact phrases customers use to describe their problem and the result they got. Quote them, do not paraphrase.” Those quotes will feed your headlines. Our guide to using AI for market research shows how to collect more of them from competitors and forums.

Step 2: Plan the page before the words (10 minutes)

Most weak pages fail on order, not wording. Decide what each section has to do before anyone writes a sentence. A homepage for a small service business usually needs these blocks, top to bottom:

1. Headline: what you do and for whom, in plain words.
2. Subheadline: the result the customer gets, or how it works in one line.
3. Main button: the one action you want (book a call, get a quote, start a trial).
4. The problem: two or three lines that show you understand their situation.
5. How it works: three steps, so the next move feels small.
6. Proof: real reviews, real client logos you are allowed to use, real numbers.
7. Offer details: what is included, and the price if you show it.
8. FAQ: the objections from your brief, answered honestly.
9. Final button: the same action as the top.

Read the plan as your customer would. If a section has no fact from the brief behind it, cut it or go find the fact.

Step 3: Draft one section at a time (30 to 60 minutes)

Asking for the whole page in one go gives you a long block of average text. Asking per section gives you choices. Start with the headline, because it shapes everything under it.

Pick one or two, then move down the plan: “Now write the subheadline and the button text for headline 4. Give three options each.” Then the problem block, then the three steps, and so on. Keep the same chat open so the brief stays in context.

Two follow-ups that improve almost every draft:

• “Which sentence in this section would a busy visitor skip? Rewrite it or cut it.”
• “Rewrite this so every sentence starts with the most important word.”

That second one comes from how people scan: they read the first words of lines and headings and decide from there (Nielsen Norman Group, First 2 Words).

Step 4: Edit for scanning and for truth (20 minutes)

This is the step that makes it yours. Print the draft or read it out loud, and check three things.

Is it concise, scannable and objective?

In a classic Nielsen Norman Group study, rewriting a site to be concise, scannable and objective (less promotional language) together measured 124 percent better usability than the original promotional version (Nielsen Norman Group, Concise, SCANNABLE, and Objective). In practice: short paragraphs, headings that say something, bullet lists for features, and fewer adjectives.

Is every claim true?

AI will happily write “trusted by hundreds of businesses” or a glowing quote from “Sarah, Marketing Director”. If that did not happen, delete it. This is not only about trust. In the United States, the FTC’s rule on consumer reviews and testimonials bans fake ones, including AI-generated reviews from people who do not exist or never used the product (Federal Register, Trade Regulation Rule on the Use of Consumer Reviews and Testimonials, 2024). Use real reviews, with permission, or leave the proof block out until you have them. Our guide to responding to Google reviews with AI is a good way to start collecting them.

Does it sound like you?

Read it out loud. Any sentence you would never say to a client gets rewritten in your words. Two or three of your own phrases per section is usually enough to make the whole page feel human.

Step 5: Add the SEO title and meta description, then publish (10 minutes)

Each page needs a title tag and a meta description. Google’s own guidance: write titles that are descriptive and concise, and avoid vague ones like “Home” (Google Search Central, title links). For descriptions, Google mostly builds the snippet from your page content, but it may use your meta description when it describes the page better, and it shortens long ones to fit the screen (Google Search Central, snippets).

Publish, then open the page on your phone. Can you tell what the business does and what to do next within ten seconds? If not, the headline or the top button needs another pass. If you are writing blog articles for the same site next, our guide on using AI to write a blog post that sounds like you uses the same voice-first approach.

Copy-paste prompts for each page

All of these assume the copy brief from Step 1 is already in the chat.

Homepage

About page

Services page

Landing page for one offer

FAQ section

A worked example (made up, to show the method)

Say you are a bookkeeper who mostly works with independent cafés. The business and every detail below are invented for this example.

Before (first AI draft with no brief): “Welcome to Ledger Bright, your trusted partner for innovative financial solutions. We help businesses of all sizes unlock their full potential.”

The brief added: customers are café owners with one to three locations. Their problem: receipts in a shoebox and a VAT deadline every quarter. They have tried doing it themselves on a Sunday night. The offer: monthly bookkeeping at a fixed monthly fee, receipts photographed on the phone, a five-line summary every month. Proof: two real client reviews the owner has permission to use.

After (headline and subheadline picked from the options):

Same tool, same afternoon. The difference is the brief. The “after” version names the customer, the pain and the next step, and nothing in it needs to be taken on faith.

ChatGPT, Claude or Gemini for website copy?

All three can do this workflow well, and the brief matters far more than the model. In our own use, Claude’s first drafts often come out a little less salesy and ChatGPT is quick at giving lots of short variations, but that is a matter of taste, so try your headline prompt in two of them and keep the one you like. For a closer look at the differences in writing work, see our comparison of ChatGPT vs Claude for writing.

Dedicated “AI website copy generator” tools exist too. Most of them are a form that feeds your answers into a similar prompt. They can save a few minutes, but you still need the brief and the edit, so start with the chat assistant you already have.

Does Google penalise AI-written website copy?

Not for being AI-written. Google’s Search team says it rewards high-quality content however it is produced, and that appropriate use of AI is not against its guidelines. What breaks the rules is using automation mainly to manipulate rankings (Google Search Central Blog, Google Search’s guidance about AI-generated content). Its guide on helpful, people-first content is a good checklist for any page: is it written for your visitors, does it show real expertise, and would they leave satisfied?

Website copy built from your own brief, your own facts and your own voice passes that test easily. Fifty near-identical location pages pumped out by a prompt do not.

Common mistakes, and the quick fix

• Buzzword soup (“seamless”, “cutting-edge”, “solutions”). Fix: put a banned-words list in every prompt, and ask “replace every abstract word with a concrete one”.
• Too long. AI pads. Fix: give a word limit per section and ask “cut 30 percent without losing any fact”.
• Invented proof. Fake numbers, fake quotes. Fix: the UNSUPPORTED check in Step 4, every time.
• Five different buttons. Fix: one main action per page, repeated at the top and the bottom.
• Same copy on every page. Fix: each page gets its own job in the plan (Step 2), and your headline should match what people searched to land there.
• Your product pages read like a catalogue. That is a slightly different job. Our guide to writing product descriptions with AI covers shop pages.

FAQ

Can AI write my whole website?

It can draft all of it. It cannot know your customers, your prices or your proof unless you tell it, and it cannot check whether its claims are true. Plan on the AI doing most of the typing and you doing the thinking and the final edit.

How long should homepage copy be?

Long enough to answer what a first-time visitor needs to decide on the next step, and no longer. For a small service business that is often 300 to 600 words, but there is no fixed rule. If a section does not help someone decide, cut it.

Is it safe to paste client details into ChatGPT or Claude?

Leave out names and anything private, and check what your company allows. Reviews you already publish are fine. Contracts, client emails and personal data are not something to paste into a free chat account without approval.

Should I write the copy first or design the website first?

Copy first, or at least the copy plan from Step 2. Design is much easier when you know which blocks the page needs and roughly how long each one is.

Can I use the same prompts for a website in another language?

Yes. Write the brief in your strongest language, then ask for the copy in the target language with “write as a native speaker in [country], not a translation”. Have a native speaker read it before you publish.

Writing for ears instead of a homepage? The same brief-first habit works for talks. Here is how to write a speech with AI that still sounds like you, with copy-paste prompts.

Your next step

Open a new chat today, paste the interview prompt from Step 1, and answer the questions. By the end of the hour you will have a copy brief, and the homepage headline is twenty minutes after that. Once that works, the next time saver is automating the rest of the small-business admin: see AI automation for small business: what to automate first.

Still choosing a writing tool for the draft? Compare ChatGPT vs Claude vs Gemini for which one fits website and marketing copy.

Want more short, practical lessons like this one? Join Launchpad, our free starting point.

Prefer to look around first? Start on our home page.

Sources

• Nielsen Norman Group: How Users Read on the Web
• Nielsen Norman Group: How Long Do Users Stay on Web Pages?
• Nielsen Norman Group: First 2 Words, A Signal for the Scanning Eye
• Nielsen Norman Group: Concise, SCANNABLE, and Objective, How to Write for the Web
• OpenAI: Prompt engineering guide
• Anthropic: Prompt engineering overview
• Google Search Central: Influencing your title links
• Google Search Central: Control your snippets in search results
• Google Search Central Blog: Google Search’s guidance about AI-generated content
• Google Search Central: Creating helpful, reliable, people-first content
• Federal Register: Trade Regulation Rule on the Use of Consumer Reviews and Testimonials (FTC, 2024)

You walked the new hire through the expense policy. You ran the workshop, sent the slides, and everyone nodded. Two weeks later the same questions land in your inbox, and you realise you never knew what actually stuck.

Here is how to make a quiz with AI: give ChatGPT, Claude or Gemini your own material (the policy, the slides, the training notes), tell it who the quiz is for and what they must be able to do, and ask for 8 to 10 questions with an answer key that quotes the line each answer comes from. Check every answer against your document. Then paste the questions into Google Forms or Microsoft Forms, switch on quiz mode, and the form scores it for you. If your plan includes Gemini in Google Forms or Copilot in Microsoft Forms, they can draft the quiz inside the form straight from your file. A first 10-question quiz takes most people around 20 minutes, and most of that is the checking.

Below: the five steps, prompts you can copy, a worked example, how to quiz from a PDF, and what to fix when the questions come out too easy or plain wrong.

Why a short quiz is worth 20 minutes

A quiz is not only a test. Pulling an answer out of your own memory helps you keep it far better than reading the same page again. Psychologists call this the testing effect or retrieval practice, and it has been studied for more than a century (Wikipedia, Testing effect). In a well-known experiment, students who took a test on a text remembered more of it a week later than students who spent that time studying it again (Roediger and Karpicke, Psychological Science, 2006).

For your work that means two things. The people who take your quiz remember the policy or the process better. And you get a list of the questions most people missed, which tells you exactly which part of your training needs a better explanation. Google Forms, for example, shows you the frequently missed questions in its quiz summary (Google Docs Editors Help, Create and grade quizzes).

What AI does well here is the slow part: turning a dense page into clear questions with believable wrong options. What it does badly is knowing whether an answer is right for your company. That part stays with you.

What you need

• A chat assistant: ChatGPT, Claude or Gemini. The free plans are enough for this. New to them? Start with our ChatGPT tutorial for beginners, the Claude tutorial for beginners or the Gemini tutorial for beginners.
• Your source material: the document, slides, process description or meeting notes the quiz is about. Without it, the AI writes generic questions about the topic in general, not about what you taught.
• A form tool with a quiz mode: Google Forms or Microsoft Forms. Both let you mark the correct answer, give points and show feedback (Google Forms help, Microsoft Forms help).

Step 1: Decide what people must be able to do (2 minutes)

Most weak quizzes test trivia: “In what year was the policy introduced?” Nobody needs that. Before you open the AI, finish this sentence: After this training, people should be able to…

• “…submit an expense claim correctly the first time.”
• “…tell a customer our refund rules without checking with a manager.”
• “…spot a phishing email and know who to report it to.”

Pick two or three of these. They become the instruction that keeps the AI on the questions that matter.

Step 2: Give the AI your material (2 minutes)

Paste the text into the chat, or upload the file. Shorter is better: one policy or one module per quiz, not the whole handbook. If your material lives in a recording, turn it into text first with our guide on how to transcribe audio with AI.

One honest warning: check what your company allows before you upload internal documents to a chat assistant. If in doubt, remove names, client details and figures first. The questions work just as well without them.

Step 3: Ask for the quiz with this prompt (3 minutes)

This prompt does the heavy lifting. Copy it, fill in the brackets, and put your material under it.

Write a quiz based ONLY on the material below.

Who takes it: [e.g. new sales staff in their first week]
After the quiz they should be able to: [your 2 or 3 goals from step 1]

Rules:
- 10 questions: 7 multiple choice (4 options), 3 short answer.
- Test understanding and real situations, not dates or trivia.
- Wrong options must be believable mistakes a real person makes.
- No "all of the above", no trick questions, no double negatives.
- Keep the correct option about the same length as the others.

After the questions, give an answer key. For every answer:
the correct option, one sentence explaining why, and the exact
line from the material it is based on.
If the material does not answer something, do not invent it.
List it at the end under "Gaps in the material".

Material:
[paste here]

The last part matters most. Asking for the exact source line turns your check in step 4 from guesswork into a quick comparison. The “Gaps” list is a bonus: it often shows you what your training document forgot to explain.

Step 4: Check every question against the source (8 minutes)

Do not skip this. AI tools can write answers that sound right and are not. Microsoft says the same about its own quiz assistant: the content it generates can be incomplete or inaccurate, so double-check it (Microsoft Support, Copilot in Forms FAQ).

For each question, ask yourself four things:

1. Is the answer in the quoted line? Find that line in your document. If it is not there, or says something slightly different, fix or delete the question.
2. Is exactly one option right? If a colleague could argue for two, rewrite one of them.
3. Would someone who skipped the training get it wrong? If the answer is common sense, the question tests nothing.
4. Does it match how things really work here? Documents go out of date. You know that the policy says 30 days but finance now wants receipts in 14.

Then let the AI review its own work once:

Review this quiz as a strict trainer. For each question tell me:
is the answer fully supported by the material, could more than
one option be argued as correct, and is it guessable without
the training? Suggest a better version for any weak question.

Step 5: Put it in Google Forms or Microsoft Forms (5 minutes)

Google Forms. Open a new form, go to Settings and turn on “Make this a quiz”. Add each question, then use “Answer key” to mark the right option and set the points. “Add answer feedback” lets you show your one-sentence explanation after someone answers. Under the respondent settings you choose whether people see missed questions and correct answers (Google Docs Editors Help).

Microsoft Forms. Choose New Quiz, add your questions, tick the Correct answer mark next to the right option and fill in the points. You can add a message per answer that people see when they choose it (Microsoft Support, Create a quiz with Microsoft Forms).

Let the form do the drafting (if your plan has it). With an eligible Google Workspace or Google AI plan, “Help me create a form” in Google Forms can generate a quiz with correct answers and points marked, and you can point it at a Doc, Slides deck or PDF from your Drive. It works on desktop and is still rolling out, so you may not see it yet (Google Docs Editors Help, Create a form with Gemini). In Microsoft Forms, Copilot can draft a quiz from a description or from pasted text when you have a Microsoft 365 account with a Copilot license (Microsoft Support). Either way, step 4 still applies: check the answer key before you send it.

Before you share it, take the quiz yourself once on your phone. You will catch the question with two right answers faster that way than by reading.

A worked example (made up, so you can see the flow)

Say you run a small agency and you want new starters to get expense claims right. Your goal from step 1: “submit an expense claim correctly the first time.” You paste your one-page expense policy under the prompt. Three of the ten questions that come back look like this:

1. You took a client to lunch. What do you need to submit
   with the claim?
   a) Only the card statement
   b) The itemised receipt and the client's name
   c) The receipt, but only if it is over the limit
   d) Nothing, client lunches go through the account manager

2. A train ticket for a project trip cost more than the
   travel limit. What do you do?
   a) Split it over two claims
   b) Ask your manager to approve it before booking
   c) Book it and explain later
   d) Pay the difference yourself

3. (Short answer) Within how many days after the expense
   must you submit your claim?

Answer key
1. b. Client meals need proof of who attended.
   Source: "Client meals: attach the itemised receipt and
   the name of the client."
2. b. Spending above the limit needs approval in advance.
   Source: "Anything above the limit must be approved by
   your manager before you book."
3. 30 days. Source: "Submit claims within 30 days."

Now the check. Question 3 quotes “30 days”, but you know finance changed it to 14 last month. So you fix the quiz and the policy document. That is the kind of thing only you can catch, and it is why step 4 matters more than step 3. (The policy lines above are invented for this example.)

How to make a quiz from a PDF or slide deck

Gemini Notebook. Google’s notebook tool (formerly called NotebookLM, renamed in July 2026) can turn your uploaded sources into a quiz. Open a notebook, add your PDF or slides, and in the Studio panel choose “Quizzes”. With the pencil icon you can set the difficulty and add a prompt for the audience and focus. While answering, there is a hint button and an “Explain” option (Gemini Notebook Help, Generate Flashcards or Quizzes). Google has since added more formats, such as short-answer, multi-select and fill-in-the-blank questions (Google blog, October 2026). It is great for testing yourself on a long report before a meeting. To collect scores from a team, a form is the better home.

ChatGPT or Claude with the step 3 prompt. Upload the PDF and change “the exact line” to “the exact line and page number”. The page numbers make your check twice as quick. For very long files, our guides on analysing long PDF files with Claude and turning a PDF into notes with Claude show how to get the material into a usable shape first.

Copy-paste quiz prompts for work

Onboarding knowledge check. Pairs well with our guide on using AI for employee onboarding.

From this onboarding document, write 8 situation-based multiple choice questions a new hire should get right by the end of week one. Each question starts with a realistic situation ("A customer asks..."). Answer key with the source line for each.

After a workshop or training session.

Here are the slides and my notes from today's session. Write 5 short questions people can answer in 3 minutes tomorrow morning to recall the key points. Mix 3 multiple choice and 2 short answer. Only use what is in the material.

Turning a process into a quiz. If you already have written procedures (see how to use AI to write SOPs), quiz the steps that go wrong most:

From this procedure, write 6 questions about the steps people most often get wrong or do in the wrong order. Include one question where the reader must put 4 steps in the right order.

Quiz yourself before an important meeting.

Quiz me on this report, one question at a time. Wait for my answer, tell me if I am right, quote the passage, then ask the next question. Start easy and get harder. Stop after 10 and list what I missed.

Multiple choice or short answer?

Multiple choice is easy to score automatically, which is why forms love it. Short answers make people work harder to recall, and research summaries report larger learning gains from short-answer questions than from multiple choice, and feedback after a quiz helps learning too (Wikipedia, Testing effect). A practical mix: mostly multiple choice for the facts that must be exact, plus two or three short answers for the things people need to explain in their own words. Always switch on the answer feedback, so a wrong answer becomes a small lesson instead of just a lost point.

Quiz or survey?

They live in the same form tools, but they do different jobs. A quiz has right answers and checks what people know. A survey has no right answers and asks what people think. Want to know whether the training was useful or what people still find confusing? That is a survey, and our guide on how to use AI to create a survey covers it. Many trainers send both: a short quiz, then two survey questions at the end.

When the quiz comes out wrong: quick fixes

• The questions are too easy. Ask: “Rewrite every question as a realistic situation where the reader has to apply the rule, not repeat it.”
• The right answer is always the longest one. A classic giveaway. Ask: “Make all four options similar in length and detail, and shuffle which letter is correct.”
• It added facts that are not in your document. Repeat the rule: “Only use the material. Delete any question you cannot back up with a quote.”
• The wrong options are silly. Ask: “Replace each wrong option with a mistake a real new employee might actually make.”
• The answer key is wrong. Fix it by hand in the form. Do not ask the AI to “double check” and trust the second answer without looking at your document.

Questions people ask

Can ChatGPT make a quiz?

Yes. ChatGPT, Claude and Gemini can all write quiz questions with an answer key from text you give them. They do not score your team’s answers, though. For that, put the questions into Google Forms or Microsoft Forms.

Can I make a quiz with AI for free?

Yes. The free plans of ChatGPT, Claude or Gemini are enough to write the questions, and both Google Forms and Microsoft Forms have a quiz mode. The built-in AI drafting inside the forms needs a paid plan: an eligible Google Workspace or Google AI plan for Gemini in Forms, and a Copilot license for Copilot in Microsoft Forms.

Can AI make a quiz from a PDF?

Yes. Upload the PDF to ChatGPT or Claude with the step 3 prompt, use the Quizzes option in Gemini Notebook, or point Gemini in Google Forms at the PDF in your Drive if your plan has it.

Should I use quiz scores to judge employees?

We would not. Keep training quizzes low-stakes: they are there to help people learn and to show you what to explain better. Microsoft also lists employment-related situations, such as evaluating performance, among the things its Copilot in Forms should not be used for (Microsoft Support).

How many questions should a work quiz have?

For a knowledge check after one topic, 5 to 10 questions is a good range: long enough to cover your two or three goals, short enough that people finish it on their phone between meetings. That is a rule of thumb from practice, not a research number.

Sources

• Google Docs Editors Help: Create and grade quizzes with Google Forms
• Google Docs Editors Help: Create a form with Gemini in Google Forms
• Microsoft Support: Create a quiz with Microsoft Forms
• Microsoft Support: Frequently asked questions about Copilot in Forms
• Gemini Notebook Help: Generate Flashcards or Quizzes
• Google blog: 3 new study features in Gemini Notebook
• Wikipedia: Testing effect
• Roediger and Karpicke (2006): Test-Enhanced Learning, Psychological Science

Your next 20 minutes

Pick the one document you explain most often. Write down what people should be able to do after reading it, run the step 3 prompt, check the answers, and put the quiz in a form. Next time someone joins, you send them the document and the quiz, and you see straight away which part still needs a better explanation.

Want more practical AI workflows like this, step by step?

Or look around first: everything we teach starts on our home page.

You have three options on the table, a deadline on Friday, and a team that keeps saying “they’re all fine, really.” Maybe it is which CRM to buy, which freelancer to hire, or which of three projects gets the budget. You have gone back and forth so often that you no longer trust your own gut.

Here is how to make a decision matrix with AI: tell ChatGPT or Claude what you are deciding and what matters to you, let it suggest the criteria and weights, then fix those weights yourself before it sees your options. In a second prompt, ask it to score each option from 1 to 5 with a one-line reason per score. Copy the table into a spreadsheet, let a formula do the math, and ask the AI which single change would flip the winner. A first matrix takes about 20 minutes, and you end up with a choice you can explain to anyone.

Below: the five steps, prompts you can copy, a worked example, the spreadsheet formula, and the two traps that make AI matrices lie to you.

What a decision matrix is, and why AI makes it easier

A decision matrix is a table that compares options against a shared set of criteria, such as cost, quality or risk (Asana, decision matrix guide). Each criterion gets a weight for how much it matters, each option gets a score per criterion, and the weighted total shows which option fits your priorities best.

The idea is old and simple. What makes people skip it is the setup: thinking of the right criteria, agreeing on weights, writing a reason next to every score. That is the part an AI assistant speeds up. It will happily suggest criteria you forgot, phrase them so they can be scored, and keep the reasons short.

What it can’t do is know what matters to you. A weight is a business judgment. If you let the AI set it quietly, it has made the most important call in the whole table. So the rule for this guide: the AI drafts, you decide the weights, a spreadsheet does the math.

What you need

• A chat assistant: ChatGPT, Claude or Gemini. The free plans are enough. New to them? Start with our ChatGPT tutorial for beginners or the Claude tutorial for beginners.
• A spreadsheet: Excel or Google Sheets. You only need one formula, and we give it to you below.
• Your facts about each option: prices, notes from demo calls, a few lines from each proposal. The more real detail you paste in, the less the AI has to guess.

Step 1: Write the decision as one neutral sentence (2 minutes)

Most lopsided matrices start with a loaded question: “Should we go with Option B, it seems cheaper?” The AI picks up your lean and builds a table that agrees with you. That is a known habit of these tools. Anthropic’s researchers found that five leading AI assistants consistently tended to tell people what they wanted to hear, a behaviour they call sycophancy (Anthropic, Towards understanding sycophancy in language models).

• Loaded: “Is HubSpot or Pipedrive better? I like Pipedrive.”
• Neutral: “Which CRM should a 4-person consultancy use to track about 200 client contacts and follow-ups, starting next month?”

Size, use and deadline are in the neutral version. Your preference is not.

Step 2: Let the AI propose criteria and weights, without the options (5 minutes)

This is the step most prompt templates skip, and it is the one that keeps the matrix honest. Ask for criteria and weights first, and do not name your options yet. Paste this:

I need to build a weighted decision matrix.

The decision: [your neutral sentence from step 1]
What matters to us, in my own words: [e.g. we are not technical, money is tight this year, everything lives in Gmail and Google Calendar, we might switch again in 2 years]

Step 1 only. Propose 4 to 6 criteria for this decision. For each one give:
- a one-line definition I can score from 1 to 5
- a weight in percent (all weights add up to 100)
- one sentence on why it deserves that weight

Do not suggest any product or option yet. Do not tell me what you would choose. Keep criteria separate, so no two of them measure the same thing.

Now read the weights like a manager, not a reader. Would you defend each one to your business partner? If “price” comes back at 40% and you know the team will quit a tool they hate, move points to ease of use. Change the numbers until they feel right, then write them down. From here on they are locked.

One check that catches most bad setups: ask “Is there any criterion here that is really a must-have, not a trade-off?” A must-have (for example “works in our country” or “meets our client’s data rules”) does not belong in the scores. It is a filter: an option that fails it is out before you start.

Step 3: Score each option, with a reason per score (5 minutes)

Now bring in the options and your facts. Open a fresh chat, or say clearly that the weights are fixed:

Here are my locked criteria and weights. Do not change them:
[paste your final list with weights]

Here are the options and what I know about each:
Option A: [your notes, prices, demo impressions]
Option B: [...]
Option C: [...]

Score every option on every criterion from 1 (poor) to 5 (excellent).
Rules:
- give a one-line reason next to every score, based only on my notes
- if my notes do not say, write UNKNOWN instead of a score and tell me what to find out
- show the result as a table: criteria in rows, options in columns, weights in their own column
- do not calculate totals and do not pick a winner

Two parts of that prompt matter most. The reasons turn a number you have to trust into a claim you can check. And the UNKNOWN rule stops the AI from quietly turning a gap in your knowledge into a low score. A missing fact is a question to answer, usually with one email to a sales rep or a look at the vendor’s help pages, not a point to deduct.

Why no totals? Chat assistants are good with words and less reliable with arithmetic. A spreadsheet never gets a multiplication wrong, so let it do that part.

Step 4: Let a spreadsheet do the math (3 minutes)

Ask the AI: “Give me this table as CSV I can paste into Google Sheets.” Paste it in so that the weights sit in column B (rows 2 to 6) and each option has its own column of scores next to it. Then, under Option A’s scores, type:

=SUMPRODUCT($B$2:$B$6,C2:C6)/100

SUMPRODUCT multiplies each weight by its score and adds the results up. It works the same way in Excel and Google Sheets. Dividing by 100 brings the total back to the 1 to 5 scale, because the weights add up to 100. Drag the formula across to the other options. If your matrix has more rows, change the 6 to your last row. Formulas giving you trouble? Our guide on how to use AI to write Excel formulas walks you through it.

Step 5: Ask what would flip the winner (5 minutes)

A total at the bottom of a column is not yet a decision. Paste the finished table back into the chat and ask:

Here is my finished decision matrix with totals: [paste]

1. Which single weight, if I moved it by 10 points, would change the winner?
2. Which single score, if it is wrong, would change the winner?
3. What should I check this week before I decide?
Argue against the current winner as hard as you honestly can.

If the answer is “nothing realistic would change it,” you have a clear winner. If one small change flips it, you have a close call, and you now know exactly which question to settle before you sign anything.

Bonus check: ask for a fresh scoring pass with the options listed in a different order. Researchers testing AI models as judges found they can favour an answer based on its position in the list (Zheng et al., Judging LLM-as-a-Judge). If the scores move when only the order changed, trust the reasons over the numbers.

A worked example: choosing a CRM for a small team

Here is what the steps produce for the consultancy from step 1. The options are kept anonymous on purpose: the method is the point, not the brands, and real scores depend on your own notes and the prices on the day you check.

After step 2 and a few edits, the locked criteria were: setup effort for a non-technical team (25%), monthly cost for 4 users (20%), Gmail and calendar connection (25%), ease of daily use (20%) and getting the data out later (10%). In step 3 the AI marked Option B’s data export as UNKNOWN. One look at B’s help pages showed it exports contacts only, not the notes, so that became a 2.

Option B wins, but only just. Step 5 showed why that matters: move 10 points from cost to the Gmail and calendar connection (35% instead of 25%, cost 10% instead of 20%) and Option A comes out on top, 4.05 against 3.80. So the real question was never “which CRM is best.” It was “how much do we live in our inbox?” For a team like this one, that is a ten-minute conversation. If almost every client conversation starts in email, A is the better pick, and the choice fits in one sentence that nobody has to win an argument over.

That is the quiet value of the method. It does not decide for you. It shows you which disagreement you are really having.

When a decision matrix helps, and when it doesn’t

It works best when you have 2 to 5 real options, several things that matter at once, and other people who need to understand the choice. Typical work examples:

• Picking software: a CRM, a booking tool, an AI voice agent platform, or an automation tool (our n8n vs Zapier comparison gives you ready-made criteria).
• Hiring or choosing a supplier: score candidates or quotes on the same criteria, so the loudest one does not win by default.
• Choosing which project goes first: impact, effort, risk and urgency as the criteria. Once you have the winner, turn it into a plan with our guide on how to use AI to create a project plan.

It helps much less when there is only one real option, when a single must-have decides everything, or when the choice is mostly about values (whether to take on a business partner, for example). Then a pros and cons list or an honest conversation does the job better.

Decision matrix, SWOT or mind map: which one do you need?

These three get mixed up, and they answer different questions:

• Mind map: “What is even on the table?” Use it first, when the problem is still a tangle. See how to make a mind map with AI.
• SWOT analysis: “What are the strengths, weaknesses, opportunities and threats of one plan?” See how to do a SWOT analysis with AI.
• Decision matrix: “Which of these options do we pick?” Use it last, when you have the options and need to choose.

A good order for a big choice: mind map to find the options, a quick SWOT on the top two if the stakes are high, and the matrix to make the call. If the chosen option is a process change, our guide on how to make a flowchart with AI helps you show the team how the new way works.

When the matrix comes out wrong

• Every option scores about the same. Your criteria are too vague (“quality”, “value”). Ask: “Rewrite each criterion so that a 1 and a 5 look clearly different for a team like ours.”
• The winner feels wrong. Take that feeling seriously. Usually a criterion is missing. Ask: “I expected Option C to win. Which criterion would I need to add for that to happen, and is it a fair one?” If it is fair, add it and re-score. If it isn’t, you just learned something about your gut.
• The AI invents facts about a product. It can describe features or prices that are out of date or wrong. Score only on what you checked yourself, and use the UNKNOWN rule from step 3.
• The scores look suspiciously like your first preference. Go back to step 1 and check your sentence for a hint of what you wanted.

Questions people ask

Can ChatGPT make a decision matrix?

Yes. It can suggest criteria, propose weights and score options with reasons, and it can give you the table as CSV for a spreadsheet. Keep the weights in your own hands and let the spreadsheet calculate the totals, as described in steps 2 and 4.

Is Claude or ChatGPT better for this?

Both handle it well, and the method matters far more than the tool. If you want to compare them for your own work, our ChatGPT vs Claude guide goes through the differences.

What scale should I use, 1 to 5 or 1 to 10?

1 to 5 is plenty for most work decisions. A wider scale invites fake precision: nobody can really tell a 7 from an 8 on “ease of use.” What matters more is that a 5 means the same level of good on every criterion.

Should I paste supplier quotes or CVs into an AI tool?

Only if your company allows it. When in doubt, leave out names, personal details and confidential figures, and describe each option in your own words. The matrix needs your judgment of the facts, not the original documents.

Sources

• Asana: What is a decision matrix? Examples and how to create one
• Anthropic: Towards understanding sycophancy in language models
• Zheng et al.: Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena
• Microsoft Support: SUMPRODUCT function
• Google Docs Editors Help: SUMPRODUCT

Your next 20 minutes

Pick the decision you have been putting off this week. Write it as one neutral sentence, run the step 2 prompt, and fix the weights before you look at a single option. You will probably see the answer forming before the spreadsheet is done.

Want more small, practical wins like this one? Join our free Launchpad.

Or look around first: everything we teach starts on our home page.

You have a new project, a workshop to plan, or a 60-page report to get your head around, and right now it is all one tangled ball in your head. You know a mind map would help. You just don’t want to spend an hour dragging bubbles around a screen before you have even started thinking.

Here is how to make a mind map with AI: write one clear question for the center, dump everything you know about it in plain words, and ask ChatGPT or Claude to sort it into an outline with main branches and sub-branches. Paste that outline into a free viewer like Markmap or the Mermaid Live Editor and it becomes a map you can fold, zoom and share. If your starting point is a document instead of your head, Google’s Gemini Notebook (the tool that used to be called NotebookLM) can build a mind map straight from your files. A first useful map takes about 15 minutes.

Below: the five steps, prompts you can copy, a worked example, how to map a PDF, and what to do when the map comes out generic.

What AI does well in a mind map (and what it can’t do)

A mind map puts one idea in the center and lets the related ideas branch out from it, with the big themes close to the center and the details further out. The term was made popular by the British author Tony Buzan, though branching diagrams like this go back centuries, and people use them for notes in meetings and planning sessions, for brainstorming and for problem solving (Wikipedia, Mind map).

The slow part was never the thinking. It was the sorting: deciding which idea belongs under which branch, spotting that two notes say the same thing, noticing the branch you forgot. AI is good at exactly that. You give it a messy pile, it hands back a tidy tree.

What it can’t do is know your situation. Ask for a mind map of “launching a new service” with no context and you get the same branches as everyone else: marketing, pricing, audience, timeline. Useful for a school essay, useless for your Tuesday. So the rule for this guide: you bring the raw material, the AI does the sorting, you make the final calls.

What you need

• A chat assistant: ChatGPT, Claude or Gemini. The free plans are enough. New to them? Start with our ChatGPT tutorial for beginners, the Claude tutorial for beginners or the Gemini tutorial for beginners.
• A free viewer that turns text into a map: Markmap reads a normal bulleted outline (Markdown) and draws it as an interactive mind map (Markmap docs). The Mermaid Live Editor does the same with Mermaid’s mind map format. Both run in the browser, nothing to install.
• Optional, for documents: Gemini Notebook, if you want a map of a report, a set of PDFs or meeting transcripts.

Why go through text at all, instead of an app that draws the picture for you? Because text is easy to change. Every fix is one prompt or one edited line, and you never lose an afternoon nudging bubbles into place.

Step 1: Write one question for the center (2 minutes)

Most weak mind maps start with a topic: “Q1 marketing”. A topic gives the AI nothing to aim at, so it returns a textbook. A question gives it a job.

• Weak: “Client workshop”
• Better: “What do I need to sort out to run a half-day workshop for 12 people at a client’s office on 3 March?”

The second version tells the AI the size, the place and the deadline. Those details are what make the branches specific to you.

Step 2: Dump what you know, then ask for the outline (5 minutes)

Type, or talk into your phone, everything that is in your head about the question. Half-sentences are fine. Worries count. “Not sure if they have a projector” is exactly the kind of line that turns into a useful branch. If you think better out loud, record a voice memo and turn it into text first; our guide on how to transcribe audio with AI shows the free options.

Then paste it into ChatGPT or Claude with this prompt:

I want a mind map for this question: [your question from step 1]

Below are my raw notes. Turn them into a mind map outline in Markdown:
- Start with "# " and a short version of the question as the center.
- Use "## " for 4 to 7 main branches.
- Under each branch, use "- " bullets for the details, max two levels deep.
- Keep every item to 2 to 6 words.
- Only use what is in my notes. If something important seems missing, do NOT add it to the map. List it at the end under "Questions for you".

My notes:
[paste your notes]

Two lines in that prompt do the heavy lifting. “Only use what is in my notes” keeps the AI from padding your map with generic branches. “Questions for you” turns its urge to fill gaps into something useful: a short list of things you may have forgotten, which you can accept or ignore.

Step 3: See it as a map (1 minute)

Copy the outline (without the questions at the end) and paste it into the left side of the Markmap editor. The map appears on the right. Click a node to fold its branch away, scroll to zoom, and drag to move around.

Prefer Mermaid? Ask the AI: “Rewrite this outline as a Mermaid mindmap”, and paste the result into the Mermaid Live Editor. Mermaid’s mind map format is simply an indented list under the word mindmap: each line indented further than the one above becomes its child (Mermaid mindmap syntax).

Working in Claude? Ask it to “show this as a mind map diagram” and it can draw it in an artifact, the panel that opens beside the chat. Artifacts are available on the free plan and Claude lists diagrams among the things it can make there (Claude Help Center, What are artifacts). If nothing appears, check that “Create and edit files” is switched on under Settings, Capabilities, as that help page explains.

Step 4: Push back once (5 minutes)

Now look at the map like a stranger would. Three things usually need fixing: a branch that is really two branches, two branches that are really one, and a detail hanging under the wrong parent. Don’t fix them by hand. Tell the AI:

Improve this mind map outline:
1. Merge any items that say the same thing.
2. Move any item that sits under the wrong branch, and tell me which ones you moved.
3. If a branch has more than 6 items, split it into two clearer branches.
4. Do not add new content.
Return the full Markdown outline again.

Then answer the “Questions for you” from step 2. Anything that matters gets added as a line in your notes, and you run the outline prompt again. Because the map is just text, a full redo takes seconds.

Step 5: Turn the map into your next move (2 minutes)

A mind map that sits in a folder did its job only halfway. The point is to see the shape of the problem, so you can act on it. Pick the one that fits:

• A plan: “Turn this mind map into a task list with an owner and a deadline for each item.” Then take it further with our guide on how to use AI to create a project plan.
• A talk or a deck: each main branch becomes a section. See how to use AI to make a presentation.
• An article or newsletter: the branches are your headings. Our guide on how to use AI to write a blog post picks up from there.
• A decision: if the map shows strengths, risks and options, a SWOT analysis with AI is the natural next step.

A worked example (made up, so you can see the flow)

Say you run a small training business and a client asked for a half-day workshop on using AI for email. Your notes, typed in two minutes, look like this: “12 people, their office, 3 March, 9 to 1. Need laptops? Ask IT if ChatGPT is blocked. Want them to leave with a working template. Coffee break. Handout. Price agreed already. Follow-up email after. Not sure about the projector. Do I need examples from their own inbox? Privacy issue.”

The outline prompt from step 2 returns something like this:

# AI email workshop, 3 March
## Logistics
- 12 people, client office
- 9:00 to 13:00
- Coffee break
- Projector to confirm
## Tech check
- Laptops for everyone?
- Ask IT: is ChatGPT allowed?
## Content
- Goal: working email template
- Practice examples
## Privacy
- Real inbox examples?
- Use made-up emails instead
## Materials
- Handout
## After the session
- Follow-up email
## Questions for you
- Who is your contact at IT?
- Is there a backup if the AI tools are blocked on the day?

One look at the map and two things jump out that the notes hid. The whole workshop depends on the “Tech check” branch, so that email to IT goes out today, not next week. And the “Privacy” branch quietly decides the “Content” branch: if you can’t use real emails, you need to write practice emails before anything else. That is the value of a mind map. It doesn’t add ideas, it shows you which ones are holding the others up.

How to make a mind map from a PDF or report

When the raw material is a document, not your head, you have two good routes.

Gemini Notebook. Google renamed NotebookLM to Gemini Notebook in July 2026; it is the same product (Google, NotebookLM is now Gemini Notebook). Create a notebook, upload your sources, and choose the Mind Map option in the chat. The map appears as a note in the Studio panel. You can expand and collapse branches, click a node to ask about that topic, and download the map. It is not available in the mobile app, and Google itself warns the output can be inaccurate, so check anything you plan to rely on (Gemini Notebook Help, Use Mind Maps).

Claude or ChatGPT with the outline prompt. Upload the PDF and use the step 2 prompt, but change “my notes” to “this document” and add: “After each item, put the page number it comes from.” The page numbers make checking fast. For long files, our guides on turning a PDF into notes with Claude and analyzing long PDF files with Claude cover the limits and the tricks.

Which one to pick? Gemini Notebook gives you a clickable map without any copy-pasting. The outline route gives you a map you can edit line by line and reuse anywhere. For a one-off overview, the first. For something you will build a plan on, the second.

Copy-paste mind map prompts for ChatGPT and Claude

Three more prompts for the jobs people use mind maps for most.

Brainstorm, without the generic answers:

Ask me 5 short questions, one at a time, about [goal]. Wait for each answer. Then build a Markdown mind map outline from my answers only, with 4 to 6 main branches and items of max 6 words.

Meeting notes into a map:

Turn these meeting notes into a Markdown mind map outline. Center: the meeting's main decision. Branches: decisions, open questions, action items (with owner), risks. Leave out small talk. Notes: [paste]

If you mostly need the action list, our guide on using ChatGPT to summarize meeting notes is the faster route.

Learn a new topic fast:

I need to understand [topic] well enough for a meeting with [who] about [what]. Make a Markdown mind map outline of what I need to know, max 5 branches. Mark anything you are unsure about with (check).

A note on the last one: when the AI writes the content itself instead of sorting yours, treat every item as a lead to check, not a fact. The “(check)” tag makes the shaky ones easy to spot.

Mind map or flowchart?

They look alike, but they answer different questions. A mind map answers “what is in this?”: the parts of a topic and how they group. A flowchart answers “what happens next?”: steps in order, with decisions. If your map starts to grow arrows like “then” and “after that”, you are describing a process. Switch to our guide on how to make a flowchart with AI, which uses the same text-first approach.

When the map comes out wrong: quick fixes

• It’s generic. Your notes were too thin, or the AI added its own ideas. Add five specific lines to the notes and keep the “only use what is in my notes” line in the prompt.
• Too many branches. More than seven main branches and nobody sees the shape any more. Ask: “Group these branches into max 5 themes.”
• Items are full sentences. Ask for “max 6 words per item, keep the full sentences in a separate list underneath”. The detail is still there when you need it.
• The Mermaid version looks odd. In Mermaid mind maps, brackets around text set the shape of a node, and the levels come from the indentation (Mermaid docs). So a label like “Budget (Q3)” or a line indented by one space too few can come out strangely. Ask the AI: “Remove all brackets from the labels and use consistent indentation of 2 spaces.”
• Markmap shows only one line. The outline lost its “#” and “-” markers when you copied it. Ask the AI to put the outline in a separate copy box and use the copy button on that box.

Questions people ask

Can ChatGPT make mind maps?

Yes, in the sense that matters: it writes the structure, and a free viewer draws it. Ask for a Markdown or Mermaid mind map outline, then paste it into Markmap or the Mermaid Live Editor. Keeping the map as text also means you can change it with a sentence later.

Is there a free AI mind map generator?

The route in this guide is free from start to finish: the free plan of ChatGPT, Claude or Gemini for the outline, plus Markmap or the Mermaid Live Editor to view it. Gemini Notebook also has a mind map feature with usage limits per plan (Gemini Notebook Help).

Can Claude create mind maps?

Yes. Ask for a mind map diagram and it can draw one in an artifact next to the chat, or give you the Markdown or Mermaid text to use elsewhere.

Should I paste client documents into an AI tool?

Only if your company or client allows it. When in doubt, make the map from your own notes with names and figures left out, as in the workshop example. The structure is what you need from the map, not the confidential details.

Sources

• Wikipedia: Mind map
• Markmap docs: What is markmap?
• Mermaid: Mindmap syntax
• Gemini Notebook Help: Use Mind Maps in Gemini Notebook
• Google: NotebookLM is now Gemini Notebook
• Claude Help Center: What are artifacts and how do I use them?

Your next 15 minutes

Pick the one thing on your plate that feels most tangled. Write the question, dump your notes, run the step 2 prompt and paste the outline into Markmap. Then look for the branch that holds the others up, and do that first.

Want more small, practical wins like this one? Join our free Launchpad.

Or look around first: everything we teach starts on our home page.

Your boss wants “a quick SWOT” for Thursday. Or you are about to raise prices, hire someone, or launch a new service, and you want to think it through properly before you bet money on it. You open a blank 2×2 grid, type “Strengths” in the corner, and stare at it.

Here is how to do a SWOT analysis with AI without getting generic filler: give ChatGPT or Claude one clear decision and your real material (numbers, customer comments, competitor pages), let it interview you before it writes anything, have it sort what you know into the four boxes with the evidence next to each point, then make it argue against its own list. Finish by turning the grid into three actions with a TOWS matrix. A solid first version takes about 45 minutes, and most of that time is you adding facts the AI cannot know.

Below: the five steps, prompts you can copy, a worked example, a version for your own career, and the mistakes that make an AI SWOT useless.

Why most AI SWOT analyses are useless (and how to fix that)

Type “do a SWOT analysis for my bakery” into any chatbot and you get a neat grid in seconds. Strengths: “quality products, loyal customers.” Threats: “competition, economic uncertainty.” It looks finished. It would also fit every bakery on the planet.

That is not the AI being lazy. It simply knows nothing about your bakery, so it fills the boxes with the most likely words. And it says them with confidence. OpenAI’s own help page warns that ChatGPT can sound sure of itself even when it is wrong, and can oversimplify complex issues (OpenAI Help, Does ChatGPT tell the truth?).

MindTools lists the classic SWOT mistakes: lists that are too long, points that are vague, not seeing your own weaknesses, and treating the SWOT as the whole plan (MindTools, SWOT Analysis). A chatbot left on its own makes every one of them.

The fix is a change of roles. You bring the facts. The AI brings the structure, the questions you forgot to ask, and the critic you do not have time to find. That split is what the steps below are built on.

What you need before you start

• A chat assistant. ChatGPT, Claude or Gemini all work. The free plans are enough for a first SWOT. New to them? Our ChatGPT tutorial for beginners and Claude tutorial for beginners get you set up.
• One decision. Not “our business” in general. Something like “should we add a lunch menu” or “should I take on retainer clients”.
• Your raw material. Whatever you have: last year’s sales by month, a handful of customer reviews, survey answers, a competitor’s pricing page, notes from a team meeting. Messy is fine.
• 45 minutes and, if you can, one colleague who will tell you the truth.

How to do a SWOT analysis with AI in 5 steps

Step 1: Name the one decision (5 minutes)

A SWOT about “the company” turns into a list of everything. A SWOT about one decision gives you something to act on. MindTools makes the same point: apply it at the right level, such as a product or product line, rather than the vague whole-company level.

Write one sentence: “We are deciding whether to ___ in the next ___ months.” That sentence goes at the top of every prompt.

Step 2: Let the AI interview you first (10 minutes)

This is the step that kills the generic grid. Instead of asking for the SWOT, ask for questions.

Answer in plain words. Voice typing helps if you think better out loud. Paste in the material you gathered too: reviews, a survey export, a spreadsheet summary. If you have a long report or PDF, our guide on analyzing long PDF files with Claude shows how to feed it in without losing the details.

Step 3: Build the grid with evidence next to every point (10 minutes)

Now ask for the SWOT, with one rule: every point has to say where it came from.

The [ASSUMPTION] tag is the useful part. It shows you, at a glance, which points are facts and which are guesses dressed up as facts. MindTools gives a good test for this: accept precise, checkable statements like a specific cost advantage, not “better value for money”.

A note on Opportunities and Threats: these come from outside, so the AI’s training data is often out of date. If your assistant has web search turned on, ask it to look up current information and show the links, then open the links yourself. OpenAI’s advice is to use search when accuracy matters and to verify by visiting the sources. For a deeper look at a market, follow our guide on how to use AI for market research first and bring the results here.

Step 4: Make the AI argue with itself (10 minutes)

Most teams are too kind to their own strengths and too vague about their weaknesses. A chatbot has no feelings to protect, so give it the job nobody in the meeting wants.

That first question is straight from MindTools: a feature is only a strength if it gives you a clear advantage. If all your competitors also deliver on time, on-time delivery is the entry ticket, not a strength.

Then do the human part. Show the grid to one colleague, or a customer you trust, and ask: “What’s wrong with this?” The AI can challenge your list. Only people who see your business from the outside can tell you what it actually feels like.

Step 5: Turn the grid into actions with a TOWS matrix (10 minutes)

A SWOT is a snapshot. On its own it does not tell you what to do, which is why so many end up in an appendix. The TOWS matrix, developed by Heinz Weihrich in 1982, fixes that by crossing the boxes (MindTools, The TOWS Matrix):

• SO: use a strength to grab an opportunity.
• ST: use a strength to blunt a threat.
• WO: fix a weakness so you can take an opportunity.
• WT: defensive moves that shrink a weakness and avoid a threat. MindTools notes these protect you from loss but will not create success on their own.

The labels keep the AI honest: an action that cannot point at a box came from nowhere. You end with three actions, not sixteen bullet points. If one of them is a bigger project, our guide on using AI to create a project plan takes it from there.

A worked example: a two-person design studio

This example is made up to show the process. The numbers are illustrative, not data.

Decision: “Should we offer monthly retainer packages instead of only one-off projects, starting in January?”

In Step 2, the AI asked eight questions. The ones that changed the picture: “How many past clients came back for a second project?” (answer: about half), “What do clients complain about?” (slow replies when both designers are busy), and “Do competitors near you offer retainers?” (they did not know, which became an open question).

After Step 4, the grid looked like this:

• S1 About half of past clients returned for a second project (studio’s own records).
• S2 Clients praise the brand workshops in reviews (12 Google reviews mention them).
• W1 Reply times slip to several days in busy weeks (client complaints).
• W2 No written process, so work depends on the two founders [critic added this].
• O1 Several returning clients asked for “someone on call” for small updates (emails).
• T1 Clients can now make simple social graphics themselves with AI tools [ASSUMPTION, check with clients].

The critic moved “creative, high-quality design” out of Strengths, because every studio says that. It also flagged that T1 rested on a guess, so the studio asked three clients. Two said yes, they now do small graphics themselves.

The three TOWS actions they kept:

1. SO (S1 + O1): offer a small retainer to the returning clients who asked for it, before building a public package.
2. WO (W1 + O1): a retainer includes a reply-time promise, so they first write a simple support process with AI (our guide on writing SOPs with AI covers this).
3. ST (S2 + T1): lead the retainer with strategy and workshops, the part clients cannot do with a template tool.

Total time at the keyboard: under an hour. The most valuable minute was the critic pointing out the missing weakness.

Personal SWOT with AI: for your career or a job change

The same method works for yourself. MindTools describes a personal SWOT for career decisions: what you do better than most, what skills you lack, what is changing in your industry, and what puts your role at risk (MindTools, Personal SWOT Analysis).

Your evidence here is your own history: performance reviews, feedback, projects you shipped. Use the same prompts, swap “us” for “me”, and change the decision line to something like “Should I apply for the team lead role?” One extra instruction helps a lot:

The result feeds straight into a self-evaluation or your interview preparation.

SWOT your competitor (carefully)

MindTools also suggests running a SWOT on your competitors. AI makes that faster: paste in their pricing page, their public reviews and their job ads, and ask for their SWOT with evidence tags. The US Small Business Administration lists what a competitive analysis should look at, including competitors’ strengths and weaknesses, your window of opportunity, and barriers to entering the market (SBA, Market research and competitive analysis).

Two cautions. Only use public information. And treat everything about a competitor’s inside (their costs, their team, their plans) as [ASSUMPTION] unless you have a real source for it.

Mistakes that make an AI SWOT worthless

• Asking for the SWOT in your first prompt. You get the generic version. Interview first.
• Accepting points without evidence. If a box is full of [ASSUMPTION] tags, you have a to-do list for research, not an analysis.
• Letting the boxes grow. Five points each, maximum. A longer list just hides the points that matter.
• Mixing internal and external. “A new competitor opened nearby” is a threat, not a weakness. The critic step catches most of these.
• Stopping at the grid. No TOWS step, no actions. The grid alone changes nothing on Monday morning.
• Pasting in confidential numbers without checking your company’s AI policy. If you are not sure, round or anonymize them. The analysis works just as well with “revenue down about 10%” as with the exact figure.

FAQ

Can ChatGPT do a SWOT analysis?

Yes, and quickly. But on its own it can only produce a generic one, because it does not know your business. Give it your facts and one decision, ask it to interview you first, and it becomes genuinely useful. Check anything it claims about the outside world, because it can be wrong while sounding certain.

Is there a free AI SWOT analysis generator?

The free plans of ChatGPT, Claude and Gemini work as a SWOT generator with the prompts above. Dedicated generator websites mostly do the same thing with a form in front. The quality depends on the facts you put in, not on the tool.

What is the best prompt for a SWOT analysis?

The one that forbids guessing. The key lines are: one decision at the top, “ask me questions first”, max 5 points per box, evidence after each point, and [ASSUMPTION] when there is none. The Step 3 prompt above has all of them.

How do I present the SWOT?

One slide with the grid, one slide with the three TOWS actions. Ask the AI to cut each point to one line for the slide. Our guide on making a presentation with AI covers the rest.

How often should I redo it?

When the decision changes or the market does. Because you saved the prompts and the evidence, updating it later takes minutes, not a new workshop.

If your SWOT leaves you with two or three options and no clear winner, score them side by side with a decision matrix built with AI.

Your next step

Pick one decision you are sitting on right now. Write the one sentence, paste the Step 2 prompt, and answer the questions. In 45 minutes you will have a SWOT that is actually about you, and three things to do next week.

Want more walk-throughs like this, where AI does the busywork and you stay in charge? Launchpad is free.

Or start on our home page and see what else you can hand off to AI this week.

Someone asks you how the invoice approval actually works, and you start talking: “Well, first it comes in, then Sam checks it, unless it’s over a certain amount, then it goes to…” Thirty seconds in, everyone is lost. You know the process. You just can’t draw it, and an hour of dragging boxes around in a diagram tool feels like a waste of a morning.

Here is how to make a flowchart with AI: describe the process in plain words (typing or a voice memo), ask ChatGPT or Claude to turn it into a flowchart written in Mermaid, a simple text format for diagrams, then paste that text into a free tool like Mermaid Live Editor or draw.io to see the chart. Claude can also draw it right in the chat. Check every box against how the work really happens, fix it with one more prompt, and export it as an image. A first version takes about 15 minutes.

Below are the five steps, prompts you can copy, a worked example, and the fixes for the errors you will hit the first time.

Why AI is good at flowcharts (and where it needs you)

A flowchart is a map of a process: steps in boxes, decisions in diamonds, arrows showing what happens next. The American Society for Quality describes it as a tool to understand, communicate, document and improve how work gets done, and lists it as one of the seven basic quality tools (ASQ, What is a flowchart).

The slow part was never the thinking. It was the drawing: placing boxes, lining up arrows, redoing everything when you remember a step you missed. That is exactly the part AI takes off your plate. You talk, it structures.

What it cannot do is know your process. It will happily invent a “manager approval” step that does not exist, or skip the awkward exception that causes half your delays. So the rule for this whole guide is simple: AI drafts the chart, the people who do the work check it.

What you need

• A chat assistant: ChatGPT or Claude. The free plans are enough for this. New to them? Start with our ChatGPT tutorial for beginners or the Claude tutorial for beginners.
• A place to see the chart: Mermaid Live Editor (free, in the browser, nothing to install) or draw.io (free, also in the browser). Or Claude itself, which can show the chart beside your conversation.
• The process in your head, or better, in the head of the person who does it every day.

Step 1: Get the process out of your head (5 minutes)

Do not try to be tidy here. The messier and more honest, the better the chart. Two ways that work:

• Type it as a list. One line per step, in the order it happens. Add the “unless” and “if” bits as you remember them.
• Talk it through. Record a two-minute voice memo while you walk through the process, then turn it into text. Our guide on how to transcribe audio with AI shows the free ways to do that.

Before you move on, decide two things, because they are the questions that make a flowchart useful (ASQ’s procedure starts with the same ones): where does the process start, and where does it end? “From the moment an invoice lands in the inbox until it is paid and filed” is a scope. “Invoices” is not.

Step 2: Ask the AI for a flowchart in Mermaid (2 minutes)

Mermaid is a way to describe a diagram in plain text. The Mermaid project says it lets non-programmers create detailed diagrams, and the makers of draw.io note that most AI tools can write it (draw.io, Insert and edit a Mermaid diagram). That is the trick: you do not ask the AI to draw a picture, you ask it to write the diagram. Text is easy to fix, a picture from an image generator is not, and small labels in generated images often come out garbled.

Paste this into ChatGPT or Claude, with your notes from step 1 at the bottom:

That last line about questions matters most. In my experience the questions the AI asks back (“Who approves when the team lead is on holiday?”) are often worth more than the first chart, because they point at the gaps nobody wrote down.

Step 3: Turn the text into a picture (3 minutes)

The AI gives you a block of text that starts with flowchart TD. Pick one of these three ways to see it:

Option A: Mermaid Live Editor (fastest)

Open Mermaid Live Editor, delete the sample in the left panel and paste your text. The chart appears on the right and updates as you edit. You can download it as an image when you are happy with it.

Option B: draw.io (best if you want to style it by hand)

In draw.io, choose Arrange > Insert > Mermaid (or the + icon in the toolbar, then Mermaid), paste the text and press Insert. draw.io turns it into normal shapes you can recolour, move and restyle, and it keeps the Mermaid source so you can edit and regenerate later (draw.io help). One catch from that same page: if you edit the source and regenerate, your colours and labels stay, but any boxes you dragged into a new position are laid out again.

Option C: let Claude draw it in the chat

Claude can show diagrams and flowcharts as an artifact, a window beside your conversation that you can edit by asking for changes, and artifacts are available on the free plan (Claude Help, What are artifacts). Just add “and show it to me as a diagram” to the prompt. Handy for a quick look. For a chart you will keep and update, I still save the Mermaid text somewhere, so the next person can edit it without starting over.

Step 4: Check it against reality (5 minutes, the step everyone skips)

This is where an AI flowchart becomes a correct one. ASQ’s advice is to have the people who perform the process draw or review it, chart how the work happens now (not how it should happen), and walk through the process after the first draft to make sure it is accurate (ASQ). With AI the drawing is cheap, so spend the saved time here.

Walk a real, recent case through the chart with your finger. Then look for these:

• Invented steps. Anything you did not mention? Delete it or confirm it.
• Dead ends. Every box except the end needs an arrow going out.
• Decisions with one exit. A diamond with only a “Yes” arrow means the “No” path is missing.
• The real exceptions. The supplier who never sends a PO number, the approver who is away. Those loops are usually where the time goes.

Then fix it with a follow-up prompt instead of editing by hand:

Step 5: Clean it up and share it

• Left to right or top to bottom? Long processes read better top to bottom (flowchart TD) on a page. Wide screens and slides suit left to right (flowchart LR). The Mermaid flowchart docs list all the directions.
• Who does what? Ask for a swimlane version: “Group the steps per role using subgraphs, one for Sales, one for Finance, one for the Customer.” Each lane becomes a labelled box around that role’s steps.
• Export as PNG for email and chat, or SVG if it goes into a document or slide and must stay sharp. Need it in a deck? Our guide on how to use AI to make a presentation covers the rest.
• Save the Mermaid text next to the image (in the same folder, or at the bottom of the document). Next time the process changes, you paste it back into the AI with the change and get a new chart in a minute.

Worked example: an invoice approval flow

This is a made-up example to show the shape of it. The notes someone typed:

What came back from the prompt in step 2, pasted into Mermaid Live Editor:

Look at those two questions. Neither was in the notes, and both are real gaps. One more prompt with the answers (“rejected invoices go back to Sam, who contacts the supplier”) and the chart matches how the team actually works. The whole thing took less time than finding the right arrow shape in a diagram tool.

Fixing the errors you will hit the first time

Mermaid is strict, so now and then the editor shows an error instead of a chart. Usually it is one of these:

• The word “end” in lowercase. The Mermaid docs warn that typing “end” in all lowercase inside a flowchart breaks it. Write “End” or “END” (Mermaid flowchart docs).
• Brackets, commas or symbols in a label. Put the whole label in double quotes, like A["Check PO (if any)"]. The prompt above already asks for this.
• The AI added extra text. Sometimes the reply has a sentence or the word “mermaid” above the chart. Paste only from flowchart down.
• Still stuck? Paste the error message back into the chat: “Mermaid shows this error, fix it and give me the full text again.”

And one that is not an error but still ruins charts: too many boxes. If your chart has more than about 15 to 20 steps, nobody will read it. Ask the AI to split it: “Make a high-level version with at most 8 steps, and a separate detailed chart for step 4.”

Which AI flowchart tool should you use?

• You want it free and fast: ChatGPT or Claude plus Mermaid Live Editor. No account needed for the editor.
• You want it polished, with your colours: ChatGPT or Claude plus draw.io. You get real shapes to restyle.
• You want to talk it through and see it at once: Claude with an artifact.
• Your company already pays for a whiteboard or diagram app: many of them now have their own AI features. Check what your plan includes before buying anything new. The Mermaid route above still works as a free backup.

The “best AI flowchart generator” is mostly the one that lets you fix mistakes quickly. That is why I like the text route: one sentence of feedback, new chart.

What to do with your flowchart next

A flowchart is rarely the final goal. It is the first draft of something more useful:

• A written procedure. Paste the Mermaid text into the chat and ask for a step-by-step procedure based on it. Our guide on how to use AI to write SOPs takes it from there.
• Training a new hire. A one-page chart of “how a request flows through the team” helps someone in their first week more than a long document. See how to use AI for employee onboarding.
• A project plan. Map the phases and decision points first, then turn them into tasks with AI for project plans.
• Finding what to automate. Once a process is on one page, the copy-paste steps and the waiting loops stand out. Those are your first automation candidates. Start with what to automate first, and when you build it, the chart becomes the blueprint for your workflow in a tool like n8n (see the n8n tutorial for beginners).
• A mind map. If your notes are still a loose pile of ideas rather than steps in order, a flowchart is the wrong shape. Start with how to make a mind map with AI to sort them into topics first.

Questions people ask

Can ChatGPT make a flowchart?

Yes. The most reliable way is to ask ChatGPT for the flowchart in Mermaid format and paste that text into Mermaid Live Editor or draw.io. You get a clean chart you can edit, instead of a picture you cannot change.

Is there a free AI flowchart generator?

The free route in this guide costs nothing: the free plan of ChatGPT or Claude to write the chart, and Mermaid Live Editor or draw.io to show and export it. Both editors run in the browser.

Can AI make a flowchart from a document?

Yes. Upload the policy, procedure or email thread and ask: “Read this and turn the process it describes into a Mermaid flowchart. Only use steps that are in the document, and list anything unclear as questions.” Then check it the same way as in step 4, because documents often describe the process as it should be, not as it is.

What is a good prompt for a flowchart?

Give the AI the scope (where it starts and ends), the steps in your own words, the format (Mermaid, top to bottom, diamonds for decisions) and the instruction not to invent steps. The prompt in step 2 does all four.

Can I make a swimlane diagram with AI?

Yes. Tell it which roles are involved and ask it to group each role’s steps in its own subgraph. It works well for handoffs between teams, which is where most delays hide.

Sources

• ASQ: What is a flowchart? (procedure and considerations)
• Mermaid: About Mermaid
• Mermaid: Flowchart syntax
• draw.io: Insert and edit a Mermaid diagram
• Claude Help Center: What are artifacts and how do I use them?

Your next 15 minutes

Pick the one process people keep asking you about. Type it out or talk it into your phone, paste it into the prompt from step 2, and look at the chart. Then show it to the person who does the work and ask, “Is this right?” That conversation alone is worth the 15 minutes.

Want more small, practical wins like this one? Join our free Launchpad.

Or look around first: everything we teach starts on our home page.

You know your business inside out. You answer the same customer questions every week. But the blog has been empty since spring, because every time you sit down to write, the blank page wins and the afternoon is gone.

Here is how to use AI to write a blog post that is actually yours: you bring the knowledge, the AI does the heavy lifting on structure and first drafts. Pick one real question a customer asked you. Talk or type your answer for ten minutes, messy is fine. Give that raw material to ChatGPT, Claude or Gemini and ask for an outline first, not an article. Fix the outline, then draft one section at a time. Add the example only you could know, check every fact, read it out loud, and publish. The AI saves you the blank page. You keep the judgement.

Plan on about 90 minutes for your first post and less once you have your voice notes set up. Below are the seven steps, prompts you can copy, a worked example and the checks that keep a made-up fact off your website.

Can you use AI to write a blog post?

Yes, and Google says so directly. In its guidance on AI content, Google Search writes that “appropriate use of AI or automation is not against our guidelines”, as long as it is not used mainly to manipulate rankings. It also says plainly that using AI “doesn’t give content any special gains. It’s just content.” (Google Search Central, AI-generated content).

So the question is not “AI or no AI”. The question is whether the post is useful, original and accurate. Google’s own page on helpful, people-first content asks things like: does this show first-hand experience, and would a reader leave feeling they learned enough? A generic article that any chatbot could write for any company fails that test. A post built on your customers’ questions and your own experience passes it.

That is why the order below matters. Your knowledge goes in first. The AI comes second.

Step 1: Pick one question a real customer asked (5 minutes)

The best blog topics for a small business are sitting in your inbox. Scroll back through the last month of emails, chat messages or call notes and write down the questions people asked more than once.

• “How long does it take?”
• “What do I need to prepare before we start?”
• “Is it worth doing X, or should I do Y?”
• “Why does it cost more than the cheap option?”

Pick one. One question, one post. If you try to cover three questions, you get three thin answers and a post nobody finishes. If you already know which search words you want to target, our guide to ChatGPT prompts for an SEO content brief shows how to plan the keyword and the gaps before you write.

Step 2: Dump what you know, in your own words (10 minutes)

This is the step most people skip, and it is the one that makes the post yours. Before you open the AI, answer the question the way you would if a customer were standing in front of you.

Easiest way: record a voice memo on your phone while you walk or drive (parked). Talk for ten minutes. Include the story of a customer who got it wrong, the thing you always warn people about, and the tip you give that nobody else seems to give. Then turn the recording into text. Our guide on how to transcribe audio with AI covers free and paid ways to do that.

If you prefer typing, bullet points are fine. Spelling does not matter. What matters is that the facts, the opinions and the examples come from you.

Step 3: Ask for an outline, not an article

Here is where most AI blog posts go wrong. People type “write a blog post about X” and get 800 words of polite filler that could belong to any company in the world.

Ask for the skeleton first. Paste your notes and use a prompt like this:

You are helping me write a blog post for my business.

About me: [one line: what you do, for whom, where]
Reader: [who searches this and what they are worried about]
The question this post answers: [the exact customer question]

Below are my raw notes. Use ONLY the facts and opinions in my notes.
Do not add statistics, studies or claims I did not give you.
If something important seems missing, list it as a question for me instead of inventing it.

Give me:
1. Three possible titles that use the reader's own words
2. An outline with 5 to 7 section headings
3. Under each heading, the points from my notes that belong there
4. A list of questions for me where my notes are thin

My notes:
[paste your notes or transcript]

Read the outline like a customer would. Is the answer to their question near the top? Is anything in there that you would never actually say? Delete it, move things around, and answer the “questions for me” list in a few lines. That ten-minute edit is worth more than any prompt trick.

If your notes are too messy to outline yet, sort them visually first. Our guide on how to make a mind map with AI shows how to group a brain dump into branches, and each main branch can become a section heading.

Step 4: Draft one section at a time

Long drafts produced in one go tend to repeat themselves and drift away from your notes. Section by section keeps the AI close to what you actually said, and keeps you reading along instead of skimming at the end.

Now write only the section "[heading]" from our outline.

Rules:
- Write like I talk to a customer: short sentences, plain words, "you" and "I".
- Use the points from my notes for this section. Keep my examples.
- No stock filler phrases like "in today's fast-paced world".
- 150 to 300 words.
- If you need a fact that is not in my notes, write [CHECK: what you need] instead of guessing.

The [CHECK] markers are the trick. They show you exactly where the AI wanted to fill a gap, so you can fill it with a real fact or cut the sentence.

Tip: set this up once so you do not repeat yourself every time. In Claude you can create a Project, upload a few of your past emails or posts and add instructions about your tone. Claude’s help centre explains that projects give each workspace its own knowledge base and instructions, and that free accounts can create up to five (Claude Help: What are projects?). ChatGPT has projects and custom instructions that do a similar job. Our AI writing assistant guide walks through building a one-page voice brief you can reuse.

Step 5: Add what only you could write

Go through the draft and look for the places where a reader would think “says who?”. That is where your experience goes.

• A real example. “Last March a customer brought in…” beats “Many customers find…”. Change names and details if you need to protect privacy.
• Your opinion. If you recommend one option over another, say so and say why. Hedged both-sides writing is what makes AI text feel empty.
• The warning nobody else gives. The mistake you see every week, the thing that costs people money.
• Your local or industry detail. Prices in your area, the rules in your trade, the season when it matters.

This is also what Google calls experience. Its helpful content guidance asks whether content clearly shows first-hand expertise, such as having actually used a product or visited a place. An AI on its own cannot do that part. You can.

Step 6: Check every fact before you publish

AI tools can sound very sure of things that are wrong. OpenAI’s own help centre says ChatGPT can produce “fabricated quotes, studies, citations or references to non-existent sources”, and advises using it “as a first draft, not a final source” (OpenAI Help: Does ChatGPT tell the truth?). Google gives the same advice for anything you put on a website: fact-check and review all AI-generated content before publishing, including the meta description and image alt text (Google Search guidance on generative AI content).

A simple check that takes ten minutes:

1. Search the draft for every number, date, name, price and quote.
2. For each one, ask: did this come from my notes, or from the AI? If it came from the AI, find a real source or delete it.
3. Open every link the AI suggested. If the page does not exist or says something else, remove it.
4. Resolve every [CHECK] marker.

You can also ask the AI to help with the hunt:

Go through this draft and list every factual claim: numbers, dates, names, prices, quotes, rules.
For each one, say whether it appears in my original notes (yes / no).
Do not fix anything. Just give me the list.

My notes:
[paste notes]

Draft:
[paste draft]

Then you do the checking. The AI finds the claims. You decide which ones are true.

Step 7: Read it out loud, then publish

Read the whole post out loud once. Anything you would never say to a customer, change it. Watch for the usual tells: the same sentence length over and over, three-item lists everywhere, and endings that summarise what you just said.

Then finish the parts around the text:

• Title: the customer’s question in their words, close to the front. Google’s title link guidance recommends titles that are descriptive and concise.
• Meta description: two sentences that say what the reader gets. Ask the AI for three options and pick one, then check it is true.
• One clear next step: book a call, get a quote, read a related post. Just one.
• Images: a real photo of your work beats a stock image. If you use AI images, our guide to the best AI tool for image generation compares the options by job.

A worked example (made up, so you can see the flow)

This is an invented example to show the steps, not a real customer.

Sam runs a small bike repair shop. Customers keep asking: “How often should I service my commuter bike?”

• Step 1: Sam picks that question.
• Step 2: On the walk home, Sam records nine minutes: the customer whose chain snapped on a hill because it had never been oiled, the rule of thumb Sam gives people who ride daily in the rain, the three things anyone can check at home, and why a cheap tune-up in spring saves a bigger bill in autumn.
• Step 3: The AI turns the transcript into a six-part outline and asks two questions: “Do you service e-bikes?” and “What does a basic service include in your shop?” Sam answers both in two lines.
• Step 4: Sam drafts section by section. One section comes back with “[CHECK: typical chain lifespan]”. Sam replaces it with what the shop actually sees, framed as experience, not a universal number.
• Step 5: Sam adds the snapped chain story and an honest line: “If you ride less than once a week, you probably do not need a full service every season.”
• Step 6: The fact list shows one claim not in the notes, a statement about a tyre pressure rule. Sam cannot find a source worth trusting, so the line goes.
• Step 7: Read out loud, two stiff sentences rewritten, title set to “How Often Should You Service a Commuter Bike?”, published with a link to book a service.

Total time: a bit over an hour. The post sounds like Sam because almost every point came from Sam.

Which AI tool should a small business owner use for blog posts?

Start with the general assistant you already use. ChatGPT, Claude and Gemini can all outline, draft and edit a blog post well when you feed them good notes. If you have not picked one yet, our ChatGPT tutorial for beginners and Claude tutorial for beginners get you set up in one sitting.

Dedicated “AI blog post generator” tools exist too. Some add templates, SEO scoring or a direct link to WordPress. They can be handy at volume, but none of them knows your customers. The steps above matter more than the tool. If a tool promises full articles from a single keyword, be careful: Google warns that using generative AI “to generate many pages without adding value for users may violate” its spam policy on scaled content abuse (Google Search guidance, see also the spam policies).

Mistakes that make an AI blog post obvious

• Starting with “write a blog post about X”. You get the average of the internet. Start with your notes.
• Letting it invent numbers. “Studies show most people…” with no link is a red flag for readers and a risk for you.
• Publishing the first draft. The first draft is raw material. The edit is where it becomes yours.
• Covering everything. A post that answers one question well beats one that lists ten tips nobody remembers.
• Posting ten a day. Volume without value is exactly what Google’s scaled content policy targets. One good post a week, built from real questions, is a steady and safe pace.

Turn one post into more

Once the post is done, you already have the hard part. Ask the AI to pull three short LinkedIn posts from it, or a paragraph for your next email to customers. Our guides on writing LinkedIn posts with AI and writing a newsletter with AI show how to do that without sounding like a robot. And if you want the publishing and sharing to happen by itself, AI automation for small business shows where to start.

If you want the fuller weekly system, one idea turned into a main piece, then into posts, an email and a short script, see our guide to AI for content creation. It starts from the same raw notes you used for the blog post, then shows how to cut that one piece into a week of content without writing everything from scratch.

Questions people ask

Will Google penalise a blog post written with AI?

Not for using AI as such. Google’s position is that it rewards helpful, original content however it is produced, and that AI used mainly to manipulate rankings is spam (Google Search Central). A post built on your own knowledge and checked for accuracy is on the right side of that line.

Do I have to tell readers I used AI?

Google suggests adding a note about how content was made where readers would reasonably wonder “How was this created?”, and advises against listing AI as the author. For a post where you supplied the knowledge and edited the draft, many businesses simply keep their own name on it. If AI did most of the writing, a short line explaining how you used it is honest and costs you nothing.

Who owns a blog post written with AI?

In the US, the Copyright Office concluded in January 2025 that AI output can be protected only where a human determined enough of the expressive elements, for example through creative arrangement or changes to the output, and that prompts alone are not enough. Using AI to assist does not rule out protection (U.S. Copyright Office, Part 2 report). Other countries have their own rules. The practical takeaway: the more of the post is genuinely yours, the stronger your position. For anything important, ask a lawyer where you are.

How long should the blog post be?

As long as it takes to answer the question properly, and no longer. Do not ask the AI to stretch it to a word count. Padding is easy to spot and it makes readers leave.

Can I use AI to write a blog post for free?

Yes. The free versions of ChatGPT, Claude and Gemini can handle every step in this guide. Paid plans mainly give you more usage, which helps if you write often.

If you are still picking between those free tools, see ChatGPT vs Claude vs Gemini for which one fits your writing work.

The same interview-first approach works beyond the blog. If your home page or service pages still sound generic, see how to write website copy with AI, step by step with prompts.

Your next step

Open your inbox, find the question you answered most often this month, and record ten minutes on it today. That recording is your first blog post, waiting to be shaped. If you want more short, practical guides like this one, start with our free Launchpad on our home page.

Need a picture to go with your post? Here is how to use ChatGPT to create images for work.

You run the business, answer the phone, chase the invoices and reply to every enquiry yourself. Everyone keeps telling you AI can take work off your plate. Nobody tells you where to begin, and the tool websites all promise everything at once.

Here is the short answer on AI automation for small business: do not start with a tool. Start with one boring task that happens every week, has a clear start and a clear finish, and would not hurt anyone if it went wrong once. Build it in a no-code tool like Zapier, Make or n8n. First connect the apps with no AI at all, then add one AI step that reads, sorts or drafts. Keep a person approving anything a customer will see. Run it for two weeks, then keep it, fix it or switch it off.

That is the whole method. Below is how to find the right first task in about 20 minutes, which tool fits your budget, a worked example you can copy, and the mistakes that make small business automation more work instead of less.

What AI automation actually means for a small business

Plain automation follows a fixed rule: when a form comes in, add a row to a spreadsheet. Same input, same action, every time. It is reliable, but it breaks the moment the input looks different from what you expected.

AI automation adds a step that can read messy input. It can tell whether an email is a sales question, a complaint or spam. It can pull the amount and due date out of an invoice that looks different from the last one. It can turn a long thread into three lines. Zapier describes most businesses as living in the middle: normal automated workflows with one or two AI steps where some judgment is needed, while the rest stays predictable (Zapier).

That middle ground is where a small business wins. You do not need an “AI agent that runs your company”. You need the enquiry that arrived at 22:00 to be logged, sorted and waiting with a draft reply when you open your laptop.

Step 1: Make a 20-minute task list

Open a notes app or a sheet and write down everything you or your team did last week that felt repetitive. Do not filter yet. Think about email, quotes, invoices, bookings, reviews, social posts, reports and copying data from one place to another.

Then give each task four quick marks:

Multiply how often by how long, and you know which tasks eat the most time. The best first pick scores high on time, has an obvious start and finish, and is low risk. QuickBooks suggests the same routine for small businesses: list your tasks, put a time estimate on each, rank them and start with the top few (QuickBooks).

One honest warning before you go further. If the task itself is a mess, with steps that change every time, automation just makes the mess run faster. Write the steps down first. Our guide to writing SOPs with AI helps you turn “how I usually do it” into a clear list in under an hour.

Step 2: Pick your first automation

These are the tasks small businesses most often start with, because they repeat a lot and the risk is low when you keep a person approving the result. Each one links to a full build on this site.

• New enquiry from your website form. Add the contact to your CRM or sheet, send yourself an alert and let AI draft the first reply. See qualifying leads with an AI workflow and n8n CRM automation for small business.
• Follow-up emails nobody remembers to send. A quote went out on Monday, nothing came back by Thursday, a polite nudge lands in your drafts. See 8 email automation examples.
• Invoices arriving by email. Pull supplier, amount and due date into a spreadsheet so you stop typing them over. See the n8n invoice workflow.
• Bookings and appointment requests. Let a workflow offer open slots and log the booking. See appointment booking automation.
• Support questions you answer every week. Sort incoming questions by topic and urgency, and draft answers from your own FAQ. See automating customer support with AI agents.
• New client onboarding. Welcome email, folder, checklist and calendar invite, all from one signed form. See automating client onboarding with Make.
• Google reviews. Get every new review in one place with a reply drafted in your tone. See responding to Google reviews with AI.

What not to automate first: anything about pricing exceptions, refunds, complaints, legal wording, hiring decisions or money leaving your account. AI can prepare those for you. The decision stays with you.

Step 3: Choose the tool. Make vs Zapier vs n8n for business automation

All three are no-code workflow tools. You connect apps with a trigger (“new form entry”) and actions (“add row”, “send to AI”, “create draft”). They differ mostly in how they charge and how much setup they ask from you. The details below come from each tool’s own pricing page, checked in November 2026. Prices change, so look again before you pay.

A few facts worth knowing before you choose:

• On Zapier’s Free plan, a Zap can only have one trigger and one action, and it checks for new data every 15 minutes. Multi-step Zaps, filters and the built-in AI step start on the Professional plan (Zapier pricing).
• On Make, a single run can use anything from two credits to thousands, depending on how many modules it touches. When you run out, scenarios pause until credits are added (Make pricing).
• n8n’s hosted plans store data in the EU, on servers in Frankfurt. Self-hosting is free but means you look after the server yourself (n8n pricing).

My honest pick for most owners who have never built a workflow: start on the free plan of Zapier or Make, prove one automation works, and only then pay. If you already know you want n8n, our n8n tutorial for beginners gets your first AI workflow running in about an hour. Want the deeper comparison? Read n8n vs Zapier and the Zapier free plan examples.

Step 4: Build it in two layers

The mistake almost everyone makes is building the clever version first. Build the dull version, check it works, then add the AI. Zapier gives the same advice: start with a trigger and an action with no AI, test it, and only then add an AI step (Zapier).

Worked example: the website enquiry

This example is made up to show the steps, not a real client. Sam runs a two-person kitchen fitting company. Enquiries come in through a contact form, sit in the inbox, and some get answered two days later.

Layer 1, no AI (about 30 minutes):

1. Trigger: new entry in the contact form (Google Forms, Typeform, Jotform or your website form).
2. Action: add a row to a Google Sheet called “Enquiries” with name, email, phone, message and date.
3. Action: send Sam an email or Slack message: “New enquiry from [name]”.

Send yourself three test enquiries. Check every column lands in the right place. Only move on when it works three times out of three.

Layer 2, add one AI step: after the row is added, send the message to ChatGPT, Claude or the tool’s built-in AI step with a prompt like this:

You help a small kitchen fitting company sort website enquiries.
Read the message below and return:
1. Type: quote request, question, complaint, supplier, or spam
2. Urgency: high, normal, or low, with one short reason
3. A one-line summary
4. A short, friendly reply draft in plain English that thanks them,
   answers only what we know for sure, and asks for the two details
   we need for a quote: room size and preferred start month.
Never promise prices or dates. If you are unsure, say so in the summary.

Message:
[form message]

Write the type, urgency and summary into the sheet. Put the reply draft in Gmail or Outlook as a draft, not a sent email. Sam reads it, changes a word if needed, and presses send.

That is a complete small business AI automation. It is not flashy. It means no enquiry waits two days, and Sam starts every morning with replies already written.

Step 5: Keep a person in the loop

The U.S. Small Business Administration’s guide gives small businesses three plain pieces of advice that fit here: start small and test low-cost tools to see if they add value, have another person review what AI produces, and avoid feeding it sensitive or proprietary information. It also recommends that a person checks every message and outreach campaign the AI writes, because AI-written outreach can be flagged as spam or read as not written with real knowledge of the person (SBA).

In practice that means:

• Draft, do not send. AI writes into drafts or a “to check” column. A person approves.
• Keep personal data to what the job needs. Do not paste bank details, ID numbers or health information into an AI step.
• Use business accounts you control for the connections, not a staff member’s personal login, so the workflow does not stop when someone leaves.
• Turn on error alerts so a failed run lands in your inbox instead of failing silently.

Once a workflow has been right for a few weeks, you can decide to let low-risk parts run on their own, such as logging and sorting. Customer replies stay with you the longest.

Step 6: Measure for two weeks, then keep, fix or stop

Before you switch it on, write down one number: how long the task takes you now, or how long enquiries wait for a reply. After two weeks, check the same number and count how many runs needed fixing.

• Keep it if it saves time and rarely needs correcting.
• Fix it if the same mistake keeps coming back. Usually the prompt needs one more rule, or the form needs one more field.
• Stop it if you spend more time checking than you used to spend doing. That is not failure. That task was not ready.

Simple maths helps you decide. If an automation saves you 15 minutes a day, that is about five hours over a month of 20 working days. Put your own hourly rate next to that and compare it with the tool’s monthly price.

Then, and only then, pick the second task from your list.

Where AI agents fit in small business workflow automation

An AI agent is different from a workflow. A workflow follows the steps you set. An agent gets a goal (“research this new lead and add a summary to the CRM”) and decides its own steps along the way. Zapier and n8n both include agent features in their plans, and agents are useful for jobs where the next step depends on what the last one found.

For a small business, agents make sense as the second or third thing you build, not the first. Get one or two plain workflows running reliably, then try an agent on a low-stakes job like a daily research summary or sorting your inbox. If you want to go that way, read choosing a no-code AI agent builder and Zapier AI automation.

Mistakes that make automation more work

• Automating five things at once. When something breaks, you cannot tell which part caused it. One reliable workflow beats five half-built ones.
• Picking the tool before the problem. Start from the task on your list, then find the tool that connects the apps you already use.
• Letting AI send straight to customers on day one. One wrong promise in a reply costs more than the time you saved all month.
• Forgetting the limits. Free plans have monthly caps. A busy week can use them up, and the workflow pauses. Check usage after the first week.
• Building on one person’s account. If the only login belongs to someone who leaves, the automation leaves with them.

Questions small business owners ask

Do I need to know how to program?

No. Zapier, Make and n8n are all built so you connect apps by choosing triggers and actions from menus. n8n lets you add your own scripts if you want to, but you never have to.

What is the best AI workflow automation software for a small business?

There is no single best. Zapier is usually the easiest start and connects the most apps (9,000+ by its own count). Make’s free plan includes more monthly actions (1,000 credits against Zapier’s 100 tasks). n8n suits longer workflows and owners who care where their data is stored. Try your first automation on a free plan and let that decide.

How does email marketing automation work for small businesses?

You set up a sequence once, and your email tool sends each message when something happens: a new subscriber gets a welcome email, a customer who bought gets a thank-you and a review request a week later. AI helps you write and personalise those emails, and you approve them before the sequence goes live. For shops, our guide to email automation for ecommerce covers the five flows to set up first.

Is it safe to put customer data through AI automation?

It can be, if you only pass what the task needs, use business accounts you control and keep a person reviewing anything that goes out. Read the data terms of each tool you connect, and leave sensitive details like payment or health information out of AI steps.

How long does the first automation take?

A simple two-layer workflow like the enquiry example is usually an afternoon, including testing. The task list from Step 1 takes about 20 minutes. Do that part today.

Sources

• Zapier: How to use AI for business automation (without a dedicated tech team)
• Zapier: Pricing and plan comparison
• Make: Pricing, credits and plans
• n8n: Plans and pricing
• U.S. Small Business Administration: AI for small businesses, benefits and risks
• QuickBooks: AI for small businesses, how to implement

Missed calls and after-hours questions are a common first agent job for service businesses. If that is the bottleneck on your task list, start with our guide to an AI voice agent for small business — cost, fit, and a simple start without a full call center.

Your next 20 minutes

Open a blank sheet and list last week’s repeated tasks. Mark how often, how long, the start and finish, and the risk. Circle the one that wins. That is your first automation, and the steps above take you from there to a working workflow this week.

Want more practical AI workflows you can use in your business this week? Start with the free Launchpad on our home page.

When the work is answering the same customer questions all day, pair those first automations with conversational AI agents for businesses so chat or SMS can own the repeat asks while your team handles the exceptions.

If marketing video is the job eating your week, read how AI UGC videos for small business work before you hire a creator for every clip.

Want a simple first step for your own income? Read how to use ChatGPT to make money with honest, low-risk ideas.

Not sure which tool to start with? Our guide to the best AI automation tool for your first workflow compares the options in plain words.

Money admin is a good first job to hand over. See how AI bookkeeping for small business works as a safe, checked routine.