Category: AI at Work

You know you should post on LinkedIn. You have real things to say about your work. Then you open the post box, stare at “What do you want to talk about?”, and close the tab again. Or you ask ChatGPT for help and get back something that starts with “I’m thrilled to share” and ends with three rocket emojis. That isn’t you, and your network would spot it in a second.

Here is how to use AI to write LinkedIn posts that still sound like you: bring the raw material yourself (one real moment from your week, the lesson, who it is for), teach the AI your voice once with a few of your own posts, ask for a short draft in that voice, then rewrite the first line and the last line by hand and post it yourself. The AI handles structure and wording. You supply the story and the final say.

That is the whole method. Below you get the voice brief, the prompts I would copy, a worked example from messy notes to finished post, and the LinkedIn rules worth knowing before you let any tool near your account.

Why AI LinkedIn posts all sound the same

When you type “write a LinkedIn post about leadership”, the AI has almost nothing to work with. So it reaches for the most average LinkedIn post it can imagine: a big opening claim, a list of five lessons, a motivational line at the end and a question nobody answers. Thousands of people send the same request every day, so thousands of near-identical posts land in the same feed.

The problem is not the tool. It is the input. AI cannot invent the thing that makes a post worth reading: what actually happened to you, what you noticed, what you would do differently. Only you have that. Give it that, and the output changes completely.

So the job splits cleanly in two:

• Your job: the moment, the lesson, the opinion, the final edit.
• The AI’s job: turning your rough notes into a clear, tight post, offering a few hook options, and cutting what does not earn its place.

Step 1: Build a voice brief once (10 minutes)

This is the step most people skip, and it is the one that stops your posts sounding like everybody else’s. You do it once, save it, and paste it in every time (or store it in your AI tool’s saved instructions, if it has them).

Find three to five posts, emails or messages you wrote yourself and liked. They do not have to be LinkedIn posts. A good email to a client works. Paste them into ChatGPT, Claude or Gemini with this prompt:

Below are a few things I wrote myself. Study how I write and give me a short style guide (max 10 bullet points) that another writer could follow to sound like me. Cover: sentence length, how I open, how I end, words I use often, words I never use, how formal I am, and whether I use humour, emojis or lists.

Then add a "never use" list of phrases that would sound fake coming from me.

[paste your samples here]

Read what comes back and fix anything that is off. Then add your own “never use” list on top. Mine would include words like “thrilled”, “humbled”, “game-changer”, “unlock” and “let that sink in”. Save the result in a note. That is your voice brief.

Step 2: Collect the raw material, not a topic

“Productivity” is a topic. “On Tuesday I cancelled our weekly status meeting and replaced it with a five-line written update, and two people thanked me” is raw material. AI can do a lot with the second one and almost nothing with the first.

Before you open the chat, write down four short lines:

1. What happened: one specific moment, decision, mistake or result from your work.
2. What you learned or think: your actual opinion, even if it is small.
3. Who it is for: the one type of reader who would care (new freelancers, team leads, people hiring their first assistant).
4. What you want them to do: try something, share their own version, or just think about it.

Messy is fine. Typos are fine. Voice notes are fine too: record yourself telling a colleague what happened, transcribe it, and paste the transcript. The way you talk out loud is often closer to your real voice than the way you type.

If you are stuck for ideas, look at the last week: a question a client asked, a tool that saved you an hour, a meeting that went sideways, a small win on your team. Each one is a post.

Step 3: The prompt that writes the draft

Now combine the voice brief and the raw material. This is the prompt I would copy:

You are helping me write a LinkedIn post. Follow my style guide exactly.

STYLE GUIDE:
[paste your voice brief]

RAW MATERIAL:
What happened: [...]
What I think: [...]
Who it is for: [...]
What I want them to do: [...]

RULES:
- Use only the facts above. Do not invent numbers, results, names or quotes.
- 120 to 200 words.
- Short paragraphs of 1 to 3 lines, written to be read on a phone.
- No hashtags in the body, max 1 emoji.
- End with one honest question or a clear takeaway, not a slogan.
- If something is missing to make this a good post, ask me first instead of guessing.

Give me the post, then 3 alternative first lines.

The two rules that matter most are “use only the facts above” and “ask me first”. Without them the AI fills gaps with made-up details, and on LinkedIn your name is on every word.

Step 4: Write the first and last line yourself

On LinkedIn, only the first couple of lines show in the feed before readers have to tap “see more”. That opening line decides whether anyone reads the rest. It is also where AI is weakest, because it loves big general statements (“Leadership is changing.”) that nobody stops for.

Look at the three alternative openers, pick the one closest to you, and rewrite it so it names a specific moment or a specific reader. A few patterns that work well:

• Start in the scene: “Last Tuesday I cancelled our weekly status meeting.”
• Name the reader’s problem: “If your team meeting is mostly people reading out updates, this is for you.”
• Say the honest opinion: “Most status meetings should be an email.”

Then do the same with the last line. Cut any motivational sign-off and replace it with the question you actually want answered, or one plain sentence about what to try.

Step 5: Push back once, then do a fake-detail check

A first draft is rarely the best one. Instead of starting over, ask the AI to critique its own work:

Read this post as a busy [your reader] scrolling on their phone. Which sentence would make them stop reading? Which line sounds generic or like AI? Rewrite only those lines, keep the rest exactly as it is.

After that, read the post out loud and check every detail against what really happened. Did the AI round up a number, add a colleague’s reaction you never mentioned, or make the result sound bigger than it was? Fix it or cut it. If you would not say it to the person sitting next to you, it does not go out.

A worked example: messy notes to finished post

Here is what this looks like for a fictional freelance bookkeeper. The raw material, typed in two minutes:

What happened: client sent me a shoebox of receipts again in March. I made a simple shared folder + told them to snap a photo of each receipt the day they get it. This quarter it took me an afternoon instead of three days.
What I think: most small business owners are not messy, they just have no system that fits their day.
Who it is for: owners of small businesses who dread tax season.
What I want them to do: try the photo habit for one month.

The kind of draft the prompt gives back, after the first line was rewritten by hand:

In March a client handed me a shoebox. Literally. Full of receipts.

It took me three days to sort.

So we tried something small. A shared folder, and one rule: snap a photo of every receipt the day you get it.

This quarter, the same books took me one afternoon.

Here's what I keep seeing: most small business owners aren't messy. They just don't have a system that fits a busy day.

A phone camera and a folder is a system.

If tax season fills you with dread, try the photo habit for one month and see how the next quarter feels.

What's the one admin task you keep putting off?

Notice what the AI did not add: no percentage, no “game-changer”, no hashtags, no extra client quotes. Every fact came from the notes. That is the difference the “only these facts” rule makes.

Turn things you already wrote into posts

The easiest LinkedIn posts are hiding in work you have already done. A newsletter section, a client case study, a long email that explained something well, your notes from a training day. Paste one in with your voice brief and ask:

From the text below, find 3 separate ideas that could each be a short LinkedIn post for [your reader]. For each: the one-sentence point, the specific example from the text that proves it, and a first line. Do not write the full posts yet.

Pick the idea you actually care about, then run Step 3 on it. If you write a newsletter, our guide on using AI to write a newsletter pairs well with this: one good issue can feed a week of posts. The same goes for client stories, which you can shape with AI case study writing first and then cut into a post.

Which AI tool should you use?

Any of the big three can do this well if you give them the voice brief and real material. The difference is smaller than people think. Pick the one you already have open all day, and save your voice brief where that tool keeps standing instructions so you do not paste it every time. If you want a side-by-side on tone and editing, see ChatGPT vs Claude for writing.

What about LinkedIn’s own AI? LinkedIn used to offer an AI writing tool for posts, but its help center now says that tool is not available at this time and that new post-writing tools are coming. Until then, a general chat tool plus your voice brief does the job.

Format it for the feed

A few practical things that help a post get read:

• Length: LinkedIn posts can hold up to 3,000 characters. You rarely need that many. A focused post of 120 to 250 words is easier to read on a phone and easier to finish.
• White space: one idea per short paragraph. Most of your readers are scrolling with a thumb.
• One point per post: if the AI gave you five lessons, keep the best one and save the rest for later posts.
• Drafts and scheduling: LinkedIn lets you save a draft and schedule a post from the post box (same help page), so you can write three posts in one sitting and spread them over the week.

Where to stop: what AI should not do on LinkedIn

Writing help is fine. Handing over your account is not. LinkedIn’s User Agreement says you must not use bots or other unauthorized automated methods to create, comment on, like or share posts, or otherwise drive inauthentic engagement. And LinkedIn’s help page on posting warns that if it detects excessive posting that looks inauthentic, or the use of an automation tool, it may limit the visibility of those posts.

In practice that means:

• Let AI help you draft. Post it yourself, from your own account.
• Do not use tools that auto-comment or auto-like in your name.
• Write your replies to comments yourself. A short real answer beats a polished generated one, and it is where the actual relationships form.
• Keep client names, private numbers and anything personal out of the chat unless your company’s AI policy allows it. The University of Massachusetts Isenberg career team gives the same advice for networking messages: write your own draft first and remove personally identifiable information, then check any facts the AI adds.

Common mistakes (and the quick fix)

• Asking for a post about a topic. Fix: give one real moment and your opinion about it.
• Skipping the voice brief. Fix: ten minutes once, and every post after that starts closer to you.
• Posting the first draft. Fix: rewrite the first and last line, push back once, read it out loud.
• Letting the AI add proof. Fix: “use only the facts above” in every prompt, and check numbers before posting.
• Five lessons in one post. Fix: one point, one post. The rest is next week’s material.
• Hashtag walls and emoji bullets. Fix: ban them in the voice brief.

If LinkedIn is only one part of a bigger publishing week, you can start from one real idea and cut it into posts, a newsletter and a short script. Our guide to AI for content creation walks through that weekly workflow, including a voice file you reuse so every format still sounds like you.

Want the same voice across more than LinkedIn? Our guide to an AI social media content generator shows how to set up one saved prompt for every platform.

FAQ

Can I use AI to write LinkedIn posts?

Yes. Nothing in LinkedIn’s rules stops you from using a writing tool to draft your own posts; LinkedIn even offered its own AI post writer for a while. What its User Agreement forbids is using bots or unauthorized automation to post, comment, like or share, or to fake engagement. Draft with AI, then post and reply yourself.

Do people use AI to write LinkedIn posts?

Plenty do, and readers have learned to spot the generic ones: the same openers, the same lists, the same sign-offs. That is exactly why the voice brief and your own raw material matter. A post built on a real moment from your week does not read like a template, whoever helped with the wording.

What is the best ChatGPT prompt for a LinkedIn post?

The one in Step 3: your style guide, four lines of raw material, and clear rules (only these facts, a word range, short paragraphs, ask before guessing). The prompt works the same way in Claude and Gemini.

Can I automate LinkedIn posts with AI?

You can automate the writing preparation: a weekly reminder, a note where you collect raw material, a prompt you reuse. Use LinkedIn’s own scheduling if you want posts to go out later. Fully automatic posting and commenting through outside bots is what LinkedIn’s terms rule out, and its help center says automation can lead to limited visibility.

How do I make AI-written posts sound less like AI?

Give it your own writing samples to copy, a list of phrases you never use, and a real story. Then rewrite the first and last line yourself and read the whole thing out loud. If a sentence makes you wince, cut it.

The voice brief you built here works beyond the feed. If you are going after a new job, reuse it to write a cover letter with AI without it sounding like everyone else’s.

Your next 15 minutes

Do this today: build your voice brief with the prompt from Step 1, then write down one thing that happened at work this week in four lines. Run the Step 3 prompt, fix the first and last line, and post it. One real post is worth more than a month of planning to post.

If you want more of these practical, do-it-this-week AI workflows for your job, start with the free Launchpad on our home page. And if you look after social accounts for a business rather than just your own profile, the ChatGPT prompts for social media managers guide is the next step. For writing beyond LinkedIn, see the AI writing assistant guide for busy professionals.

Posts land better when the profile behind them is clear. If yours still reads like an old CV, our guide on how to use AI to write a LinkedIn profile walks through the headline, the About section and your experience with copy-ready prompts.

Job hunting as well? See how to use AI to write a resume to tailor your CV to each job ad without it sounding generic.

The deadline is in nine days. The funder’s guidelines are a 14-page PDF, your notes are spread across three documents, and the blank “Statement of Need” box is staring at you. You have heard ChatGPT can write a grant proposal. You have also heard funders can tell.

Here is how to use AI to write a grant proposal without losing what makes it yours: let the AI read the funder’s guidelines and turn them into a checklist, feed it your real facts (the need, your numbers, your team), have it draft one section at a time from those facts, then check every claim, every figure and every reference yourself before you submit. You bring the ideas and the proof. The AI handles the structure and the first draft of the wording.

That split matters more for grants than for almost any other document, because some funders now have written rules about it. Below you will find the step-by-step method, prompts you can copy, what the big funders say about AI, and the mistakes that get proposals rejected.

Can you use AI to write a grant proposal?

Usually yes, for help. Rarely for the whole thing. Check the rules of your funder first, because they differ a lot:

• NIH (US health research): since the 25 September 2025 receipt date, NIH will not consider applications that are substantially developed by AI, or that contain sections substantially developed by AI, to be the applicant’s original ideas. It also warns that AI use can lead to plagiarism and fabricated citations (NIH notice NOT-OD-25-132).
• NSF (US science): proposers are encouraged to say in the project description whether and how they used generative AI (NSF notice on generative AI in merit review).
• UKRI (UK research and innovation): applicants may use generative AI, but must not put other people’s sensitive or personal data into it without consent, must make sure nothing in the application is fabricated, falsified or plagiarised, and are expected to be open about substantive use. Polishing your English or cutting the word count counts as minimal use that you do not need to declare (UKRI policy on generative AI in applications).

Most local foundations and community funds have no written AI policy at all. That does not make a generic, AI-sounding proposal a good idea. It just means the reviewer decides with their gut. So the safe rule for every funder is the same: the ideas, facts and numbers come from you, and you can explain every sentence if someone asks.

What AI is good at in grant writing (and what it is not)

Think of the AI as a fast, tireless assistant who has never met your organisation.

Good jobs for AI:

• Reading a long call for proposals and listing every requirement, word limit and scoring criterion.
• Turning your rough notes into a clean first draft of one section.
• Cutting a 900-word section down to the 500 the form allows.
• Checking your draft against the funder’s criteria and pointing out gaps.
• Rewriting jargon into plain language a non-expert reviewer will follow.

Bad jobs for AI:

• Inventing your need statistics. It will happily produce a convincing number with no real source.
• Writing your references. Made-up citations are one of the risks NIH names in its notice.
• Deciding what your project is. If the AI came up with the idea, a strict funder may not see it as yours.
• Building the budget from scratch. It can format and check your budget, but the costs must come from real quotes and salaries.

Step 1: Turn the funder’s guidelines into a checklist (10 minutes)

The most common reason a proposal fails is not weak writing. Candid, the nonprofit behind GuideStar and the Foundation Directory, asked its community which red flags funders watch for. The most cited one was simply not following the funder’s instructions, and budgets that do not add up came next (Candid: 7 red flags for funders).

This is where AI saves you real time. Upload the guidelines PDF (or paste the text) into ChatGPT, Claude or Gemini and use this prompt:

You are helping me prepare a grant application.
Read the attached funder guidelines and give me:
1. A checklist of every required section, with its word or page limit.
2. Every formatting rule (font, margins, file type, page count).
3. The scoring or selection criteria, quoted word for word.
4. All required attachments.
5. The deadline and any eligibility rules.
Quote the guidelines for each item and give the page number.
If something is unclear or missing, say so. Do not guess.

The “quote it and give the page number” line is what makes this safe. You can check any item in seconds, and you will spot it if the AI got something wrong. Keep this checklist open for the rest of the job. If your guidelines run long, our guide to summarizing long documents with ChatGPT shows how to keep the AI accurate on big files.

Step 2: Write a fact sheet only you can write (20 minutes)

A generic proposal comes from generic input. Before the AI writes a single sentence, put your raw facts in one document. Bullet points are fine, typos are fine.

• The need: what problem in your community or field this solves, and the real data that shows it (with the source of every number).
• The project: what you will actually do, step by step, and over what time.
• Who does the work: a named person or role for every activity.
• Outcomes: what will be different for people when you are done, and how you will measure it.
• Your track record: past projects, results, partners, letters of support.
• Budget facts: salaries, quotes and price sheets you already have.
• Why this funder: how your project matches their stated priorities.

Candid’s guide to proposal writing lists the same building blocks funders expect: an executive summary, a statement of need, a project description with activities, outcomes, evaluation and sustainability, information about your organisation, and a budget (Candid Learning: How do I write a grant proposal?). Your fact sheet feeds every one of them.

Before you paste anything: remove personal details about clients, patients or staff, and anything confidential you do not have permission to share. UKRI tells applicants that the confidentiality of what you type into a generative AI tool is not guaranteed. In ChatGPT you can also switch off “Improve the model for everyone” under Settings, Data controls, so your chats are not used for training (OpenAI Help: Data controls). If your organisation has its own AI policy, that comes first.

Step 3: Draft one section at a time

Never ask for “a grant proposal for my project”. Ask for one section, give it the facts it needs, and tell it what it may not do. Here is a prompt for the statement of need:

Write the Statement of Need for a grant application to [funder name].
Limit: [number] words.
The funder's priorities, quoted from their guidelines: [paste].
Use ONLY the facts below. Do not add statistics, studies or
examples that are not in this list. If a claim needs a number
I have not given you, write [NEED DATA] in that spot.
Write in plain, warm, specific language. No buzzwords.
Facts:
[paste the need section of your fact sheet]

The [NEED DATA] trick is the most useful line in that prompt. Instead of quietly inventing a figure, the AI leaves a visible gap that you fill with a real, sourced number.

Repeat the same pattern for each section: project description, goals and outcomes, evaluation plan, organisational background. Paste the matching part of your fact sheet each time. Keep each section in its own chat message so you can redo one without touching the others.

Turn activities into outcomes

Candid points out that funders often see proposals that mix up activities, outputs and outcomes. Their short version: activities are plans, outputs are counts, outcomes are changes. AI is good at sorting these once you give it the rule:

Here is my list of what the project will do: [paste].
Sort it into three columns: Activities (what we do),
Outputs (how many, how often), Outcomes (what changes for people).
For each outcome, suggest how I could measure it.
Do not invent targets. Mark any outcome that has no measure yet.

Step 4: Make it sound like you

The first draft will be clear but flat. Reviewers read a lot of proposals, and Candid lists “not personalized to the funder” among the red flags. Three fixes that work:

• Add one real story. One person, one moment, with permission and without identifying details. Only you have it.
• Use your own words for the core idea. Write the two or three sentences that explain your project yourself, then let the AI tidy them, not replace them.
• Ask for a style match. Paste a section from a proposal you wrote before and say: “Rewrite this draft so it sounds like the sample. Keep every fact. Do not add anything.”

Claude and ChatGPT each have their own writing habits. If one sounds stiff for your organisation, try the same prompt in the other. Our comparison of ChatGPT vs Claude for writing covers which tends to suit which voice.

Step 5: Use AI as a tough reviewer

This is the step most people skip, and it is where AI helps most. Paste your full draft and the scoring criteria from Step 1:

Act as a strict reviewer for [funder name].
Score my draft against each criterion below from 1 to 5.
For every score under 5, quote the weak sentence and say
what is missing. Then list:
- any requirement from the checklist I have not covered
- any activity with no named person responsible
- any budget line that does not match an activity
- any claim that has no source
Criteria: [paste]
Checklist: [paste]
Draft: [paste]

You will not agree with every point, and that is fine. The value is in the gaps it finds: the missing sustainability paragraph, the activity with nobody assigned, the evaluation plan that measures attendance but not change.

Step 6: Check the budget and the facts yourself

Do not let the AI do arithmetic you will not check. Keep the budget in a spreadsheet, where the totals are formulas, and ask the AI only to compare the narrative with it: “List every cost or activity mentioned in this narrative and tell me if it appears in this budget table.” If you want help with the formulas themselves, see how to use AI to write Excel formulas.

Then do a final human pass with three questions for every sentence:

1. Is this true, and could I show the proof?
2. Is every number traced to a real source I have opened myself?
3. Does every reference actually exist and say what I claim it says?

If your funder asks you to disclose AI use, or you choose to, a short honest line works: what you used it for (structure, editing, checking against criteria) and that all ideas, data and final text were reviewed by your team.

A worked example

Say you run a small community kitchen and want funding for weekend cooking classes for teenagers. Here is how the method plays out:

• Step 1: The AI turns the local foundation’s 10-page call into a checklist and flags something easy to miss: they want a sustainability plan for after the grant ends.
• Step 2: You write the fact sheet: how many teens are on your waiting list, your own attendance records, the two volunteer chefs by name, a quote from the hall rental.
• Step 3: The need statement draft comes back with two [NEED DATA] markers where it wanted local figures you had not given it. You fill one from your own records and remove the other claim.
• Step 4: You add a short story about a teenager who now cooks for her family on Sundays, shared with permission.
• Step 5: The reviewer prompt notices the budget has an ingredient line but the narrative never says who buys the ingredients. You add one sentence.
• Step 6: You check the totals in the spreadsheet and submit.

The AI saved the slow, dull parts: reading the call, structuring, trimming, checking. The parts that win funding, the real need and the real people, stayed yours.

Mistakes to avoid

• Asking for the whole proposal in one go. You get a long, generic document that fits no funder.
• Trusting statistics or references from the AI. Check every one against the original source.
• Pasting sensitive data. Client names, health details and unpublished research ideas stay out unless your organisation’s rules allow it.
• Sending the same text to every funder. Reuse your fact sheet, but rewrite the “why this funder” part every time.
• Skipping the rules check. One missed attachment or word limit can sink a strong proposal.

Frequently asked questions

Can funders tell if a grant proposal was written by AI?

Some say they try. NIH writes that it uses technology to detect AI-generated content. Detection tools are not perfect, but a reviewer does not need one to notice text that is vague and could be about any organisation. Specific facts, real people and your own voice are the best protection.

Is it cheating to use ChatGPT for a grant application?

Not when the ideas and facts are yours and you follow the funder’s rules. UKRI, for example, treats things like improving your English or cutting the word count as minimal use. Letting the AI invent the project or the evidence is where it becomes a problem.

Which AI is best for grant writing?

ChatGPT, Claude and Gemini can all do the steps above. Pick one that can read long PDFs, because you will upload guidelines and earlier proposals. If you are new to them, start with our ChatGPT tutorial for beginners.

Can I use AI for a letter of inquiry too?

Yes, with the same method. A letter of inquiry is a short version of the proposal, so build it from the same fact sheet and keep it within the length the funder asks for.

Do I need special grant writing software?

Not to start. The method on this page works with a general AI chat tool and a spreadsheet. Dedicated tools may add funder databases or templates, but the rules about facts, sources and your own ideas stay the same.

Your next step

Take the guidelines of the next grant on your list and run the Step 1 prompt today. It takes ten minutes, and you will know exactly what the funder wants before you write a word. Writing a business case rather than a grant? Our guide on using AI to write a business plan follows the same facts-first idea.

If you want more practical wins like this one, using AI for the real work on your desk, start with the free Launchpad on our home page.

If you also send client proposals (not just grants), see our guide to AI proposal writing software for when a chat AI is enough and when dedicated tools help.

You already use Zapier to move things from A to B. A form comes in, a row gets added, someone gets a Slack ping. But the moment a step needs a bit of judgment (“is this email urgent?”, “summarise this thread”, “write a polite nudge”), the Zap stops and the job lands back on your desk.

Here is the short answer on Zapier AI automation: add an AI by Zapier step in the middle of a normal Zap. The trigger and the actions stay fixed, and the AI step does only the part that needs reading or writing: it summarises, sorts, extracts or drafts, and hands clean fields to the next step. Start with the Standard model tier, return named output fields, and let the AI make drafts, not final sends, until you trust it. You can test an AI step on the Free plan, but running it live needs a paid Zapier plan (details below).

Below: which of Zapier’s AI options to use when, what each one costs in tasks, two builds you can copy (a ChatGPT summary to a Slack channel, and an AI follow-up email in Gmail), and the guardrails that keep it from embarrassing you.

What Zapier AI automation actually means

Zapier describes AI automation as AI models placed inside your workflows, so software can handle work that needs judgment and not only work that follows fixed rules. Their own summary of the pattern is the most useful sentence on the topic: a deterministic workflow with AI judgment inside it. The same trigger follows the same steps every time, and AI only handles the moments that need interpretation, like reading a support ticket or drafting a reply (Zapier, What is AI automation?).

That is the mindset to keep. You are not handing your day to a robot. You are taking the one annoying step in a process you already run, and letting AI do that step.

Good first candidates are jobs where you read something and then type a short answer:

• Reading a long email or form answer and deciding which bucket it goes in.
• Turning a messy thread or meeting transcript into three bullet points.
• Pulling names, dates and amounts out of free text into separate fields.
• Writing a first draft of a reply, a follow-up or a status update.

Zapier’s AI options, and when to use each

Zapier now has a few different doors into AI. They overlap, so here is the plain version.

1. AI by Zapier (the AI step inside a Zap)

This is the one you will use most. It is a built-in action you add like any other step. You write a prompt, insert fields from earlier steps by typing /, pick a model, and test it with Preview. You can use a built-in connection or connect your own account with OpenAI, Anthropic, Google Gemini, Azure OpenAI or Amazon Bedrock. There are ready prompt templates for Summarize, Write, Classify and Extract (Zapier Help, Use AI by Zapier).

The setting that matters most: Output Fields. Define the fields you want back (for example category, summary, reply_draft) and each one becomes something you can map into the next step. Without output fields you get one combined block of text, which is harder to use downstream.

2. AI by Zapier with tools (what used to be Zapier Agents)

Zapier folded its separate Agents product into the AI step. Agent instructions now live in the prompt, the tools an agent could use live in a tools picker inside the step, and run history shows up in Zap History like every other automation (Zapier, Zapier Agents is now AI by Zapier).

With tools switched on, one AI step can look things up and act, for example find a customer’s latest order and draft a reply with the delivery date. For any tool that creates, updates or deletes data you can turn on Require approval before running: the run pauses and Zapier emails you a review link before anything changes. Tools need a Professional, Team or Enterprise plan and an Advanced or Premium model, or your own API key (Zapier, Add apps, knowledge and web tools to an AI by Zapier step).

3. Zapier Copilot (AI that builds the Zap for you)

Copilot does not run inside your workflow. It helps you build it. You describe the workflow in plain language at the top of the Zap editor (“When X happens, do Y, then do Z”), and it sets up the trigger, actions and field mappings. You can choose Auto-build or Ask as you build, where it asks before each change. It is in open beta and listed for every plan, Free included (Zapier Help, Use AI to generate Zap workflows).

Zapier’s own advice is to start with Ask as you build while you are new, and to test each step yourself, with a small sample first, before a bigger run uses more tasks (Zapier Help, Best practices for Copilot). In practice the outline is the easy part; the field mapping is where you still want your own eyes.

4. The ChatGPT (OpenAI) app

The older route. It connects Zapier straight to your OpenAI account with an API key. You need an OpenAI account with prepaid billing, which is separate from a ChatGPT subscription; a ChatGPT Plus plan does not cover it. Usage is billed by OpenAI (Zapier Help, Get started with ChatGPT (OpenAI) on Zapier). Use it if you specifically want your own OpenAI setup. For most people starting today, the AI by Zapier step is the simpler choice, because there is no separate key or bill to manage.

Quick pick

• You want one smart step in a fixed process: AI by Zapier, Standard tier, no tools.
• The AI needs to look something up or act in another app: AI by Zapier with tools, approval on for anything that writes.
• You know what you want but not how to wire it: Copilot, in Ask as you build mode.
• You already pay for the OpenAI API and want it billed there: the ChatGPT (OpenAI) app, or AI by Zapier with Bring Your Own Key.

What it costs: tasks, tiers and the Free plan

This is the part people skip and then get surprised by. AI by Zapier steps are counted in tasks, and the model tier sets the multiplier (Zapier Help, AI by Zapier model tier pricing):

• Standard: 1 task per run. No tools. Fine for simple prompts.
• Advanced: 3 tasks per run, tools allowed.
• Premium: 5 tasks per run, tools allowed. Zapier’s help page lists this as the default for new steps.
• Your own AI account: 1 task rate, and you pay your provider for the model use.

Every successful tool call adds the same rate again. Zapier’s own example: a Premium step that calls two tools uses 15 tasks. A step pauses and asks for your approval if one run hits 75 tasks, a limit you can raise to 500.

So check the tier on every new AI step. If the job is “summarise this” or “pick a category”, switch it down to Standard. Leaving a simple summary on Premium quietly costs five times as many tasks for every run.

And one honest limit: on the Free plan you can preview an AI by Zapier step on the Standard tier, but you cannot publish a Zap that contains one or add tools (Zapier Help). You can test the idea for free. To run it for real, you need a paid plan. If you are still deciding whether Zapier is worth paying for at all, our Zapier free plan examples show what runs without AI, and n8n vs Zapier compares the two on cost.

Build 1: send a ChatGPT summary to a Slack channel

The classic. Long things come in (a support email, a form answer, a call transcript, a new doc), and your team wants the gist in Slack without opening anything.

The Zap:

1. Trigger: whatever produces the long text. A new email with a certain label in Gmail, a new form response, a new file in a folder.
2. Action: AI by Zapier. Tier: Standard. Insert the body text with /. Add output fields summary, action_needed (yes or no) and owner.
3. Filter (optional): only continue if action_needed is yes, so the channel does not fill up with noise.
4. Action: Slack, send a channel message. Map summary and owner into the message, plus a link back to the original.

A prompt that works for this:

You summarise incoming messages for a busy team channel.

Message:
[insert the email or form body here with /]

Return:
summary: max 3 short bullet points. Facts only, no opinions.
action_needed: "yes" if someone on our team has to do something, otherwise "no".
owner: the team or person who should act (sales, support, billing, or "unclear").

If the message is too short or empty, write "nothing to summarise" in summary.
Never invent names, dates or amounts that are not in the message.

Why “ChatGPT summary” works here even without the ChatGPT app: inside AI by Zapier you can pick an OpenAI model, or connect your own OpenAI account with Bring Your Own Key. If your team already summarises threads by hand, the prompt in our Slack thread action items guide slots straight into this step.

Test it with three real messages: a normal one, a very short one, and one with no clear action. If all three come out right, publish.

Build 2: a Gmail follow-up email written by AI

You sent a proposal, a quote or an intro, and nobody answered. You mean to follow up. You forget. This Zap writes the nudge for you and leaves it in Drafts, so you still decide what goes out.

The Zap:

1. Trigger: a row in a Google Sheet reaches its follow-up date, or a Gmail label like “waiting” has sat for a few days (a schedule trigger plus a search works for this).
2. Action: AI by Zapier. Tier: Standard. Give it the original subject, the person’s first name and one line on what you sent. Output fields: subject and body.
3. Action: Gmail, create draft. Map subject and body, and the recipient from the sheet or the original email.

Write a short, friendly follow-up email.

Recipient first name: [name]
What I sent them: [one line, e.g. "the quote for the new website"]
Sent on: [date]

Rules:
Max 4 sentences. Warm, not pushy.
Do not apologise for following up.
End with one easy question they can answer in a line.
Do not invent deadlines, discounts or details that are not given above.
Return subject and body separately.

Why a draft and not a send? Because the AI does not know that the client called you yesterday. A draft costs you ten seconds of reading and saves you the awkward “we spoke on the phone already” reply. Once a few weeks of drafts come out right, you can decide whether some types are safe to send automatically.

We have a full step-by-step version of this for proposals, with CRM options and the second and third nudge, in how to automate proposal follow-up emails with Zapier, and a tool-neutral version in automate follow-up emails without code.

Three more Zapier AI automation ideas worth trying

• Sort the inbox you share with others. New email in a shared address, AI step classifies it (sales, support, invoice, spam) with a category output field, then Paths send each category somewhere different.
• Pull data out of free text. A form with an open “tell us about your project” box. The AI step extracts budget, deadline and company name into fields, and the next step adds a clean row to your sheet or CRM.
• Score and route leads. AI reads the form answers against your own rules and returns a score plus one line of reasoning. We built this one out in full in our AI lead qualification workflow.

Zapier’s own customer stories follow the same shape, for example a voice note in Slack that AI by Zapier turns into a blog draft, which then goes through Slack for approval before it is published (Zapier). AI does the reading and writing, people approve, the fixed steps do the rest.

Guardrails before you switch it on

• Drafts first. Anything that goes to a client starts as a draft or a Slack message for a person to check.
• Approval on every tool that writes. If your AI step can create, update or delete things in another app, turn on Require approval before running for that tool.
• Tell it what not to do. “Never invent names, dates or amounts” in the prompt prevents most of the embarrassing output.
• It does not remember. Each run starts fresh; AI by Zapier does not learn from earlier runs. Put the rules in the prompt, every time (Zapier Help, Limitations).
• It cannot read a URL by itself. The plain AI step does not fetch websites; pass the text in from an earlier step, or switch on the browse tool on a plan that supports tools.
• Check Zap History weekly. Every run shows the tools called, the inputs and outputs, the tier and the tasks used. Five minutes a week tells you if the prompt is drifting or the cost is creeping.
• Company data rules still apply. If client data is not allowed to go to an outside AI model, keep AI out of that Zap.

FAQ

Does Zapier have built-in AI?

Yes. AI by Zapier is a built-in step you add to any Zap, and Copilot is a built-in assistant that helps you build Zaps from a plain-language description.

Do I need a ChatGPT subscription to use AI in Zapier?

No. The AI by Zapier step works with a built-in connection. You only need an OpenAI API account (with prepaid billing, separate from ChatGPT Plus) if you use the ChatGPT (OpenAI) app or bring your own key.

Can I use Zapier AI automation on the Free plan?

You can try it: Copilot is available on Free, and you can preview an AI by Zapier step on the Standard tier. Publishing a Zap with an AI by Zapier step, and adding tools, needs a paid plan.

What happened to Zapier Agents?

Zapier moved Agents into the AI by Zapier step. You add tools to the step, set approval per tool, and the runs show in Zap History.

Is Zapier the right tool for AI automation, or should I use n8n or Make?

If you want lots of ready app connections and little setup, Zapier is the easy start. If you run many workflows and cost per run matters, n8n is worth a look. Our n8n vs Zapier comparison walks through that choice.

Your next step

Pick one thing you read and retype every week. Build the Slack summary or the follow-up draft above around it, on the Standard tier, and let it run for a week while you check every result. That one Zap teaches you more than ten tutorials.

If you want to learn this alongside other people doing the same, with practical AI lessons for real work, start free on our home page with RobotsHelp Launchpad.

Running a small team and wondering where AI fits beyond Zapier? Our guide to AI automation for small business covers what to automate first and how to start without a big budget.

Sources

• Zapier: What is AI automation? A complete guide
• Zapier Help: Use AI by Zapier to analyze and return data
• Zapier Help: AI by Zapier model tier pricing
• Zapier: Add apps, knowledge and web tools to an AI by Zapier step
• Zapier: Zapier Agents is now AI by Zapier
• Zapier Help: Use the power of AI to generate Zap workflows (Copilot)
• Zapier Help: Best practices for using Zapier Copilot
• Zapier Help: How to get started with ChatGPT (OpenAI) on Zapier

If you are still comparing options, see the best AI automation tool for beginners and run the short test it describes.

A supplier sends you a 12-page agreement in German. A client in Madrid wants your onboarding guide in Spanish by Friday. You paste a paragraph into a translator, get something that is roughly right, and start to wonder if the whole document will come out the same way: understandable, but not something you would sign or send.

Here is how to use AI to translate a document properly: pick the tool by what matters most. If the layout has to survive, upload the file to a document translator (Word’s built-in Translate, Google Translate or DeepL). If the wording, tone and terms have to be right, use ChatGPT or Claude with a prompt that tells it who will read it, which terms to keep, and to flag anything it is unsure about. Then check the result before anyone relies on it, and use a certified human translator for anything official.

That is the whole method. Below you get the tool choice, the prompt, how to handle long PDFs and Word files, how to keep confidential documents safe, and where AI translation stops being good enough.

Two kinds of AI translation, and when to use which

Most frustration comes from using the right tool for the wrong job. There are two families, and they are good at different things.

Document translators: keep the layout

These take a whole file and give you back the same file in another language, with headings, tables and bullet points where they were.

• Microsoft Word. Go to Review, then Translate, then Translate Document. Word opens a translated copy in a new window and leaves your original alone. It needs a Microsoft 365 subscription or Office 2021 and an internet connection (Microsoft Support: Translate text into a different language).
• Google Translate. On a computer, open Google Translate and choose Documents. It takes .docx, .pdf, .pptx and .xlsx files up to 10 MB, and PDFs up to 300 pages. Text inside images and scanned PDF pages is not translated (Google Translate Help: Translate documents).
• DeepL. A free account translates Word, PowerPoint and PDF files up to 5 MB or 100,000 characters, and those translated files are not editable without a paid plan (DeepL Help Center: File formats).

Use these when the document is long, the structure matters, and “correct and readable” is good enough: an internal manual, a supplier’s spec sheet, a report you only need to understand.

Chat assistants: get the wording right

ChatGPT and Claude read the document and write the translation like a person would, so you can steer them. You can tell them the reader is a nervous first-time customer, that “onboarding” must stay in English, that the tone is informal Dutch “je” and not formal “u”. A document translator cannot take any of that into account.

• ChatGPT accepts DOCX, PDF and TXT, plus spreadsheets and PowerPoint files. Documents can be up to 512 MB and 2 million tokens per file, and free users can upload 3 files per day (OpenAI Help: Uploading files to ChatGPT).
• Claude accepts PDF, DOCX, TXT, RTF, ODT and more, up to 20 files per chat. It reads text and visuals in PDFs up to 100 pages, and text only from 101 to 1000 pages (Claude Help Center: Upload files to Claude).

Use a chat assistant when the words will be read by a customer, a partner or a boss: a proposal, a website page, a policy, a sales email sequence.

My own habit, working between Dutch and English most weeks: I run the long, dry stuff through a document translator, and anything with my name on it goes through a chat assistant with the prompt below. Then I read it out loud. That last step catches more than any tool.

Step 1: Get the document ready (2 minutes)

Bad input gives bad translations. Before you upload anything:

• Check that the text is real text. Try to select a sentence in the PDF. If you cannot, it is a scan, and Google Translate will leave it untranslated. ChatGPT and Claude can usually read a clear scan, but check names and numbers extra carefully.
• Google Docs users: ChatGPT cannot take a .gdoc file. Export it as PDF or DOCX first.
• Write down your fixed terms. Product names, job titles, legal terms your company uses, words that must stay in the original language. Five to ten lines is plenty. This is your glossary.
• Know your reader. One sentence: who reads this, and what do they need to do after reading it?

Step 2: The prompt to paste

Upload the file to ChatGPT or Claude, then paste this. Fill in the square brackets once and save it, so every document your team translates sounds the same.

Translate the attached document from [source language] into [target language].

Reader: [who reads it, e.g. new customers in Spain who are not technical].
Tone: [e.g. friendly and clear, informal "tú" form].

Rules:
* Translate everything, including headings, tables and footnotes. Do not summarise or skip parts.
* Keep the same structure: same headings, same order, same lists.
* Keep these terms exactly as written: [glossary terms].
* Keep all names, numbers, dates, amounts and SKUs exactly as in the original. Convert date formats to the [target country] style only if I say so.
* Where a sentence is ambiguous or has no clean equivalent, translate it as best you can and mark it with [CHECK] plus a short note in English on why.

Start with the first [3 pages / 2,000 words]. I will say "next" for the rest.

Three lines do most of the work. “Do not summarise or skip parts” stops the model from quietly shortening long sections. “Keep all names, numbers, dates” protects the details that cause real damage when they change. And the [CHECK] marks turn the model’s doubts into a short list for you, instead of a confident guess buried in paragraph nine.

Step 3: Long documents, in chunks

Even when a tool accepts a big file, asking for 40 pages in one answer is where things go wrong: the reply gets cut off, or later sections come back shorter than the original. Work in pieces.

1. Ask for the first few pages, as in the prompt above.
2. Type “next” for each following chunk, in the same chat, so your rules and glossary stay in place.
3. Every few chunks, ask: “List any term you translated in two different ways so far.” That catches drift, like “invoice” becoming two different words halfway through.
4. Paste the chunks into one document at the end and compare the length with the original. A section that is much shorter is the first place to look for something skipped.

For very long PDFs in Claude, remember the split: up to 100 pages it also reads charts and images, from 101 pages it reads the text only. If your figures carry labels that need translating, upload those pages separately.

Step 4: Check it before anyone relies on it

AI translations read smoothly, and that is exactly the risk. OpenAI’s own help page says ChatGPT can produce incorrect or misleading output, sometimes with confidence, and advises treating it as a first draft and verifying important information (OpenAI Help: Does ChatGPT tell the truth?). A fluent sentence that says the wrong thing is worse than a clumsy one, because nobody stops to question it.

A check that fits in ten minutes:

• Numbers, names and dates. Put the original and translation side by side and compare every figure. This is boring and it is the most important part.
• Every [CHECK] mark. Decide each one yourself, or ask a colleague who speaks the language.
• Back-translate the risky parts. Open a new chat, paste the translated version of a key paragraph, and ask for a translation back into your language. If the meaning shifted, you have found a problem. It does not prove the translation is perfect, but it is a quick way to spot a big error.
• A native speaker for anything public. Five minutes of their time on the final version is worth more than any prompt.

Translating a PDF or Word document without losing the formatting

This is the most common question, and the honest answer is: chat assistants are not great at it. ChatGPT and Claude give you the text in the chat, and you rebuild the layout yourself.

Two ways to get both good wording and the original layout:

1. Translate the file first, polish second. Run the document through Word’s Translate Document or Google Translate so the layout comes back intact. Then paste the key sections (intro, anything customer-facing) into ChatGPT or Claude with your tone and glossary, and replace those parts in the translated file.
2. Ask for a file back. If your ChatGPT or Claude plan can create files, ask it to “return the translation as a .docx with the same headings and tables”. Check the result carefully: complex tables, text boxes and headers often move.

For a scanned PDF, neither family will keep the layout well. Ask a chat assistant for the text, check it, and drop it into a clean template.

Confidential documents: what to switch on first

Contracts, HR files and client data deserve a minute of thought before you upload them.

• On a personal ChatGPT Free, Plus or Pro account, your chats can be used to improve OpenAI’s models by default. You can turn that off under Settings, then Data controls, then “Improve the model for everyone”. Temporary chats are not used for training and may be kept for up to 30 days for safety (OpenAI Help: Data controls in ChatGPT).
• On ChatGPT Business or Enterprise, OpenAI does not use workspace content for training by default (same source).
• Follow your company’s rules first. If your employer has a policy on AI tools or an approved tool, that wins. When in doubt, remove names, account numbers and addresses before uploading, and put them back afterwards.

When AI translation is not enough

Some documents need a human translator, no matter how good the AI version reads.

• Official and legal submissions. For example, US immigration rules require any foreign-language document submitted to USCIS to come with a full English translation that the translator certifies as complete and accurate, plus a statement that they are competent to translate (8 CFR 103.2(b)(3)). Other authorities have their own rules, so check before you submit.
• Contracts you will sign. Use AI to understand the document, so you know what to ask. Let a lawyer or professional translator confirm the version that counts.
• Medical, safety and product instructions where a wrong word can hurt someone.

Even then, the AI draft saves time: you arrive at the professional with your questions ready.

A worked example

Say you run a small webshop and want your two-page returns policy in Spanish. Here is how the steps play out:

1. Prep: the policy is a Word file. Glossary: the shop name stays as is, “store credit” becomes one fixed Spanish term you choose, the 30-day window and the address stay exactly the same.
2. Prompt: reader is “Spanish customers who want to send something back and are a bit annoyed”, tone “calm, friendly, informal tú form”.
3. Result: the model returns the policy plus two [CHECK] marks, for example on a phrase it could translate in two ways. You pick one for each.
4. Check: you compare the numbers, back-translate the refund paragraph, and ask a Spanish-speaking friend or customer to read it once.
5. Publish: paste it into your shop. Save the prompt and glossary, so the shipping page next week sounds the same.

Common mistakes

• Pasting a long document with no instructions. You get a translation for nobody in particular, often slightly shortened.
• Trusting fluency. Smooth sentences are not the same as correct ones. Check the facts.
• A new chat for every chunk. Your glossary and tone get lost, and terms drift. Keep one chat per document.
• Forgetting the reader’s country. Spanish for Spain and for Mexico differ, so do Portuguese for Portugal and Brazil. Say which one in the prompt.
• Uploading client data without thinking. Check your data settings and your company rules first.

Frequently asked questions

Can ChatGPT translate documents?

Yes. Upload a Word file, PDF or text file and ask for a translation into the language you need. It is strongest on wording and tone. It gives you the text in the chat rather than a formatted copy, unless your plan lets it create files.

Is there a free way to translate a document with AI?

Yes. Google Translate’s Documents option is free for files up to 10 MB, a free DeepL account allows smaller Word, PowerPoint and PDF files, and ChatGPT’s free plan allows 3 file uploads per day. Limits change, so check the help pages linked above.

Can Claude translate a PDF?

Yes. Upload the PDF and use the same prompt. Claude reads text and visuals in PDFs up to 100 pages and text only from 101 to 1000 pages.

Is AI translation accurate enough for work?

For understanding a document and for most everyday business writing, it is usually a strong first draft. For anything official, legal or safety-related, treat it as a draft that a qualified human checks. The tools themselves tell you to verify important information.

Which is better for translation, ChatGPT or Claude?

Both do this well when you give them a clear reader, tone and glossary. Try the same page in both and keep the one that sounds more natural to you. Our comparison of ChatGPT vs Claude for writing covers the differences in style.

Legal wording is the part of a translation where mistakes cost the most. If you are working with contracts or similar paperwork, read how to use AI for legal documents safely before you rely on any translated clause.

Where to go from here

Translation is often step one. Once the text is in your language, you can let AI do the rest of the work: analyze a long PDF with Claude, summarize a long document for work, or make your reply email sound more professional before you send it back.

Today, take one real document from your inbox. Write a five-line glossary, run the prompt, check the numbers. You will know within ten minutes whether this saves you time.

If you want more small, practical wins like this one, using AI for the real tasks on your desk, start with the free Launchpad on our home page.

You have 40 products waiting for a description, the supplier text is stiff and half in capitals, and every time you ask ChatGPT for help you get “elevate your everyday” with no real detail. It feels like the AI is the problem. It usually isn’t.

Here is how to use AI to write product descriptions that actually sell: give it the facts first, not the product name. Fill in a short fact sheet (what it is, what it is made of, sizes, who it is for, what buyers keep asking), paste that into ChatGPT, Claude or Gemini with a prompt that sets your tone and length, then check every claim before you publish. The AI does the wording. You supply the truth.

That is the whole method. The rest of this page shows the fact sheet, the prompt, a worked example, how to do it from a photo or a spreadsheet, and what Google says about AI-written product text.

Why AI product descriptions all sound the same

When you type “write a product description for a linen apron”, the AI has nothing to go on except the words “linen” and “apron”. So it fills the gap with the safest, most average sentences it knows. That is where “crafted with care” and “perfect for any occasion” come from.

The fix is not a cleverer prompt. It is better input. A description that sells answers the questions a buyer has when they cannot touch the product: How big is it? What is it made of? Will it work for me? How do I look after it? If those answers are in your input, they end up in the output.

Step 1: Build a fact sheet (5 minutes per product)

Before you open any AI tool, write down the raw facts. Bullet points, no style, typos are fine. This is the part only you can do.

• What it is: the plain name, as a customer would say it.
• Specs: size, weight, materials, colours, capacity, what is in the box.
• Who it is for: one or two real types of buyer, and the moment they use it.
• Who it is NOT for: this cuts returns and builds trust.
• Customer words: phrases from your reviews, support emails or DMs. “Finally fits over my jumper” is better copy than anything an AI will invent.
• Care and practical stuff: washing, assembly, compatibility, delivery quirks.
• Claims you can prove: certifications, test results, warranty. Only the ones you have on paper.

Google’s own advice for shopping listings points the same way. Its Merchant Center guide says to be specific and accurate, to list the most important details in the first 160 to 500 characters, and to include things like size, material, intended age range and special features (Google Merchant Center Help: Description).

Step 2: The prompt to paste

Open ChatGPT, Claude or Gemini, paste this, and drop your fact sheet at the bottom. Change the bits in square brackets once, then reuse it for every product.

You write product descriptions for [shop name], which sells [what you sell] to [who buys].

Tone: [e.g. friendly and plain, like a shop owner talking to a regular]. No hype words such as "elevate", "game-changer", "must-have" or "perfect for any occasion".

Write a product description using ONLY the facts below. If something is not in the facts, do not add it. If a buyer would obviously want to know something that is missing, list it as a question at the end instead of guessing.

Format:
* One opening sentence that names the product and who it is for.
* A short paragraph (2 to 3 sentences) on the main benefit, in the buyer's words.
* 4 to 6 bullet points with the specs that matter most.
* One line on care or practical use.
Keep the first 2 sentences under 300 characters and put the most important detail there, it is what shoppers read first.

FACTS:
[paste your fact sheet here]

Two lines in that prompt do most of the work. “Using ONLY the facts below” stops the AI from inventing features. “List it as a question instead of guessing” turns the gaps into a to-do list for you, not a made-up spec on your shop.

Step 3: Push back once, then check every claim

The first draft is usually close, but not finished. Send one follow-up instead of rewriting it yourself:

Read this as a buyer who is on the fence. Which sentence would make them doubt it, and which question is still unanswered? Rewrite those parts. Then list every factual claim in the description next to the fact it came from.

That last instruction is your safety check. If a claim has no matching fact, delete it. This matters more than style. In the US, the FTC’s guidance for small businesses is clear that ads must be truthful and that you need evidence for your claims before the ad runs, including claims that are only implied (FTC: Advertising FAQs, a guide for small business). A product page counts. “Keeps coffee hot for 12 hours” needs to be true for your product, not just plausible to an AI.

A worked example (made-up product)

Here is the method on an invented product, so you can see what changes. The fact sheet:

Linen cross-back apron. 100% washed linen, mid weight. One size, adjustable by how you cross the straps. Two front pockets, one fits a phone. Colours: oat, charcoal. Buyers: home bakers, potters, café staff. Not for heavy grease work (linen stains). Reviews say: "no neck strap, finally", "fits over a jumper". Machine wash 40°C, line dry, gets softer with washing.

What a bare prompt (“write a description for a linen apron”) tends to give you:

Elevate your kitchen with our premium linen apron, crafted with care for the modern home chef. Perfect for any occasion, this must-have piece combines style and function.

What the fact-sheet prompt gives you, after one push-back:

A washed linen apron for home bakers, potters and café staff who are tired of a strap pulling on their neck.

The straps cross at the back instead, so the weight sits on your shoulders, and there is room to wear it over a jumper. It gets softer every time you wash it.

* 100% washed linen, mid weight
* One size, adjusts by how you cross the straps
* Two front pockets, one fits a phone
* Oat or charcoal
* Machine wash at 40°C, line dry

Not the best pick for heavy grease work: linen shows oil stains.

Questions for you: What are the exact length and width? Is the linen certified (for example OEKO-TEX)?

Notice the questions at the end. The AI did not invent a size or a certificate. It asked. You fill those in, delete the question line, and publish.

Write a product description from a photo

No fact sheet yet, just a picture? You can upload a product photo to ChatGPT (PNG, JPEG or non-animated GIF, up to 20MB) and ask it to describe what it sees. Claude and Gemini accept images too.

Look at this product photo. List what you can actually see: shape, colour, materials that are visible, parts, details. Mark anything you are guessing with (guess). Do not write a description yet.

Use that list as the start of your fact sheet, then fix the guesses. OpenAI’s own help page warns that the model may generate incorrect descriptions of an image in some cases and that it does not read the original file name or metadata (OpenAI Help: ChatGPT image inputs FAQ). So a photo can show “brushed metal”, but only you know whether it is stainless steel or coated aluminium.

Write descriptions in bulk from a spreadsheet

Once the prompt works for three or four products, do the rest in batches. Most shop systems can export products to a CSV file.

1. Export your products. Keep one row per product with columns for name, specs, materials, audience and notes.
2. Paste 10 rows at a time into the chat with your prompt and this extra line: “Write one description per row. Keep the product name as a heading so I can match them.”
3. Read every description. Batches are where small mix-ups happen, such as one product’s size landing in another product’s text.
4. Paste the results back into a new column and re-import.

Why 10 and not 200? Smaller batches are easier to check properly, and checking is the step you cannot skip. If you need help cleaning up the spreadsheet first, the same approach as using AI to write Excel formulas works for splitting and merging spec columns. When you are ready to run this without copying and pasting at all, an automation tool can loop through the rows for you; the n8n tutorial for beginners shows how a first workflow like that is set up.

Using the AI built into your shop

If you sell on Shopify, you may not need a separate chat window. Shopify’s help pages say its assistant, Sidekick, can write text inside the pages you edit, including product descriptions (Shopify Help: AI writing in your Shopify admin).

The same rule applies there. A built-in tool only knows what is on the product page. Fill in your specs and notes first, and give it your tone, or you get the same generic copy, just faster.

Will Google penalise AI-written product descriptions?

Not for being written with AI. Google’s Search team has said that appropriate use of AI or automation is not against its guidelines, and that what it rewards is helpful, original content, however it is produced (Google Search Central: guidance about AI-generated content).

What does get sites in trouble is volume without value. Google’s spam policies name “scaled content abuse”, including using generative AI tools to generate many pages without adding value for users (Google Search Central: spam policies). Pasting 500 product names into a tool and publishing whatever comes out is close to that line. Feeding it real facts and checking the result is not.

A few practical points if your products also go to Google Shopping:

• The description can be up to 5,000 characters, but the first 160 to 500 characters matter most.
• Google asks you not to include comparisons with other products, such as “even better than X”.
• Merchant Center has an optional structured description field where you can mark that the text was created with generative AI.

All three come from the same Merchant Center description guide.

Common mistakes to avoid

• Starting from the product name only. No facts in, no facts out.
• Letting the AI fill in numbers. Battery life, dimensions and “lasts for years” must come from you.
• Health or safety claims. “Eases back pain” or “non-toxic” need proof. If you don’t have it, cut it.
• Copying the supplier text into the prompt and asking for a rewrite. You end up with the same facts in slightly different words, while your own knowledge (customer questions, what gets returned) stays out of it.
• One tone for everything. Write the tone line once, save the prompt, and reuse it so the shop sounds like one person.
• Skipping the read-through. Read each description as a buyer. If you would not trust it, they won’t either.

Frequently asked questions

Can I use AI to write product descriptions?

Yes. Google does not penalise text for being written with AI, and every major chat tool can do it. The part you keep is the facts and the final check, because you are responsible for what the page claims.

Which AI is best for product descriptions?

ChatGPT, Claude and Gemini all handle this well when the input is good. The difference is mostly tone, so try the same fact sheet in two of them and keep the one that sounds more like you. Our comparison of ChatGPT vs Claude for writing goes deeper on that.

How long should a product description be?

As long as it takes to answer the buyer’s questions, with the most important details at the top. For a simple product that is often a short paragraph and a few bullets. Google’s Merchant Center guide puts the key details in the first 160 to 500 characters.

Can AI write product descriptions from a picture?

It can describe what it sees, which is a good start. It cannot see materials, sizes or certificates for sure, so treat the photo as the start of your fact sheet, not the finished text.

Should I tell customers a description was written with AI?

Rules differ by country and marketplace, so check the ones where you sell. Google Merchant Center already lets you mark AI-generated descriptions in its structured description field. Either way, the facts on the page must be right.

Once your product pages read well, the rest of the site deserves the same care. Here is how to write website copy with AI for your home page, about page and service pages.

Your next step

Pick your three best-selling products today. Write the fact sheets, run the prompt, push back once, check the claims. If the new versions read better than what is live, do the next ten tomorrow.

If you want more small, practical wins like this one, using AI for the real tasks on your desk, start with the free Launchpad on our home page.

It’s Thursday afternoon. The newsletter goes out tomorrow at 9, you have four half-ideas in a notes file, and every time you open a blank draft you end up answering Slack instead. You know AI can write. You also know the last AI email you received read like a brochure, and you deleted it.

Here’s the short answer. To use AI to write a newsletter, don’t ask it to “write a newsletter about X”. Give it your raw material first (notes, links, a customer question, what happened this month), plus two of your past issues so it can copy your voice, and ask for a draft in a fixed structure: one main story, two short items, one clear ask. Then you edit for 15 minutes: check every fact, cut the fluff, and add one line only you could write. AI does the first 80 percent. You do the part that makes people open the next one.

Want one practical AI workflow like this every week, free?

Can AI write a newsletter people actually read?

Yes, if you feed it. No, if you don’t.

A language model with no input writes the average of every newsletter it has seen. That average is polite, generic and forgettable. Readers notice. In usability research on newsletters, Nielsen Norman Group found that people now call even newsletters they signed up for “spam” when the content feels random, impersonal or full of promotional hype. Most of them don’t unsubscribe. They just stop opening.

So the risk with AI isn’t that it writes badly. It’s that it writes blandly, and bland is exactly what gets ignored. The fix is simple: the AI writes the sentences, but the substance (the news, the opinion, the example from your week) comes from you.

Google takes the same view for content in general. Its guidance on AI-generated content says it rewards helpful, original content however it is produced. The tool doesn’t matter. Whether the result is useful does.

Step 1: Collect your raw material (10 minutes, all month)

The best AI newsletter drafts start long before you open ChatGPT. Keep one running note, in your phone or a doc, and drop things into it as they happen:

• A question a client or customer asked you this month.
• A link you read and had an opinion about.
• A small win or a mistake at work, with one number in it (“we cut proposal time from 3 hours to 1”).
• Anything new: a product, a price change, an event date.
• One sentence on what you want readers to do after reading.

Messy is fine. “client asked if AI can do invoices, said yes but showed n8n, she was surprised it took 20 min” is perfect input. That one line has a real question, a real tool and a real reaction. AI can turn it into a strong paragraph. It can’t invent it.

If your raw material is a long report or a meeting transcript, shrink it first. Our guide on summarizing long documents with ChatGPT shows how to pull the three points worth sharing out of 40 pages.

Step 2: Teach the AI your voice once

This is the step most people skip, and it’s why their AI newsletter sounds like everyone else’s.

Paste in two or three of your past issues (or, if you have none yet, a few emails you wrote that you liked) and ask the AI to describe your style. Anthropic’s own prompting best practices say clear instructions plus a few well-chosen examples are one of the most reliable ways to get the output you want. Your old issues are those examples.

Below are 3 newsletters I wrote. Describe my writing style in 8 short bullet points: sentence length, tone, how I open, how I close, words I use often, words I never use, how I use humor, how long each section is.

Then write a "style guide" paragraph I can paste at the top of future prompts so you write like me.

[paste newsletter 1]
[paste newsletter 2]
[paste newsletter 3]

Read the bullets. Correct anything that’s wrong (“I never use exclamation marks”, “I always end with a question”). Then save the style guide somewhere you can reuse it:

• Claude: create a Project for your newsletter and put the style guide and your past issues in it. Claude’s Help Center explains how Projects keep files and custom instructions together, so every new chat in that project starts with them.
• Gemini: make a Gem with the style guide as its instructions. Google’s help page on using Gems in Gemini Apps covers the setup.
• ChatGPT: a Project or a custom GPT does the same job. Or just keep the style guide in a note and paste it at the top of each prompt. Low-tech, works everywhere.

You do this once. Every issue after that starts in your voice instead of the machine’s.

Step 3: Ask for a draft in a fixed structure

Newsletters that people keep reading usually have a shape readers can rely on. Pick yours and tell the AI to stick to it. A format that works well for professional newsletters:

1. Opening line: one sentence about something real from your week.
2. Main story: 150 to 250 words on one idea, with one concrete example.
3. Two quick items: a link with your one-line take, a tip, or a short update.
4. One ask: reply, book a call, read a post, register. Only one.

Here is the prompt. Swap the brackets for your own material:

[paste your style guide]

Write this month's newsletter for [who reads it, e.g. "owners of small accounting firms"].

Use ONLY the raw material below. Do not add facts, numbers, quotes or links that are not in it. If something is missing, write [CHECK] instead of guessing.

Structure:
* Opening: 1 sentence, about something real from the notes
* Main story: 150 to 250 words, one idea, one concrete example
* Two quick items: 2 to 3 sentences each
* One call to action: [what you want them to do]
* Sign off as [your name]

Also give me 5 subject lines under 50 characters and a 1-sentence preview text.

Raw material:
[paste your notes]

Two lines in that prompt do most of the work. “Use only the raw material” keeps the draft about your actual month. “Write [CHECK] instead of guessing” gives the model permission not to make things up, which is one of the techniques Anthropic lists in its guide to reducing hallucinations.

What a draft looks like (a real example)

Say you run a small bookkeeping practice and your notes say:

* 3 clients asked about the new VAT deadline
* tried AI to sort receipts for client M, saved maybe 2 hrs, still checked every line
* office closed 24 Dec to 2 Jan
* want people to book their year-end review before 15 Dec

A good AI draft of the main story will open with the VAT question three clients asked, explain what changed in plain words (with a [CHECK] where the exact date should go, because the notes don’t contain it), then use the receipt experiment as the example: what you tried, the two hours saved, and why you still checked every line. The quick items become the holiday closing and the year-end review. The ask is the review booking.

Nothing in that draft is invented. It just turned four messy lines into something readable in under a minute. That is the job.

Step 4: Edit like the human you are (15 minutes)

Never send the first draft. Go through it with this checklist:

• Every [CHECK]: fill it in from a real source, or delete the sentence.
• Every fact and number: does it match your notes? AI sometimes rounds, merges or “improves” numbers.
• Every link: open it. If the AI wrote a link you didn’t give it, remove it.
• The fluff test: delete any sentence that would fit in anyone’s newsletter. “Exciting times ahead” goes. “We closed 40 year-end files in November, a record for us” stays.
• Add one line only you could write: an opinion, a small story, a reaction from a client. This is the line people reply to.
• Read it on your phone: short paragraphs, nothing wider than the screen, the ask easy to spot.

If a section sounds stiff, paste just that paragraph back and ask: “Rewrite this the way I’d say it to a client over coffee. Keep every fact.” Our page on making an email sound more professional has the reverse prompts, for when a draft is too casual.

Subject lines and preview text

The subject line decides whether any of your work gets read. Ask the AI for five or more options, then pick with two rules:

• It must be true. Clickbait that the email doesn’t deliver on is exactly what readers call spam. In the US it can also break the law: the CAN-SPAM Act (15 U.S.C. 7704) prohibits deceptive subject lines in commercial email.
• It should name the main story. “Your year-end review: book before 15 Dec” beats “November update” every time, because the reader knows what’s inside.

Use the preview text (the grey line after the subject in most inboxes) to add the second reason to open, not to repeat the subject.

Which AI is best for writing a newsletter?

Honestly, the tool matters less than the input. ChatGPT, Claude and Gemini all write a solid newsletter draft when you give them real notes and a style guide. A few practical differences:

• Claude is a good fit if you want to keep a big folder of past issues and reference material in a Project.
• Gemini is handy if your notes and drafts already live in Google Docs and Gmail.
• ChatGPT is the most common choice and works fine with the prompts above.
• Your email platform’s built-in AI (most newsletter tools now have one) is convenient for subject lines and small rewrites inside the editor. For the main draft, a full chat assistant with your style guide usually gives you more control.

If you’re still choosing a general assistant, our overview of an AI writing assistant for busy professionals compares them for everyday work writing.

Turn one issue into more content

Once the newsletter is done, you have the hardest part of a week’s content already written. Ask the AI:

From the newsletter below, write:
* 1 LinkedIn post (under 150 words) based on the main story
* 3 short social posts, one per idea
Keep my voice. No hashtags unless I add them. No facts that are not in the newsletter.

[paste final newsletter]

For more on this, see our ChatGPT prompts for social media managers. And if the newsletter announces something big, like a launch or a new hire, the same raw material can feed a press release written with AI.

Automate the boring part (optional)

When the writing routine works, you can automate the collection step. A simple setup: forward any email or link worth sharing to one label or folder during the month. On the day you write, export or copy that folder into your notes and run the prompt. If you want to go further, our list of email automation examples shows how filters, rules and tools like Zapier can do the sorting for you.

Keep the human steps human, though. Automating the gathering saves you time. Automating the sending of an unchecked AI draft to your whole list is how a wrong date reaches 2,000 people.

Mistakes that make an AI newsletter obvious

• No raw material. “Write a newsletter about productivity” gets you the average of the internet.
• Too many topics. Five stories of equal weight means none of them land. One main story, two small ones.
• The same opening every month. “I hope this finds you well” in issue after issue tells readers nobody is behind it. Change the first line every time, from something real.
• Several calls to action. Read this, book that, follow us, reply. Pick one.
• Pasting private data. Don’t put client names, contracts or personal details into a chat tool your company hasn’t approved. Anonymise first (“client M”, “a firm in Utrecht”).
• Forgetting the basics of email rules. Marketing newsletters need consent and an easy way to unsubscribe. In the UK, the ICO’s guide to electronic mail marketing explains the consent rules. Your newsletter platform handles most of this, but the list and the content are your responsibility, not the AI’s.

FAQ

Is it okay to use AI to write my newsletter?

Yes. Readers care whether it’s useful and sounds like you. Use AI for structure and first drafts, keep the facts, opinions and examples yours, and check everything before you send.

How long should a newsletter be?

Long enough for one main idea, done well. For most professional newsletters that’s somewhere around 300 to 600 words. If your readers open it on their phone between meetings, shorter usually wins.

Can I use ChatGPT to write a newsletter for free?

Yes, the free versions of ChatGPT, Claude and Gemini can all draft a newsletter with the prompts on this page. Paid plans mostly add higher limits and features like Projects, which make saving your style guide easier.

How do I stop my AI newsletter from sounding like AI?

Three things: give it your past issues as examples, give it real notes instead of a topic, and add one line yourself that only you could write. Then cut every sentence that would fit in anyone else’s newsletter.

Should I tell readers I use AI?

That’s your call and depends on your field. What matters most is that every fact is checked and the opinions are genuinely yours. If you’d be uncomfortable explaining how it was written, that’s a sign to put more of yourself in it.

A good newsletter issue often has two or three ideas that work as short posts too. If LinkedIn is where your readers are, here is how to use AI to write LinkedIn posts that sound like you.

Some newsletter ideas deserve more room than one issue. When a topic keeps growing, turn it into a full article with our guide on how to use AI to write a blog post that still sounds like you.

If you want a weekly rhythm that turns one idea into a newsletter, social posts and a short script, not just one format at a time, see how to use AI for content creation. It covers the voice file, the main piece, and how to split that piece across the week.

Your next step

Open your notes app today and start the running list. Next week, run the style prompt on your last three emails. The week after, you’ll write your newsletter in 30 minutes instead of an afternoon.

We share one workflow like this every week, built for people who want to use AI at work without becoming engineers. Start on our home page, or sign up right here:

The project went well. The client said so on the last call. And the case study you promised yourself you’d write has now been sitting on your to-do list for three weeks, because every time you open a blank document you don’t know where to start.

Here’s the short answer. To write a case study with AI, don’t ask ChatGPT or Claude to write one from nothing. Give it the real material first: a client interview transcript, the before and after numbers the client agreed to, and your own project notes. Ask the AI to pull out the challenge, the solution, the results and the best direct quotes, then draft it in a simple Challenge, Solution, Results shape. You check every number and every quote, and the client approves it before it goes anywhere. The AI does the sorting and the first draft. You do the truth.

Want one practical AI workflow like this every week, free?

Why case studies stall, and what AI actually fixes

Most case studies don’t fail because people can’t write. They fail at three points:

• The blank page. You know the project was good, but turning a messy three-month engagement into a clean story feels like a whole afternoon.
• Scattered material. The results are in a spreadsheet, the nice thing the client said is in a Slack message, and the timeline lives in your head.
• The approval loop. You write it, send it to the client, and it waits in their inbox.

AI is genuinely good at the first two. It can read a 40-minute interview transcript and hand you the five strongest quotes in a minute. It can take bullet points and turn them into readable paragraphs. What it can’t do is know what happened. If you give it gaps, it fills them with confident, plausible sentences that may be false. That’s the one rule for this whole process: the AI only works with facts you give it.

Using AI to write is not a problem for search either. Google’s own guidance says it rewards helpful, original content however it is produced, and that using automation mainly to manipulate rankings is what breaks its rules (Google Search Central). A case study built on a real client and real results is about as original as content gets.

Step 1: Collect the raw material (15 minutes)

Before you open any AI tool, put these in one document or folder:

• The client interview. A 20 to 30 minute call is enough. Record it with the client’s permission and use the built-in transcript. Microsoft Teams can produce a live transcript during the meeting (Microsoft Support), and most other meeting tools have something similar.
• Before and after numbers. Hours spent per week, response times, error rates, revenue, whatever you actually measured. Only numbers the client knows about and agrees with.
• Your project notes. What you did, in what order, which tools, what went wrong and how you fixed it.
• Anything the client already said in writing. A thank-you email or a message after go-live is often the best quote you’ll get.

If you already turn meetings into notes with AI, the same habit works here. Our guide on summarizing meeting notes with ChatGPT covers that part.

What not to paste: client financials they haven’t agreed to share, personal data of their customers, or contract details. Check your AI tool’s data settings before you upload anything sensitive. ChatGPT, for example, lets you switch off “Improve the model for everyone” in its data controls (OpenAI Help Center). If the client hasn’t agreed to be named yet, replace their name with “the client” in what you paste.

Step 2: Let AI prepare the client interview

A good case study depends on good answers, and good answers depend on good questions. Most people ask “so, how did it go?” and get “great, really happy”. Ask AI to help you dig for the specific stuff.

I'm writing a customer case study about a project I did for a client.
Project: [one or two lines on what you did]
Client type: [e.g. 6-person design studio]

Write 8 interview questions for a 25-minute call with the client.
* Cover: the situation before, what they had tried, why they picked us, what changed, what they would tell a peer.
* Ask for specifics: numbers, moments, examples. No yes/no questions.
* Keep each question short enough to read out loud.

Questions that tend to pull out the best material sound like this:

• “What was a normal Monday like before we started?”
• “What had you already tried, and why didn’t it stick?”
• “Was there a moment you noticed it was working? What happened?”
• “If a friend in your industry asked about this, what would you tell them?”

That last one is gold. People answer it in their own words, and it often becomes the headline quote.

Step 3: Extract the story from your material

Now paste the transcript and your notes into ChatGPT, Claude or Gemini and ask it to sort, not write. This is where you save the most time.

Below is a client interview transcript and my project notes.
Use ONLY what is in this text. Do not add facts, numbers or quotes.

Give me:
* The challenge: what was going wrong before, in the client's words where possible
* What they tried before
* Why they chose to work with me
* What I did, step by step
* Results: every number mentioned, exactly as stated, with who said it
* The 5 strongest direct quotes, word for word, with the timestamp
* Missing information: questions I should ask the client to fill the gaps

[paste transcript]
[paste notes]

The last line matters most. Instead of guessing, the AI tells you what is missing (“the transcript mentions faster invoicing but gives no number”). You send those questions to the client in one short email.

Claude handles long transcripts well, and if you write case studies regularly, a Claude Project lets you keep your template, your past case studies and your instructions in one place so every new one starts from the same base (Claude Help Center). We go deeper on that in how to use Claude for work.

Step 4: Draft it with a shape buyers can scan

People rarely read a case study top to bottom. They look for “is this company like mine?” and “what was the result?”. Nielsen Norman Group’s research on web reading found that most users scan pages rather than reading word by word (Nielsen Norman Group). So the shape does half the work:

1. Headline with the result: “How a 6-person studio stopped chasing invoices” beats “Case study: Studio X”.
2. Summary box: client, challenge, solution, result, in four short lines.
3. The challenge: the before picture, with a client quote.
4. The solution: what you did, in plain steps. No jargon.
5. The results: the numbers, bolded, plus one quote about how it feels now.
6. What’s next: one or two lines, and a way to contact you.

Then ask the AI to draft into that shape:

Using only the extracted facts above, write a customer case study.
Structure: headline with the main result, a 4-line summary box (Client, Challenge, Solution, Result), then sections Challenge, Solution, Results, What's next.

Rules:
* 600 to 900 words
* Use the client's quotes word for word. You may cut a quote shorter, never change its meaning
* Every number must come from the facts above. If a number is missing, write [NUMBER NEEDED]
* Plain, warm language. No buzzwords like "game-changer", "seamless" or "revolutionize"
* Write for a reader who runs a business like the client's

The [NUMBER NEEDED] rule is the trick. It stops the AI from inventing a “37% improvement” to make the paragraph sound better.

A worked example

This example is made up to show the shape. Your case study uses your client’s real words and numbers.

Say you’re a freelance automation consultant. You set up an invoice reminder workflow for a small design studio. In the interview, the owner says: “I used to spend Friday afternoons going through who hadn’t paid. Now I don’t think about it.” Your notes say the studio sends around 30 invoices a month and the owner estimated about three hours a week on chasing before.

A first AI draft from that material might open like this:

How a 6-person design studio stopped spending Friday afternoons chasing invoices

Client: independent design studio, 6 people
Challenge: about 3 hours a week spent checking and chasing unpaid invoices
Solution: an automated reminder workflow connected to their invoicing tool
Result: [NUMBER NEEDED: hours per week now]. "Now I don't think about it."

Notice what it did right: it used the owner’s exact words, kept the “about” in “about three hours”, and flagged the missing after-number instead of making one up. Your job is to get that number from the client, and to read the rest line by line.

Step 5: Fact-check, then get the client’s approval

This step is not optional, and it’s where your reputation lives. A case study is a public statement about a real business. If a quote is twisted or a number is inflated, the client sees it first.

There’s also a legal side. In the US, the FTC’s Endorsement Guides say endorsements must reflect the honest opinions and experience of the person giving them (FTC), and since October 2024 an FTC rule explicitly bans fake testimonials, including AI-generated ones (FTC press release). Rules differ by country, but the safe habit is the same everywhere: real quotes, real numbers, written approval.

Make the check fast with one more prompt:

Go through the case study draft and list every sentence that contains:
* a number
* a direct quote
* a claim about the client's business or results
For each one, show where it came from in the transcript or notes. If you can't find a source for it, say "NOT IN SOURCE".

Anything marked “NOT IN SOURCE” gets cut or confirmed with the client. Then send the draft with a short note: “Here’s the draft. Can you check the quotes and numbers, and let me know if you’re happy for us to use your name and logo?” Keep that reply email. It’s your written approval.

If the client says yes to the story but no to their name, you can still publish an anonymized version (“a 6-person design studio in Rotterdam”). It’s less strong, but it’s honest and still useful.

Step 6: Turn one case study into four useful pieces

Once it’s approved, the hard work is done. Ask the AI to reuse it:

From this approved case study, write:
1. A LinkedIn post (under 150 words) that tells the story from the client's problem, with one quote
2. A 3-sentence version I can paste into proposals
3. A short follow-up email to a prospect in the same industry
4. Text for one presentation slide: headline, 3 bullet points, 1 quote
Use only the facts and quotes in the case study.

The proposal snippet is the one people forget, and often the most valuable. If you write proposals with AI, our page on ChatGPT for proposal writing shows where proof like this fits. And if a client result is big enough for news, the same material can feed an AI-written press release.

Which AI tool should you use?

Honestly, any of the big ones will do this well, because the quality comes from your material, not the model. A few practical differences:

• ChatGPT: good all-rounder, easy for the extract and draft steps, and handy for the social and email versions.
• Claude: comfortable with long transcripts, and Projects keep your template and past case studies together.
• Gemini: practical if your transcripts and notes already live in Google Docs and Drive.
• Copilot: the natural choice if your company runs on Microsoft 365 and Teams transcripts.

What about a dedicated “AI case study generator”? Most of them are a prompt like the ones above with a form on top. They can be convenient, but they still need your real facts, and you still need to check the output. Try the prompts here first before paying for another subscription.

A finished case study is also strong proof for funding applications. If you are applying for one, see how to use AI to write a grant proposal to turn those results into a Statement of Need.

Common mistakes to avoid

• Asking AI to “write a case study about how I helped a client save time”. You’ll get a generic story with invented numbers. Always start from the transcript.
• Polishing quotes until they sound like marketing. Real people don’t say “leveraged a streamlined solution”. Keep their words. Trim, don’t rewrite.
• Burying the result. If the best number sits in paragraph six, most readers never see it. Put it in the headline and the summary box.
• Making it about you. The hero of a case study is the client. You are the guide who helped.
• Skipping approval because “they’ll be fine with it”. Most clients are. Ask anyway, in writing.

FAQ

Can AI write a case study for me?

It can write the draft, and do it well, if you give it real material: a transcript, notes and numbers. It can’t write a true case study from nothing, because it doesn’t know what happened. Treat it as a fast assistant that organizes and drafts, not as the source.

How long should a business case study be?

Long enough to show the before, the change and the result, and short enough to read in a few minutes. Around 600 to 900 words works for most service businesses, plus the summary box for people who only skim.

Do I need the client’s permission?

Yes. Get written approval for the quotes, the numbers and the use of their name and logo. Even if your contract doesn’t require it, it protects the relationship, and in many places it’s also the legally safer route.

What if the client won’t share numbers?

Use what they’re comfortable with. “Invoices now go out the same day the project closes” is a real result without a percentage. Ask the AI to rewrite the results section around concrete changes instead of metrics, and never let it guess numbers.

Can I use the same prompts for a UX or portfolio case study?

Mostly, yes. Swap the client interview for your own process notes, research findings and screenshots, and change the structure to problem, process, decisions and outcome. The same rule applies: only facts you provide.

A finished case study also gives you strong proof for your website. To turn those results into clear page copy, read how to write website copy with AI.

Your next 30 minutes

Pick the client project you’re proudest of from the last six months. Send them a short message asking for a 25-minute call. While you wait for a reply, run the interview question prompt from Step 2. That’s the case study started, and the hardest part, getting going, is behind you.

If you’d like to build more workflows like this one, from meeting notes to proposals, start on our home page or grab the free Launchpad below.

Get one practical AI workflow in your inbox every week.

When the case study feeds a client pitch, our guide to AI proposal writing software covers when chat tools are enough and when dedicated software helps.

You have real news. A new product, a second location, a partnership you worked on for months. Now someone says “send out a press release” and you’re staring at a blank page, wondering how the big companies make it sound so official. You don’t need an agency for this. You need your facts in order and a good first draft.

Here’s the short answer. To use AI to write a press release, first write a short fact sheet yourself (what happened, who, when, where, why it matters, one real quote, one contact). Paste it into ChatGPT, Claude or Gemini with a prompt that asks for the standard press release structure and forbids hype words. Then run two checks: a hype pass and a fact pass. The AI handles the format and the wording. You own every fact.

Want one practical AI workflow like this every week, free?

Why most AI press releases get ignored (and how to avoid it)

Press releases are not dead. In Cision’s 2025 State of the Media survey of more than 3,000 journalists, 72% called press releases the most useful resource PR teams can give them. The same survey found that 72% worry about factual errors in AI-generated PR content.

That’s the whole story in two numbers. Reporters want releases. They don’t trust sloppy AI ones. Among newspaper journalists, Cision reports that half are opposed to receiving AI-generated pitches or press releases.

Muck Rack’s 2025 State of Journalism survey points the same way: the top reasons journalists delete a pitch are that it’s irrelevant to them (86%) or overly promotional (71%). And what does a chatbot do when you ask it to “write an exciting press release”? It reaches for “revolutionary”, “game-changing” and “industry-leading”. Exactly the tone that gets deleted.

So the trick is not to ask AI for a press release. The trick is to give it facts and rules, and let it do the part it’s actually good at: structure, clear sentences and the right length.

Step 1: Write a 5-minute fact sheet (this is the part AI can’t do)

Before you open any chat, answer these in plain notes. Half sentences are fine.

• The news in one line: what is new, today, that wasn’t true yesterday?
• Who: your company, plus any partner, with names spelled correctly.
• When and where: the date of the announcement and the city it comes from.
• Why it matters, and to whom: the customer, the town, the industry. Be specific.
• Two or three facts with proof: numbers you can back up, a price, an opening date, how many people it affects.
• One quote in your own words: say out loud why you did this, record it on your phone and type what you said. Rough is fine.
• Media contact: a real name, email and phone number.
• Boilerplate: three sentences about your company that you’d be happy to see printed.

If you can’t fill in “why it matters, and to whom”, stop here. That’s not a writing problem, and no AI will fix it. Find the angle first, then write.

Step 2: Use the drafting prompt

Paste your fact sheet into a new chat, then paste this underneath. It works in ChatGPT, Claude, Gemini and Copilot.

Act as an experienced PR writer who used to be a journalist. Above is my fact sheet.

Write a press release using ONLY the facts above. Do not add numbers, names, dates or claims that are not in my notes. If something important is missing, ask me instead of guessing.

Structure:
* Headline: under 12 words, states the news plainly
* Subheadline: one sentence with the most useful detail
* Dateline: [CITY], [DATE]
* First paragraph: who, what, when, where and why, in 2 to 3 sentences
* Second paragraph: the most important supporting facts
* One quote, based on my words, kept sounding like a real person
* Practical details: price, date, address, link, availability
* About [company]: use my boilerplate
* Media contact
* End with ###

Rules:
* Keep it to one page
* No hype words: revolutionary, game-changing, cutting-edge, leading, innovative, excited to announce, thrilled
* Short sentences. Write for a busy reporter who reads only the first paragraph.

The line “ask me instead of guessing” is the most important one. Without it, many chatbots fill gaps with plausible numbers. With it, you’ll often get two or three good questions back, like “Is the opening date confirmed?” Answer them, and the draft gets better.

Step 3: Run the hype pass

Even with rules, adjectives sneak back in. Ask for a second look:

Read the press release again as a skeptical local reporter.
* List every word or sentence that sounds like marketing instead of news.
* For each one, suggest a plain replacement that uses a fact from my notes.
* Tell me which sentence a reporter would most likely copy into their story. If none, say so and tell me what is missing.

That last question is a good test. A reporter with little time often lifts the first paragraph and the quote. If neither works on its own, rewrite them before anything else.

Step 4: Run the fact pass (never skip this one)

This is where AI press releases go wrong, and where journalists lose trust. Ask the AI to show its work:

Make a table of every factual claim in the press release: names, numbers, dates, prices, places, titles.
For each claim, quote the line from my fact sheet it came from.
Mark any claim that is NOT in my fact sheet with "CHECK".

Then you go through the table yourself. Check the spelling of every name, every date against the calendar, every number against your own records. Anything marked “CHECK” either gets a source from you or gets deleted. Two minutes of checking protects your name with the reporter for the next five releases.

Step 5: Make the quote sound like a person

AI quotes all sound the same: “We are thrilled to bring this exciting solution to our valued customers.” Nobody talks like that, and a reporter has no reason to print it.

Use the voice memo from Step 1. Ask the AI to tighten what you actually said, not to invent a new quote:

Here is what I said out loud about why we did this: [paste your words].
Turn it into a quote of 2 to 3 sentences. Keep my words and my opinion. Remove filler words only. No new claims.

A good quote says something the facts can’t: a reason, a story, an opinion. “Half our customers drove 40 minutes to reach us. Now they don’t have to.” That’s a quote someone will print.

An example: before and after

A made-up example, so you can see the difference. A bakery opens a second shop. Here’s the kind of opening paragraph a chatbot writes when you just say “write a press release about our new bakery location”:

Sunrise Bakery, a leading artisanal bakery, is thrilled to announce the exciting launch of its innovative new location, bringing its passion for exceptional baked goods to even more valued customers.

And here’s what the same prompt gives you after the fact sheet, the rules and one hype pass:

AMERSFOORT, 14 November 2026. Sunrise Bakery opens its second shop on Langestraat on Saturday 21 November, three years after starting from a single oven in Leusden. The new shop bakes on site from 6:00 and adds a lunch counter with 20 seats.

The first version tells a reporter nothing. The second one answers who, what, when and where in two sentences, and they could paste it straight into a short local news item.

Step 6: Write the pitch email (shorter than you think)

A press release on its own doesn’t get coverage. The email you send it with does the work. In the same Muck Rack survey, 69% of journalists preferred pitches under 200 words, and half said one follow-up is fine.

Write a pitch email to [reporter name] at [outlet], who covers [beat]. Here are two of their recent articles: [titles or short summaries].
* Under 150 words
* First line: why this matters to THEIR readers, not to us
* Subject line: plain, no capitals, no clickbait
* Mention the press release is pasted below the email, not attached
* Offer one concrete thing: an interview, photos, or a visit

Notice the prompt asks you to name the reporter and their recent stories. That’s the relevance part, and irrelevance is the top reason journalists give for deleting a pitch. If you write outreach emails often, our guide to ChatGPT prompts for cold email outreach goes deeper on the first email and the follow-up.

Which AI is best for writing a press release?

Honestly, the prompt matters more than the tool. All the main assistants can follow the structure above. Some practical differences:

• ChatGPT is quick and handles the structure well, but tends to add enthusiasm. Your hype pass will catch it.
• Claude often writes a calmer, more natural tone, which helps with quotes.
• Gemini or Copilot are handy if your notes already live in Google Docs or Word.

If you’re choosing between them for writing in general, we compared ChatGPT vs Claude for writing side by side.

What about dedicated “AI press release generators”? Many are a form with fields (headline, quote, boilerplate) on top of the same kind of model. They’re fine if you like filling in boxes, but the fact sheet and the two passes on this page do the same job in the chat tool you already use.

Where to send it, and one SEO trap

• Your own website. Put the release on a news page on your site. It’s the version you control, and you can link to it from your pitch.
• Directly to the right reporters. Five well-chosen local or trade reporters beat a list of 500 random ones.
• A wire service, if it fits. Useful for companies that need broad distribution, but it costs money and doesn’t replace a personal pitch.

The trap: some people stuff press releases with keyword links to push their Google rankings. Google’s spam policies list links with optimized anchor text in press releases distributed on other sites as an example of link spam. Keep links in a distributed release plain and few, like your company name and homepage. Write the release for reporters, not for search engines.

What not to paste into an AI chat

Press releases often deal with news that isn’t public yet: an acquisition, a funding round, a launch date. Before you paste anything confidential, check how your AI account handles your data. OpenAI explains the options, including turning off training on your chats, in its data controls FAQ, and business plans usually have stricter settings. If your company has an AI policy, follow it. When in doubt, use placeholders like [PARTNER] and [AMOUNT] in the chat and fill them in yourself afterwards.

Common mistakes

• No news. “We updated our website” is not a press release. Ask the AI: “Would a local reporter care about this? Be honest.” It’s a useful gut check.
• Burying the news. If the first paragraph is about your company history, the reporter has already moved on.
• Trusting the numbers. Every figure in the release must come from you. The fact pass exists for this.
• The same release to everyone. Keep one release, but change the pitch email for each reporter.
• Forgetting the contact. A reporter on deadline who can’t reach you will write a different story.

And when the release is out and a reporter replies, answer in your own voice. If you want help polishing a reply, these prompts to make an email sound more professional keep it friendly without the stiff tone.

FAQ

Can AI write a press release for me?

It can write a solid draft in a few minutes, and it’s good at the standard format. It can’t know your news, your numbers or why it matters to a specific reader. Give it a fact sheet, and check every fact before you send anything.

Is there a free AI press release generator?

You don’t need a separate tool. The free versions of ChatGPT, Claude and Gemini can all follow the drafting prompt on this page. Paid generators mostly add templates and distribution, and distribution is the part worth comparing if you need it.

Will journalists know it was written with AI?

They notice the signs: hype words, vague claims, quotes nobody would say out loud, and factual errors. A release built from your own facts, your own quote and a hype pass reads like a careful human wrote it, because a careful human checked every line.

How long should a press release be?

Aim for one page. The news belongs in the first paragraph, so a reporter who reads only that still gets the story. Everything after it is supporting detail.

Should I attach the press release or paste it?

Paste it into the email below your short pitch, and link to the version on your website. Many people prefer not to open attachments from senders they don’t know.

Write your first one this week

Pick a piece of real news, even a small one. Write the fact sheet, paste the drafting prompt, run the hype pass and the fact pass, and send it to three reporters who actually cover your area. That’s an hour of work, and most of it is thinking, which is exactly the part that should stay yours.

If your news is a client win, the same facts can also become a story for your own site. Here is how to write a case study with AI.

And if you share news like this with your list every month, here is how to use AI to write a newsletter people actually read.

Want more workflows like this, one practical win at a time? Start on our home page, or join Launchpad right here:

It’s 8:40. Your inbox already has 31 unread emails, there’s a call at 10, a report due by end of day, and a list in your head that keeps getting longer while you stare at it. You don’t need another productivity app. You need someone to look at the mess and say “do this first”.

Here’s the short answer. To use AI to plan your day, dump every task into ChatGPT, Claude, Gemini or Copilot in three minutes, give it your fixed meetings and your real working hours, and ask it to build a time-blocked plan with only three must-do items, a time estimate on every block and buffer between them. Then you push back once: cut what doesn’t fit and move the rest to tomorrow. The AI does the sorting. You make the calls.

Want one practical AI workflow like this every week, free?

Why your to-do list isn’t a plan (and what AI fixes)

A to-do list tells you what exists. A plan tells you what happens at 9:30. The gap between those two is where most days go wrong.

Two pieces of research explain why. First, we are bad at guessing how long our own work takes. In the classic “planning fallacy” studies by Buehler, Griffin and Ross (1994), fewer than half of the participants finished their tasks in the time they predicted. The optimism disappeared when people were asked to connect the estimate to similar past tasks. That is exactly the question an AI can ask you every morning.

Second, open tasks keep pulling at your attention. Masicampo and Baumeister (2011) found that unfinished goals caused intrusive thoughts during an unrelated task, and that making a specific plan for those goals removed that effect. A written plan with a time slot for the report is not just tidier. It lets you stop thinking about the report while you’re on the 10 o’clock call.

So the job for AI is clear: turn the pile into a sequence, put honest times on it, and make the trade-offs visible before 9 instead of at 4.

Step 1: Do a 3-minute brain dump

Open a new chat and type everything that’s on your mind for today. Don’t sort, don’t format. Half sentences are fine. Include the small stuff (“reply to Sanne about the invoice”) and the big stuff (“Q4 report, still need numbers from finance”).

Add three facts the AI can’t guess:

• Your fixed points: meetings, calls, school pickup, anything with a time on it.
• Your real hours: not 9 to 5 on paper, but when you actually start and stop.
• Your energy pattern: are you sharp in the morning and useless after lunch, or the other way round?

Speaking it out loud works too. Most AI apps on your phone let you dictate, and talking for two minutes on the way to your desk is often faster than typing.

Step 2: Use the planning prompt

Paste this under your brain dump. It’s built to stop the two classic problems: a plan that’s too full, and estimates that are too optimistic.

Act as my no-nonsense planning assistant. Above is everything on my mind for today.

My working hours today: [08:30 to 17:30]
Fixed appointments: [10:00 to 10:45 client call, 14:00 to 14:30 team stand-up]
My best focus time: [mornings]

Build a time-blocked plan for today:
* Pick a maximum of 3 must-do items. Tell me why you chose them.
* Put my hardest task in my best focus time.
* Give every block a time estimate, and add 25% to anything I have never done before.
* Leave a 10 to 15 minute buffer after every meeting.
* Batch small tasks (emails, quick replies, admin) into one or two blocks.
* Keep at least 60 minutes unplanned for surprises.
* Anything that doesn't fit goes into a "Tomorrow or later" list. Don't squeeze it in.

Before you plan, ask me up to 3 short questions if something is unclear (deadlines, who is waiting on what).

The last line matters. A good assistant asks “is the report due to your boss or to the client?” before it guesses. If the AI skips the questions and plans straight away, that’s fine too. Just read the plan with that in mind.

Step 3: Push back once (the 2-minute reality check)

The first plan always looks good. That’s the problem. Before you accept it, ask the question from the planning fallacy research:

Here is how long similar tasks took me last time:
* [the monthly report took about 3 hours, not 2]
* [inbox clean-up usually takes 45 minutes]

Redo the time estimates with this in mind. If the day no longer fits, tell me which item to move and what the cost is of moving it.

This is where AI earns its place. It has no ego about the plan, so it will tell you plainly that four big items don’t fit in a day with three meetings. You decide what moves. Sometimes the honest answer is a quick message to someone: “I’ll have this to you tomorrow at 11.” The AI can draft that message too.

An example: a messy Tuesday, before and after

The brain dump (as typed):

q4 report needs finance numbers, call client 10am, prep slides for thursday, reply to 3 supplier mails, expense claims, check new hire contract, stand-up 2pm, gym??, call mum back

The plan the AI returned (shortened):

Must-do today: 1) Q4 report draft  2) New hire contract check  3) Client call

08:30 to 08:45  Email finance for the Q4 numbers (unblocks the report)
08:45 to 09:45  Q4 report: write structure and text, leave number gaps
09:45 to 10:00  Buffer + prep notes for the client call
10:00 to 10:45  Client call
10:45 to 11:00  Buffer: write call follow-ups
11:00 to 11:45  New hire contract check
11:45 to 12:15  Supplier emails (3) + expense claims
12:15 to 13:00  Lunch
13:00 to 14:00  Open slot for surprises / finance numbers into report
14:00 to 14:30  Stand-up
14:30 to 15:45  Q4 report: finish and send
15:45 to 17:30  Open (gym at 17:30 if the report is sent)

Tomorrow or later: Thursday slides (start Wednesday morning), call mum back during lunch or this evening

Look at what changed. The first task is a two-minute email that unblocks the biggest job. The slides moved to tomorrow with a reason. The vague “gym??” became a reward tied to finishing the report. Nothing here is clever. It’s just sorted, which is the part that’s hard to do when you’re the one drowning.

Let AI see your calendar (so you don’t type it)

Typing your meetings every morning gets old fast. The big assistants can read your calendar directly once you connect it:

• Claude: the Google Workspace connectors let Claude look at your Google Calendar, including events and availability, when you ask about your schedule. On Team and Enterprise plans an admin has to switch it on first. If you want the full setup, our guide on how to use Claude for work walks through connecting email, calendar and files.
• Gemini: Google’s help page explains how to create and manage calendar events with Gemini. Add @Google Calendar to your prompt to point it at your calendar. Keep Activity has to be on for the connection to work.
• Copilot in Outlook: Microsoft’s chat-assisted scheduling lets you schedule meetings and block time straight from the Copilot chat in Outlook.

With a connected calendar, your prompt shrinks to: “Look at my calendar for today, here’s my brain dump, build the plan using the rules from yesterday.” One tip: let the AI suggest focus blocks, but put them in your calendar yourself the first few weeks. You’ll learn fast which estimates you trust.

Make it a morning habit with a scheduled task

The best plan is the one you actually make every day. ChatGPT can remind you. Its scheduled tasks run a prompt at a time you choose, one-time or recurring. On the free plan, recurring tasks can run at most once a day, which is exactly what you need here.

Ask it something like:

Every weekday at 08:15, ask me for my brain dump for today. When I answer, build my plan with these rules: max 3 must-dos, hardest task in the morning, 25% extra on new tasks, buffer after meetings, 60 minutes unplanned, the rest to "Tomorrow or later".

Now the habit runs itself. You get a nudge, you talk for two minutes, you get a plan.

When the day breaks at 11: replan in 60 seconds

Every plan meets reality. A client calls, a colleague is sick, the finance numbers don’t arrive. The old way is to abandon the plan and run on panic. The better way is to go back to the same chat:

It's 11:20. I lost 90 minutes to an urgent client issue. Done so far: [finance email, client call]. Replan the rest of my day with the same rules. Tell me what drops off and draft a short message for anyone who is waiting on something I'm moving.

Because the chat already knows your morning, the new plan takes a few seconds. This is the part most planner apps don’t do well, and it’s where talking to an AI feels like having a calm colleague next to you.

End the day with a 3-line review

At the end of the day, give the AI three lines: what got done, what didn’t, and how long the big task really took. Ask it to carry the unfinished items into tomorrow’s list. After a week, you have exactly what the planning fallacy research says helps: your own track record. Your estimates get better because they’re based on what happened, not on what you hoped.

If you track your work for your manager or team, the same notes feed nicely into a weekly update. We showed that in using Gemini for automated status updates.

What not to paste into an AI planner

Your day plan often contains names, client details and sometimes health or HR matters. A few simple habits:

• Use first names or roles (“new hire”, “client A”) instead of full names and details you’d never put in an email to a stranger.
• If you use ChatGPT for work, check its data controls to decide whether your chats may be used to improve models.
• If your company provides Copilot or Gemini through your work account, use that one for work planning. It’s what your IT team has approved.

Common mistakes (and quick fixes)

• Letting it plan 10 must-dos. Then nothing is a priority. Keep the hard limit of three.
• No buffer. A plan with back-to-back blocks breaks at the first overrun. The buffer rule is not optional.
• Planning tasks that are really projects. “Write the strategy” is not a 2-hour block. If something is bigger than a day, plan it first. Our guide to creating a project plan with AI breaks it into steps you can drop into a day.
• Accepting the first draft. The pushback step is where the plan becomes realistic. Two minutes, every time.
• Using it as a guilt list. If items keep moving to “tomorrow” for a week, ask the AI: “Which of these should I delegate, drop or say no to?” That’s a real answer too.

FAQ

Can AI really plan my day?

It can do the sorting, the time estimates and the replanning well, especially when you give it your fixed meetings and real past timings. It can’t know that your boss will be in a bad mood or that the report matters more than it looks. That’s why you stay the one who decides. Think of it as a fast first draft of your day.

What’s the best AI to plan my day?

The one that already sees your work. If your calendar is in Outlook and your company has Copilot, start there. If you live in Google Calendar, Gemini or Claude with the calendar connected saves typing. If you just want a quick plan from a brain dump, any of them, including ChatGPT, will do the job with the prompt above.

Is there a free AI daily planner?

You don’t need a separate app to start. The free versions of ChatGPT, Claude and Gemini can all turn a brain dump into a time-blocked plan with the prompts on this page. Calendar connections and higher limits can depend on your plan, so check what your account includes.

How long does this take each morning?

Plan on about five minutes: three for the brain dump, one for the AI, and two to push back and adjust. Less once you have a scheduled task and a connected calendar.

Start tomorrow morning

Pick one day this week. Do the brain dump, paste the planning prompt, push back once, and see how 4 p.m. feels compared to a normal day. If it helps, make it a scheduled task and let the habit run itself.

Want more workflows like this, one practical win at a time? Start on our home page, or join Launchpad right here:

You have the idea. Maybe you have the first clients. What you don’t have is the document a bank, a partner or your own sceptical brain keeps asking for. You open a blank page, type “Executive Summary”, and stare at it for twenty minutes.

Here’s the short answer. To use AI to write a business plan, don’t ask it to “write a business plan for my coffee shop”. Give it a short brief with the facts only you know (who pays you, what for, what it costs you, what you need), let it interview you about the gaps, then have it draft one section at a time in a format you picked: a one-page lean plan or a full traditional plan. You do the numbers yourself, and you finish by asking the AI to attack the plan like a lender would. ChatGPT, Claude or Gemini can get you to a solid first draft in an afternoon. The thinking stays yours. The typing gets a lot lighter.

Want one practical AI workflow like this every week, free?

What AI is good at in a business plan (and where it will fool you)

AI is very good at three jobs here: turning your messy notes into clear sections, asking the questions you forgot, and spotting where your story doesn’t hang together. That is most of the pain of writing a plan.

It is bad at one thing that matters a lot: facts about your market. OpenAI says it plainly in its own help article “Does ChatGPT tell the truth?”: the model can produce incorrect facts and even fabricated studies or citations, and it can sound confident while doing it. Their advice is to use it as a first draft, not a final source, and to check data and references yourself.

So a market size figure that shows up in your draft without a link you clicked yourself doesn’t go in the plan. Treat every number the AI offers as a question (“where would I check this?”), never as an answer.

There’s a second trap. Harvard professor William Sahlman wrote in Harvard Business Review that detailed financial projections for a new company are an act of imagination, and that most business plans spend far too much ink on the numbers. AI makes that worse if you let it, because it can produce five years of tidy spreadsheets in seconds. Tidy is not the same as true. Spend your effort on who the customer is, why they buy and what could go wrong.

Step 1: Pick your format before you prompt

The U.S. Small Business Administration describes two common formats in its guide “Write your business plan”:

• Lean startup plan. Usually one page, built from a handful of blocks: key partnerships, key activities, key resources, value proposition, customer relationships, customer segments, channels, cost structure and revenue streams. The SBA says it can take as little as an hour.
• Traditional plan. Detailed and often dozens of pages: executive summary, company description, market analysis, organization and management, service or product line, marketing and sales, funding request and financial projections. The SBA notes lenders and investors commonly ask for this one.

Rule of thumb: if you’re asking a bank or investor for money, write the traditional plan. If the plan is for you and a partner to get clear on the business, start lean. You can always grow the lean version into the full one later, and AI is great at that expansion step.

Telling the AI which format you want is the single biggest upgrade to your prompt. Without it you get a generic mix of both.

Step 2: Write a 10-minute founder brief

This is the part nobody can do for you, and it is why AI drafts usually sound empty. The model knows what a business plan looks like. It doesn’t know your business. Open a note and answer these in plain sentences. Bullet fragments are fine.

The last line matters more than it looks. A plan for a bank loan officer and a plan for a co-founder are different documents, and the AI can only write for the reader you name.

Step 3: Let the AI interview you first

Paste your brief and ask for questions, not a draft. This is where AI saves you the most time, because it is good at noticing what is missing.

Answer honestly, including “I don’t know yet”. Those “don’t know” answers become your to-do list. Some of them, like how big your local market is or what competitors charge, need real research. Our guide on how to use AI for market research shows how to do that with sources you can check, and the SBA’s guide lists free public data such as the Census Business Builder.

Step 4: Draft one section at a time

Asking for the whole plan in one go gives you a long, smooth, shallow document. Section by section gives you something you can actually defend. Start with the middle (market, customers, offer) and leave the executive summary for last, because it summarizes everything else.

Then repeat for each section you need. The [NEEDS DATA] tags are the trick. They stop the model from filling holes with plausible fiction, and they show you exactly what to look up.

A quick tool note. ChatGPT, Claude and Gemini all handle this well. Claude is pleasant for long documents if you paste the whole draft back in later, and it is what we use in how to use Claude for work. Pick the one you already pay for. The brief matters far more than the brand.

Step 5: Do the numbers yourself, with AI as a checker

This is where most AI business plans go wrong. Don’t ask the model to “create financial projections”. Give it your numbers and ask it to lay them out and check the math.

The SBA’s guide includes a simple break-even formula every plan should show: fixed costs divided by (price per unit minus variable cost per unit) gives your break-even point in units. Here is a made-up example to show the idea. Say your fixed costs are 3,000 a month, you sell a service at 500, and each sale costs you 100 to deliver. 3,000 divided by 400 is 7.5, so you need 8 sales a month before you make a cent.

Then check the math in a spreadsheet. Chatbots still slip on arithmetic now and then, and a lender will check. If you’d rather work from a sheet you already have, our guide to analyzing a spreadsheet with ChatGPT walks through uploading it and asking good questions.

Honest projections with your reasons written next to them beat confident five-year curves. That’s Sahlman’s point too: readers care about the people, the opportunity and the risks you see, not about how many cells you filled.

Step 6: Ask the AI to attack your plan

Once the draft is done, switch roles. This is the step that turns an AI draft into a plan you trust.

You will not like every answer. Good. It is much cheaper to hear it from a chatbot at your kitchen table than from a loan officer across the desk.

Want a second angle on the same plan? Run a quick SWOT analysis with AI on your business before you send it out. It shows where your strengths line up with real opportunities, and which threats your plan still has no answer for.

A worked example: a freelance bookkeeper’s lean plan

To make this concrete, here is an illustrative example (not a real client). Sam is a freelance bookkeeper who wants to stop taking every job and focus on small hair and beauty salons.

Her prompt without a brief: “Write a business plan for a bookkeeping business.” The result: a generic page about “offering high-quality financial services to businesses of all sizes”, a competitor section with no names, and growth figures nobody can check.

Her prompt with a brief: She told the AI she has six salon clients, charges a fixed monthly fee, gets most new clients through referrals from one supplier, and has room for about ten more clients on her own. She asked for a lean plan for herself and a possible part-time hire.

The draft came back with a clear value proposition (bookkeeping that understands booth rental and tips), a customer segment narrow enough to market to, and three [NEEDS DATA] tags: how many salons are within her area, what other bookkeepers charge them, and how many referrals the supplier sent last year. The attack step then pointed out that the whole plan leans on one referral source. That became the first risk in her plan, with a fix: test a second channel before hiring.

Same tools, same afternoon. The difference was the brief.

Free AI business plan generators vs a normal chatbot

If you search for an AI business plan generator, you’ll find many tools, free and paid. Some are useful, especially ones that give you a good template and a finance sheet that adds up. Under the hood, most of them do what you just did: ask you questions and fill sections.

So the honest test for any tool is simple. Does it ask you for your real facts before writing? Does it let you edit every section? Does it show you where its numbers come from? If yes, use whatever you like. If it writes ten pages from one sentence about your idea, you are back to a generic plan with a nicer cover.

A general chatbot plus the prompts on this page gets you the same result without a new subscription. If you’ve built a plan before, you’ll notice the work looks a lot like using AI to create a project plan: brief first, structure second, your judgement last.

What not to paste into a chatbot

A business plan holds sensitive things: client names, bank details, your salary, a partner’s personal finances. Before you paste, check your settings. OpenAI’s data controls page explains that on personal ChatGPT plans, data sharing for model training is on by default and can be turned off under Settings, Data controls, “Improve the model for everyone”. It also says business and enterprise workspaces aren’t used for training by default.

Simple habit: replace client names with “Client A”, round personal figures, and leave out account numbers completely. The AI doesn’t need them to write a good plan.

Applying for funding instead of a loan or investor? The same facts-first method works for grants too. Here is how to use AI to write a grant proposal section by section, including what funders say about AI.

Common mistakes

• No brief. One sentence in, ten generic pages out.
• Letting AI invent the market. Every number needs a source you opened yourself.
• Writing the executive summary first. Write it last, from the finished sections.
• Five-year projections with no reasons. Twelve honest months with your assumptions written down are more convincing.
• Skipping the attack step. The weak points are still there. You just find out later.
• Sending the AI’s wording as is. Read it out loud. If it doesn’t sound like you, a reader who meets you will notice.

FAQ

Can AI write a business plan for me?

It can write the draft, and it can save you hours of typing and structuring. It can’t know your customers, your costs or your goals unless you tell it, and it can get facts wrong. Think of it as a fast co-writer who needs a good brief and a careful editor.

What is the best AI to use to write a business plan?

ChatGPT, Claude and Gemini all do this well. Use the one you already know. A clear brief, a chosen format and the section-by-section method matter far more than which model you pick.

Can I use an AI-written business plan for a bank loan?

Yes, as long as everything in it is true and you can explain every number. The lender reads the plan, not the tool you used. Check the figures yourself, add real sources, and make sure you could answer the questions from Step 6 in a meeting.

How long does it take?

The SBA says a lean plan can take as little as an hour. With AI doing the drafting, a lean plan is realistic in one sitting. A traditional plan usually takes a few sessions, mostly because of the research and the numbers, which is time well spent.

What should a business plan include?

For a traditional plan, the SBA lists executive summary, company description, market analysis, organization and management, service or product line, marketing and sales, funding request and financial projections. Use only the sections your reader needs.

Stuck choosing between two business ideas, markets or pricing models before you write the plan? Make a decision matrix with AI first, so the plan starts from a choice you can defend.

Your next step

Open a note now and fill in the founder brief from Step 2. Ten minutes. Then paste it into the chatbot you already use with the interview prompt from Step 3. By tonight you’ll have a list of honest questions about your business, and that list is worth more than any polished template.

If you want more workflows like this, built for people who use a computer all day and want AI to take the boring parts, start on our home page or grab the free Launchpad right here.

Writing client proposals as well as plans? See AI proposal writing software for how to pick the right stack for your deals.

Still testing an idea before you write the plan? Here is how to use ChatGPT to make money and get your first paying customer.