Author: Pete

You have a hunch. A new service, a new customer group, maybe a new country. Before you spend money on it, you want to know if the market is real, who else is already there, and what customers complain about. And you’d like that answer this week, not after a consultant’s invoice.

Here’s the short answer. To use AI for market research, turn your hunch into one decision question, let a deep research tool (ChatGPT deep research, Gemini Deep Research, Claude Research or Perplexity Research) read the web and write a cited report, then add what customers say in their own words, check every number against its source, and squeeze it all into a one-page brief. The AI does hours of reading in minutes. You do the judging, and you still talk to a few real customers.

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

What AI is good at in market research, and where it fails

AI research tools are very good at the boring half of desk research: finding pages, reading them, comparing them and writing it up with links. OpenAI describes its deep research tool as “agentic”: it plans the search, evaluates sources, refines its queries and then writes a report with citations (OpenAI Academy).

It is not a crystal ball. OpenAI itself writes that deep research “can sometimes hallucinate facts” or make incorrect inferences, and may struggle to tell authoritative information from rumors (OpenAI, Introducing deep research). So treat every report as a well-read first draft. Useful. Not proof.

And it can’t do the one thing that matters most: tell you whether your future customers will pay you. That still comes from conversations, a test offer, or a small ad budget.

How to use AI for market research in 5 steps

Step 1: Turn your hunch into one decision question (10 minutes)

“Research the coaching market” gets you a generic essay. A decision question gets you something you can act on. Write down three things before you open any tool:

• The decision: launch it or not, which customer group first, what price range.
• The scope: country or region, type of customer, time frame (for example the last two years).
• The output you want: a table of competitors, a list of customer problems, a short recommendation.

Bad: “Tell me about the bookkeeping market.”
Good: “Should I offer a fixed-price monthly bookkeeping package to freelance designers in the Netherlands, and at roughly what price do competitors sell something similar?”

Step 2: Let a deep research tool do the reading (5 to 30 minutes, mostly waiting)

Normal chat answers from what the model already knows plus a quick search. A deep research mode runs many searches that build on each other and hands you a longer, cited report. All four big assistants have one now:

• ChatGPT deep research: pick Deep research from the + tools menu. ChatGPT proposes a research plan you can edit, you can limit which websites it uses (Sites > Manage sites), and you can download the report as Markdown, Word or PDF (OpenAI Help Center). OpenAI says a run may take 5 to 30 minutes.
• Gemini Deep Research: choose Deep Research in the Gemini text box, check the plan (click Edit plan to change it), then Start research. Google Search is a source by default and you can add Gmail, Drive or uploaded files. Google says a report usually takes about 5 to 10 minutes (Gemini Apps Help).
• Claude Research: on a paid Claude plan, click + and then Research, with web search turned on. Claude runs several searches that build on each other and can also look in connected tools like Google Docs (Claude Help Center).
• Perplexity Research: select Research in the mode selector of the search box. Free users get limited access, Pro gets more, and Perplexity says most tasks finish in under 3 minutes. You can export the report to PDF or a document (Perplexity Help Center).

Not sure which one you already have, or which one fits the rest of your work? Our ChatGPT vs Claude vs Perplexity comparison walks through that. Then paste this prompt in and fill the brackets:

I'm deciding: [your decision question].
Market: [country/region], customers: [who exactly], time frame: [e.g. last 2 years].

Write a report for a busy owner, max 2 pages, with:
1. The 5 to 8 most relevant competitors or alternatives, in a table:
   name, what they sell, who it's for, price if public, source link.
2. The main customer problems and buying reasons you can find,
   with a link for each one.
3. Signs of demand: growing or shrinking, and the evidence.
4. Risks and reasons NOT to do this.
5. Three questions I should ask real customers before deciding.

Rules: link a source for every number. If you can't find a
number, write "not found" instead of estimating. Mark anything
that is your own inference with [INFERENCE].

That last block of rules matters more than the rest. It makes the gaps visible instead of letting the AI paper over them.

When the plan appears, read it before you press start. If it wanders into a topic you don’t care about, edit it out. Two minutes there saves you a report full of the wrong thing.

How to use Perplexity AI for market research

Perplexity is handy when you want quick, source-heavy answers to lots of small questions, because every answer shows its sources up front. A good rhythm: run one Research query for the big picture, then use normal searches for follow-ups like “What does [competitor] charge for [service]?” or “Which [industry] forums discuss [problem]?” Open the source behind every price before you copy it into your notes.

Step 3: Map the competitors yourself, with AI as your assistant (20 minutes)

The report gives you names. Now go one level deeper on the three or four that matter most. Open their websites, copy the text of their pricing and “about” pages, and paste it into ChatGPT or Claude with:

Here is the text from the pricing and about pages of [competitor 1],
[competitor 2] and [competitor 3].
Make a table: target customer, main promise, price and what's included,
what they DON'T offer. Only use the text I pasted. If something
isn't in the text, write "not stated".
Then name one gap none of them covers well.

Because you pasted the source yourself, the table is built on what’s really on their sites today, not on what a model remembers from last year.

Step 4: Hear the customer in their own words (20 minutes)

Competitor pages tell you what sellers promise. Reviews, forum threads and support questions tell you what buyers actually feel. Collect 30 to 100 real comments: public reviews of competitors, threads from industry forums or Reddit, or your own customer emails and survey answers. Then ask:

Below are customer comments about [type of product/service].
Group them into the 5 most common complaints and the 5 most common
reasons people buy. For each group: how many comments fit it, and
2 short quotes in the customer's exact words.
Don't add anything that isn't in the comments.

Those exact quotes are gold later, for your offer, your landing page and your ads. If you have a big survey export, our guide to analyzing a spreadsheet with ChatGPT shows how to handle the file, and for long reports there’s summarizing long documents with ChatGPT.

Before you paste anything from your own customers, check what your company allows. Client emails and survey answers with names in them are personal data. Strip the names, or use the business plan your workplace has approved.

No survey answers to work with yet? Our guide on how to use AI to create a survey helps you write a short one that people actually finish. Once the answers are in, here is how to use AI to analyze survey data without losing what people actually said.

Step 5: Check demand and every number (15 minutes)

This is the step most people skip, and it’s the one that keeps you from building a plan on a made-up statistic.

• Click every source behind a number. Does the page really say that? Is it from this year or from 2019? Is it about your country? If the source is missing or says something else, the number goes.
• Be careful with market-size figures. Big “the market is worth X billion” numbers are often from paid reports the AI can’t read, quoted second-hand. Ask for the method instead: “Estimate how many [customers] exist in [region], show your calculation, and link a source for each input.” A rough count you can check beats a big number you can’t.
• Check search demand with Google Trends. Compare your main term with a couple of alternatives over the same time range and location (Trends Help). Remember the numbers are relative, scaled from 0 to 100, and Google says Trends is not a scientific poll, so use it as one data point among others (Google Trends FAQ).

Turn it into a one-page brief you can decide on

A 15-page report you never reread is not research. It’s homework. Paste your checked notes back into the chat and ask:

From my notes below, write a one-page decision brief:
- The decision in one line
- What we know (only checked facts, with links)
- What we still don't know
- My 3 options, with the main risk of each
- What I would test first, in under 2 weeks and with a small budget
Plain language. No buzzwords.

Need to present it to a partner or your team? Our guide on how to use AI to make a presentation takes that brief to slides.

Before you pick one of those 3 options, it helps to see them side by side. Our guide to a SWOT analysis with AI shows how to sort your checked notes into strengths, weaknesses, opportunities and threats, and turn them into a short list of moves.

Ready to turn that research into a full plan for a lender, a partner or yourself? Here is how to use AI to write a business plan, section by section.

Market research vs market analysis: same workflow?

Mostly, yes. Market research is collecting the facts: who the competitors are, what customers want, how big the demand is. Market analysis is what you do with them: comparing options, spotting a gap, choosing a price. Steps 2 to 4 above are the research. Step 5 and the brief are the analysis. AI helps with both, but the analysis is where your own judgment has to lead, because you know your costs, your skills and your clients.

One thing this guide is not about: stock market research. Picking investments is a different job with different risks, and we don’t cover it here.

An example: from hunch to test in one afternoon

Say you’re a freelance bookkeeper and you keep getting questions from designers. Here’s how the five steps could look:

1. Question: “Should I offer a fixed-price monthly package for freelance designers in my country, and at what price?”
2. Deep research: one report on who already serves freelancers this way, their public prices and what freelancers say about bookkeeping.
3. Competitors: you paste the pricing pages of the three closest offers and get a clean table, plus one gap (for example, nobody explains the package in plain language for creative people).
4. Customer voice: 50 forum comments from freelancers about bookkeeping, grouped into complaints like “I never know what I’ll pay” and “I don’t understand what my bookkeeper sends me”.
5. Check: you click every price, drop the two that don’t match the source, and look at search interest in Google Trends.

The brief ends with a small test: message ten designers you know, offer the package at the price range you found, and see how many say yes. That test is the real market research. AI just got you there with a lot less reading.

Not sure which assistant to run these steps in? Our ChatGPT vs Gemini for research comparison runs the same research task through both so you can pick one.

FAQ

Can I use ChatGPT for market research for free?

You can do a lot in a normal free chat: brainstorm questions, group customer comments you paste in, build competitor tables from page text. Deep research usage varies by plan and ChatGPT shows your remaining tasks in a counter (OpenAI Help Center). Perplexity also gives free users limited access to its Research mode.

Is AI market research accurate?

It’s as accurate as the sources it finds and your check afterwards. The tools give you citations precisely so you can verify them. Use the “not found” and [INFERENCE] rules from the prompt, and click before you trust.

Which AI tool is best for market research?

The one you already pay for is usually good enough. If you live in Google Workspace, Gemini can pull in your Gmail and Drive. If you want fast answers with sources visible up front, Perplexity is quick. ChatGPT and Claude are both strong for the report and the follow-up analysis. The workflow matters more than the brand.

Can AI replace talking to customers?

No. It can tell you what people say in public. It can’t tell you whether the people you can reach will pay your price. Use the AI work to ask much better questions in those conversations.

Your next step

Pick the one business question that has been sitting in your head the longest, write it as a decision, and run it through the Step 2 prompt today. You’ll know in half an hour whether it deserves a real test. For more workflows like this, from campaign ideas with ChatGPT to small automations, start with the free Launchpad on our home page.

The meeting is tomorrow at ten. You have the notes, the numbers and a rough idea of what you want to say. What you don’t have is the two hours it takes to turn all of that into slides that don’t look thrown together.

Here’s the short answer. To use AI to make a presentation, don’t start by asking for “a deck about X”. Give ChatGPT, Claude or Gemini your real material first (notes, a report, a spreadsheet), ask for a storyline of 6 to 10 slides with one message per slide, fix that outline, and only then let a slide tool build it: Copilot in PowerPoint, Gemini in Google Slides, Canva, or Claude, which can hand you a finished .pptx file. Then you check every number and cut the text by half. The AI does the structure and the first draft. You keep the judgment.

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

Why “make me a presentation” gives you a generic deck

If you type one line into an AI slide maker, it has nothing to work with except the topic. So it fills the gaps with safe, general statements and stock photos. The slides look tidy and say very little.

The fix is simple. A presentation is a story first and a design second. Get the story right in plain text, where it’s fast to change, and the design step becomes almost automatic. That’s also how the AI tools themselves are built: Canva’s AI shows you an outline you can edit before it generates anything (Canva Help Center), and Copilot in PowerPoint may ask who the audience is and how long the deck should be before it starts (Microsoft Support).

How to use AI to make a presentation in 5 steps

Step 1: Feed it your real material (5 minutes)

Collect what you already have: meeting notes, the report the deck is based on, a few rows of data, last quarter’s deck. Paste it in or upload the file. Then add three facts the AI can’t guess:

• Who is watching: your manager, a client, the whole team.
• What you want them to do afterwards: approve a budget, pick an option, stop worrying about the project.
• How long you have: 10 minutes is a very different deck than 45.

That last line in your prompt does more for the result than any clever wording.

Step 2: Ask for the storyline, not the slides

Copy this into ChatGPT, Claude or Gemini and fill in the brackets:

I need to give a [10]-minute presentation to [my manager and two team leads].
After it, I want them to [approve moving the launch by two weeks].
Below are my notes and data.

Make an outline of 6 to 10 slides. For each slide give:
1. A headline that states the point as a full sentence (not a topic label)
2. Max 3 short bullets
3. One suggestion for a chart, table or image, if it helps
4. 2 to 3 sentences of speaker notes

Use only facts from my notes. If something is missing, write
"[CHECK: ...]" instead of making it up. End with a list of
questions my audience is likely to ask.

[paste notes / data here]

Two things in that prompt matter most. “Headline as a full sentence” turns “Q3 numbers” into “Q3 sales are up, but returns doubled”, which is what people remember. And the “[CHECK]” line stops the AI from quietly inventing a figure to fill a gap.

Step 3: Fix the outline before anything gets designed

Read the headlines only, top to bottom. Do they tell the story on their own? If you’d skip a slide while presenting, delete it. If a slide makes two points, split it. Ask the AI for changes in plain words: “merge slides 3 and 4”, “the client doesn’t care about our internal process, cut it”, “start with the decision we need”.

This is the step most people skip, and it’s the one that makes the deck sound like you.

Step 4: Turn the outline into slides with the tool you already use

Pick the route that matches where you work. You don’t need all of them.

• PowerPoint with Copilot: start a new presentation (or open your company template first), open Copilot and paste your outline as the prompt. Copilot builds the deck with your template, and on work licences it can also start from a Word or PDF file (Copilot in PowerPoint FAQ). It needs a Microsoft 365 plan that includes Copilot.
• Claude: in the same chat, say “Now turn this outline into a PowerPoint file.” Claude can create .pptx files you download or save to Google Drive, and that works on the Free plan too (Claude Help Center). If you don’t name a file format, it may build the deck as an artifact instead, so say “PowerPoint file” out loud.
• Google Slides with Gemini: open your presentation, click Ask Gemini and describe the slide. You can point it at a Drive file with @, for example “Create a slide that summarizes @Project notes”. It generates one slide at a time in your current theme (Google Docs Editors Help), so paste one outline slide per prompt. It needs an eligible Workspace or Google AI plan.
• Canva: on the homepage choose Canva AI, then Design, then Presentation, paste your outline as the prompt, adjust the outline Canva shows you and click Generate presentation (Canva Help Center). Good when looks matter more than your company template. It counts toward your monthly Canva AI usage.

If you’re new to the chat tools themselves, our Claude tutorial for beginners and Gemini tutorial for beginners get you set up in a few minutes.

Step 5: Do the human pass (10 minutes)

Now open the deck and go through it slide by slide:

• Check every number against your source. Search the deck for “[CHECK” and fill those in yourself.
• Cut text. Microsoft’s own advice is short bullets of one line each and no font smaller than 18 pt, because you want people listening to you, not reading the screen (Microsoft Support). AI drafts almost always carry too many words.
• Swap generic images for your own screenshot, chart or photo where you can. A real screenshot of your dashboard beats a stock photo of a handshake.
• Read the speaker notes out loud once. Change anything you’d never actually say.

Which AI tool is best for presentations?

Honestly, the one you already pay for and work in. The quality of the story comes from steps 1 to 3, and every tool above does those well enough. The difference is in the last mile:

• Your company has a strict template: PowerPoint with Copilot, or the Claude for PowerPoint add-in on paid Claude plans, which builds slides from your existing template’s layouts (Claude docs).
• Your team lives in Google Workspace: Gemini in Slides, because it can pull straight from your Drive files.
• You want it to look designed with little effort: Canva.
• You want a file fast and you’re on a free plan: Claude’s file creation, then open the .pptx in PowerPoint, Keynote or Google Slides.

How to use AI to make a PowerPoint presentation for free

You can do the whole thing without a paid plan. Use the free version of ChatGPT, Claude or Gemini for steps 1 to 3 (the storyline is the valuable part anyway). Then either ask Claude for the .pptx file, or copy the headlines and bullets into a blank PowerPoint or Google Slides deck yourself. With a clean outline, building 8 simple slides by hand is a short job, and you’ll know every line on them.

Free plans have usage limits, so do your outline in one focused chat instead of ten scattered ones.

An example: from messy project notes to an 8-slide update

Say you run a small website project and you need a 10-minute update for a client. Your input is a page of notes from the last three status calls plus a short list of open tasks.

With the prompt above, a typical outline comes back like this: one slide with the decision you need (“We recommend moving launch to 14 March”), one with what’s done, one with what’s blocking, one with two options and their trade-offs, one with the new timeline, and a closing slide with the next step. Plus speaker notes and a list of likely client questions, such as “Does this change the budget?”

The bits where your notes were vague come back as “[CHECK: confirm hosting migration date]”. Those are exactly the places where an AI would otherwise have guessed. You fill them in, cut the bullets, and paste the outline into your tool of choice. The structure is done before you open a slide.

If your source is a long report rather than notes, summarize it first with ChatGPT for long documents. If it’s a pile of meeting notes, turn them into a report and build the deck from that.

A deck usually needs a clean logo on the title slide too. If you do not have one, see the best AI tool for logo design for options that fit a work budget.

Mistakes that make AI slides look like AI slides

• Letting it invent numbers. If a figure isn’t in your source, it shouldn’t be on a slide. Your audience will ask where it came from.
• Topic headlines. “Market overview” says nothing. “Two competitors cut prices this quarter” says something.
• Too many slides. AI is happy to give you 20. For a 10-minute slot, 6 to 10 is plenty.
• Pasting confidential data into a personal account. Use the work account or tool your company has approved for client and financial data, or strip names and figures first.
• Presenting a deck you haven’t rehearsed. The slides were fast. Spend part of the time you saved saying them out loud once.

Need to show a process on a slide, like an approval flow or a customer journey? Skip drawing boxes by hand. Our guide on how to make a flowchart with AI shows how to get a clean diagram you can export as an image and drop into your deck.

Questions people ask

Can Claude make a PowerPoint presentation?

Yes. Ask it to create a PowerPoint file and it gives you a .pptx to download or save to Google Drive. On paid plans there’s also an add-in that works inside PowerPoint itself.

Can AI make a Google Slides presentation?

Gemini in Google Slides generates and edits slides inside your deck, one slide per prompt, if your Workspace or Google AI plan includes it. Without it, you can open a Claude-made .pptx in Google Slides or paste an AI outline in yourself.

How long does it take?

Depends on your material. Most of the time goes into steps 3 and 5, fixing the story and checking facts, and that’s the time well spent. The typing and layout are what the AI takes off your plate.

Is it OK to use an AI-made deck at work?

Treat it like a draft from a fast junior colleague: useful, but you sign off on it. Check your company’s policy on which AI tools may see work data, check every fact, and make sure the opinions on the slides are yours.

Your next step

Take the next presentation on your calendar, paste your notes into the prompt above, and only look at the headlines it gives you. If those tell the story, you are most of the way there. For more workflows like this, from using Claude for everyday work to small automations, start with the free Launchpad on our home page.

You’re the one everybody asks. How do I issue a refund, where’s the client onboarding checklist, what’s the login for the supplier portal. It all lives in your head, and you know you should write it down. You just never get two quiet hours to do it.

Here’s the short answer. You can use AI to write SOPs in about 20 minutes of your own time for a simple process: do the task once while you talk it through out loud, turn that recording into a transcript, paste the transcript into ChatGPT or Claude with a fixed SOP prompt, then have someone else follow the draft and fix what trips them up. The AI does the formatting and the tidying. You still supply the knowledge and the final check.

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

Why AI is good at SOPs (and where it isn’t)

A standard operating procedure is mostly structure: a purpose, who it’s for, what you need, numbered steps, what can go wrong. The US Environmental Protection Agency’s guide on writing SOPs puts the style in one line: write them “in a concise, step-by-step, easy-to-read format”, in active voice and present tense, and keep them short (EPA QA/G-6, Guidance for Preparing SOPs).

That is exactly the kind of rewriting ChatGPT and Claude do well. Give them a rambling explanation and they hand back clean numbered steps.

What they can’t do is know your process. If you ask “write an SOP for processing refunds” with nothing else, you get a generic document that sounds right and matches nobody’s real work. The same EPA guide says SOPs should be written by the people who actually do the work. So the trick is simple: you bring the raw knowledge, AI brings the shape.

Step 1: Capture the process while you do it (5 minutes)

Don’t sit down to “write”. Do the task for real and narrate it, like you’re showing a new colleague over your shoulder. Say what you click, why you click it, and what you check before moving on. Mention the annoying bits too: “if the invoice is in dollars, stop and ask finance first.”

Ways to capture it:

• Phone voice memo. On an iPhone 12 or later, the Voice Memos app transcribes speech for you, and you can copy the whole transcript with one tap (Apple Support: Voice Memos transcription).
• Screen recording with your voice. Many video call and screen recording tools can produce a transcript. Grab that text.
• Screenshots plus notes. For a click-heavy process, take a screenshot at each key screen and type one line per screenshot.
• An existing mess. An old email where you explained it to someone, a Slack thread, a half-finished doc. That’s raw material too.

Messy is fine. “Uhm, then you go to, no wait, first you check the order number” is perfect input. Cleaning that up is the AI’s job.

Step 2: Turn the transcript into an SOP with one prompt

Paste your transcript into ChatGPT or Claude together with this prompt. Fill in the brackets.

You are helping me write a standard operating procedure (SOP).
Below is a transcript of me doing the task "[task name]" and explaining it out loud.

Turn it into an SOP with these sections:
1. Purpose (one sentence)
2. Who does this and when (trigger)
3. What you need before starting (access, tools, files)
4. Steps: numbered, one action per step, active voice, present tense
5. Checks: what to verify before the task counts as done
6. Common problems and what to do
7. Who to ask if stuck

Rules:
- Use ONLY what is in the transcript. Do not invent steps, tools or names.
- If something is unclear or missing, do not guess. List it under
  "Open questions for the author" at the end.
- Write for someone with basic computer skills who has never done this task.
- Keep it short. No intro, no conclusion.

Transcript:
[paste here]

Two lines in that prompt do most of the work. “Use only what is in the transcript” stops the model from filling gaps with plausible fiction. “Open questions for the author” turns those gaps into a short to-do list for you, instead of hidden mistakes for your colleague.

If your transcript is long, ChatGPT and Claude both handle it fine. For really long material, like a 40-page manual you want to split into several SOPs, see our guide to getting ChatGPT to summarize long documents for work.

Step 3: Answer the open questions, then fix the draft

Read the “open questions” list first. Typical ones: “Which folder are signed contracts saved in?” or “What is the refund limit before approval is needed?” Answer them in the same chat in plain words and ask for an updated version.

Then read the steps yourself, once, slowly. Look for three things:

• Merged steps. “Log in and create the invoice” is two steps. Split them.
• Missing decisions. Every “if this, then that” you said out loud should be in there.
• Invented details. A button name or menu path you never mentioned. Delete it or correct it.

Step 4: Let someone else test it

This is the step people skip, and it’s the one that makes an SOP actually work. The EPA guidance recommends that draft SOPs be tested by someone other than the writer before they’re finalized. Hand the draft to a colleague, a freelancer, or your future self in three months, and watch where they get stuck. Every question they ask is a missing line.

A good test: could someone with basic computer skills follow it alone, without phoning you? If yes, it’s done.

Some processes are easier to follow as a picture, especially when there are a few yes or no decisions in them. You can turn the same steps into a diagram: here is how to make a flowchart with AI from a process you have already written down.

Step 5: Keep your SOPs in one place AI can reuse

Once you have five or ten SOPs, two things help.

One home. A shared folder or wiki with one SOP per process and a clear name (“SOP: Issue a customer refund”). Add a “last checked” date at the top. Processes change, and an outdated SOP is worse than none.

One AI workspace. In Claude you can create a Project, upload your SOP template and a couple of finished SOPs as project knowledge, and add instructions like “always use this format”. Every new chat in that project then follows your house style without re-pasting the prompt. Projects are available on free Claude accounts too, with a limit of five projects (Claude Help Center: What are projects?). ChatGPT has a similar Projects feature. New to either tool? Our Claude tutorial for beginners covers the setup in your first 30 minutes.

An example: from voice memo to SOP

Say you run a small agency and onboarding a new client is always “ask Sanne, she knows”. Sanne records herself once while setting up a real client: creating the folder, sending the welcome email, adding them to the invoicing tool, booking the kickoff call. That’s maybe six minutes of talking.

Run through the prompt above, that turns into a one-page SOP: a dozen or so numbered steps, a “before you start” list (access to the drive, the email template, the calendar link) and three open questions Sanne answers in two minutes. The next time a client signs, a junior colleague does it from the SOP. And once the steps are written down and stable, you can see which ones a tool could do for you. That’s how you get from “documented” to “automated”, like in our guide to automating client onboarding with Make.com.

Hiring rather than signing a client? The SOPs you write this way are exactly what a new starter needs in their first weeks. Our guide on how to use AI for employee onboarding shows how to fold them into a 30-60-90 day plan and a new-hire FAQ.

AI SOP generators: do you need one?

Search for “AI SOP generator” and you’ll find dedicated tools, some of which record your screen and build a step-by-step guide with screenshots automatically. They’re handy for click-heavy software processes where a picture of every screen saves explaining.

For most small teams, though, ChatGPT or Claude plus the prompt above is enough. You already pay for it or use the free tier, there’s nothing new to learn, and the SOP lands in whatever doc tool you already use. Try the free route on three processes first. If you find yourself pasting screenshots all day, then a dedicated tool is worth a look.

Quick answers

Can ChatGPT write SOPs on its own?

It can write a generic SOP template from a one-line request. It can’t write your SOP, because it doesn’t know your systems, limits and exceptions. Give it a transcript or notes of the real process and it does a good job.

Is it safe to paste internal processes into AI?

Leave out passwords, personal customer data and anything confidential. Use placeholders like [CUSTOMER NAME] or [ADMIN LOGIN]. If your company has a policy on AI tools, follow it, and check whether your plan uses chats for training in the tool’s data settings.

What’s the best format for an SOP?

Short and scannable: purpose, trigger, what you need, numbered steps with one action each, checks, problems, contact. For a process with many branches, a simple flow chart helps, and the EPA guide recommends one too. You can ask the AI to draft the chart as a list of decisions first.

How often should I update SOPs?

Whenever the process changes, and otherwise on a fixed rhythm. The EPA guide suggests a periodic review, for example every one to two years. For fast-moving software processes, a quick check every few months is more realistic.

Can I turn meeting notes into an SOP?

Yes, if the meeting was someone explaining a process. Use the same prompt. For regular meeting notes, our guide on using ChatGPT to summarize meeting notes is the better fit.

Hiring someone to take over a process? The same interview approach works for the role itself: see how to use AI to write a job description.

Turning a bigger piece of work into steps, owners and dates instead of a single procedure? See how to use AI to create a project plan.

Interviewing for that role next? See how to use AI to create interview questions that check whether a candidate can really do the work, with a simple scoring guide.

Need to announce a new process in a team meeting? Turn the SOP into spoken talking points with how to write a speech with AI, so the rollout still sounds like you.

Your next step

Pick the one task people ask you about most this week. Record yourself doing it, run the prompt, answer the open questions, and hand it to someone to test. That’s one less interruption, every week, from now on. If you’d like a new hands-on AI workflow for your work every week, join the free Launchpad on our home page, or open robotshelp.me to see what’s inside.

It’s the last week of the month. There are 30 supplier invoices spread across your inbox, two are phone photos of paper bills, one is in German, and you’re typing totals into a spreadsheet again. You’ve seen “AI invoice processing” everywhere, but the tool lists all look the same.

Here’s the short answer. AI invoice processing software reads an invoice (a PDF, a scan or a photo), pulls out the supplier, invoice number, dates, VAT and total, often the line items too, and sends that to your accounting software or a sheet so you only check and approve. For most small businesses the right pick is a capture app that plugs straight into the accounting software you already use. Build your own in n8n only when you need custom routing, and use a raw document AI API only if a developer is involved. One check comes before all of that: if your suppliers are switching to structured e-invoices, part of the problem disappears on its own.

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

What AI invoice processing software actually does

Strip the marketing away and every tool in this category does the same three jobs.

1. Capture. Invoices come in by email forward, upload, phone photo or a shared inbox.
2. Extract. An AI model reads the document and fills fixed fields. Microsoft’s invoice model, for example, returns fields like vendor name, invoice ID, invoice date, due date, subtotal, total tax, amount due, IBAN and line items, and it handles phone photos, scans and digital PDFs (Microsoft Learn: invoice model). Google’s Invoice Parser can pull up to 46 fields, including VAT amounts and supplier tax IDs (Google Cloud: Document AI processors).
3. Hand off. The fields go into Xero, QuickBooks, your ERP or a Google Sheet, ideally with the original PDF attached, so you approve instead of type.

The difference between tools is almost never “can it read an invoice”. They all can. The difference is where the data ends up, how you correct mistakes, and what it costs per document once you’re past the free trial.

The four kinds of tools (and who each one is for)

1. Bookkeeping capture apps

These sit next to your accounting software. You snap or forward an invoice, the app reads it and creates the bill for you. Hubdoc, for example, extracts supplier names, amounts, invoice numbers and due dates and creates transactions in Xero and QuickBooks Online with the source document attached (Hubdoc). Dext does the same job and accepts documents by mobile, email, WhatsApp or desktop (Dext).

Good for: freelancers and small teams that already use a cloud accounting package. Less good for: anything outside bookkeeping, like routing an invoice to a project manager for sign-off.

2. Accounts payable automation platforms

One step up. These add approval flows, purchase order matching, supplier onboarding and payment runs. Nanonets is one example; it positions itself around AI agents for accounts payable and reconciliation (Nanonets).

Good for: a finance team handling hundreds of invoices a month with several approvers. Less good for: a one-person business. You’ll pay for features you never open.

3. Build your own workflow (n8n, Make, Zapier plus an AI model)

You connect your inbox to an AI model and a sheet yourself. n8n’s Extract From File node pulls the text out of a PDF (n8n docs), a model turns it into fields, and the result lands wherever you want. We walk through that exact build in our n8n workflow for automatic invoice processing.

Good for: people who want invoices routed by project, client or cost centre, or posted into a tool no capture app supports. Less good for: anyone who doesn’t want to maintain a workflow when a supplier changes their layout.

4. Document AI APIs

Microsoft Document Intelligence and Google Document AI are cloud services you call from your own code. You get raw fields plus a confidence score per field, and you build everything around it.

Good for: a company with a developer and a high, steady volume. Less good for: everyone else. There’s no inbox, no approval screen and no accounting sync out of the box.

How to choose: six questions that decide it

1. Where must the data end up? If the answer is Xero or QuickBooks, start with a capture app that syncs with it natively. Everything else is extra plumbing.
2. How many invoices a month? Twenty is a capture app or a small DIY flow. Several hundred with approvals is AP automation territory.
3. Do you need line items or just totals? Totals and VAT are enough for bookkeeping. Line items matter when you rebill clients or match purchase orders, and that is where tools differ most.
4. Which languages and countries? Check the vendor’s supported language list against your real suppliers. Microsoft, for instance, publishes its list and even extracts the Dutch KVK number as its own field (invoice model schema).
5. Where is the data stored? Invoices carry names, addresses and bank details. If you’re in the EU, ask where documents are processed and kept, and whether the vendor signs a data processing agreement.
6. What does a correction cost you? No tool is right every time. The best one for you is the one where fixing a wrong field takes two seconds, not a support ticket.

Test it on your own invoices before you pay

Vendors quote accuracy figures (Dext’s homepage says over 99%). Treat any number like that as their test, not yours. The only test that counts is your own messy pile. It takes about 30 minutes:

Run the same 10 through two tools and the choice usually makes itself. If you want a free baseline first, you can ask Claude to pull the same fields from a PDF into a table; here’s how to extract data from a PDF with Claude. It won’t replace a tool for monthly volume, but it shows you what “good extraction” looks like on your own documents.

Check this first: are your invoices becoming e-invoices?

A lot of invoice software exists to read PDFs. Structured e-invoices don’t need reading: the data already arrives as fields. And in parts of Europe they’re becoming the rule.

• Belgium: since 1 January 2026, invoices between Belgian VAT-registered businesses must be structured e-invoices; a PDF by email no longer counts (Belgian government e-invoice FAQ).
• Germany: since 1 January 2025, businesses must be able to receive e-invoices (German Federal Ministry of Finance FAQ).
• The wider EU: the Commission tracks each country’s rules on its eInvoicing pages (European Commission: eInvoicing).

What that means for your choice: pick software that can both receive structured e-invoices and read the PDFs and photos that will keep coming from smaller or foreign suppliers. Paying only for PDF reading could leave you with half a solution in a year or two.

What it won’t do for you

• Decide what gets paid. Keep approval with a person, at least for anything above an amount you pick.
• Spot every odd invoice. A changed bank account on a familiar supplier’s invoice should always get a phone call to a number you already have, whatever the software says.
• Fix a messy inbox. If invoices arrive in five places, set up one forwarding address first. That alone saves more time than most features.

Choosing a tool for contracts instead of invoices? The same test-it-on-your-own-files approach works there too: see our guide to AI contract review software.

Quick answers

Can I just use ChatGPT or Claude to process invoices?

For a handful a month, yes: upload the PDF and ask for the fields in a table. For a steady flow you want something that runs on its own, which is why people end up with a capture app or a workflow like the ones in our email automation examples.

What does it cost?

Capture apps usually charge a monthly plan with a document limit; APIs charge per page. Prices change often, so compare the vendors’ own pricing pages against your real monthly volume, not their headline plan.

Do I need to code?

No for capture apps and AP platforms. No for n8n, Make or Zapier either, though you’ll spend an afternoon setting it up. If you’ve never built a workflow, start with our n8n tutorial for beginners.

Your next step

Pick the kind of tool that matches where your data must end up, run the 10-invoice test on two options, and keep a person on the approve button. That’s the whole decision. If you’d like a new hands-on AI workflow for your work every week, join the free Launchpad on our home page.

You have seen the photos: a little Mac mini on a shelf, running an AI agent all day. Now you are on Apple’s order page, looking at 16GB, 24GB, 32GB and a Pro chip, and wondering how much of that you actually need. Fair question. The price gap between the cheapest and the most expensive Mac mini is bigger than the cost of a year of most AI subscriptions.

Here is the short answer. If your OpenClaw agent uses a cloud model like Claude, ChatGPT or Gemini, the base Mac mini is enough: Apple M6, 16GB memory, 256GB storage, $899 in Apple’s US store. OpenClaw itself only needs a recent Node version, and the OpenClaw Mac app needs macOS 26.2 or later. The heavy thinking happens on the AI company’s servers, not on your desk. Pay for more memory only if you want to run the AI model on the Mac mini itself. Then memory is the one spec that decides everything, and 32GB is where it starts to feel usable. This guide to OpenClaw Mac mini specs shows you which box fits which job, so you don’t overspend or come up short.

Want one practical AI lesson like this every week?

What OpenClaw itself needs (it’s less than you think)

OpenClaw’s own install docs keep the list short: Node 24.16 or newer, or Node 26.1 or newer, and the installer sets up Node for you if it’s missing (OpenClaw docs: Install). There is no minimum chip and no minimum memory in that list.

Two macOS details matter before you buy, especially a used Mac mini:

• The OpenClaw Mac app (the menu bar app with the dashboard and chat window) needs macOS 26.2 (Tahoe) or later.
• The Terminal install runs on Node, whose Mac builds are made for macOS 13.5 or later. Older versions may still run it, but they are not supported (OpenClaw docs: macOS app).

Why so little? The part of OpenClaw that runs on your Mac mini is the Gateway. It keeps your settings, your chats and your connections to apps like Telegram or Slack, and it passes each job to the AI model. With a cloud model, the model does the hard work somewhere else and sends back the answer. Your Mac mini is mostly waiting, routing and running a browser now and then.

The current Mac mini line-up, in one table

Apple sells two chips in the Mac mini right now: the M6 and the M5 Pro. These are the numbers from Apple’s own spec page (Apple: Mac mini technical specifications). Prices are from Apple’s US store on 26 October 2026 and change by country.

Every model has Ethernet (2.5Gb, 10Gb as an option), Wi-Fi 7 and the same small 12.7 cm square box. The M5 Pro adds Thunderbolt 5 ports. For an OpenClaw box, the rows that matter are memory and storage. The rest is nice to have.

One thing to know before you order: you choose the memory when you buy. Apple only offers it as an option at checkout, so there is no adding more later. Storage you can always extend with an external drive. Memory you can’t.

Memory: the one OpenClaw Mac mini spec that decides it

Everything comes down to one question: where does the AI model run?

Cloud model (Claude, ChatGPT, Gemini): 16GB is fine

This is how most people start, and it’s what OpenClaw’s own setup steers you to first. The Mac mini runs the Gateway, maybe a browser for web tasks, and a few connected apps. 16GB covers that comfortably. If you also plan to run Docker, several browser sessions at once, or a second user account that stays logged in, 24GB gives you breathing room.

The real cost of a cloud setup is not the hardware. It’s the AI plan or API usage behind it. If that part is new to you, our guide on what happens when you hit ChatGPT and Claude usage limits shows where the caps are.

Local model (the AI runs on the Mac mini): memory first, then speed

Running the model yourself means nothing leaves your desk, and there is no per-message bill. The trade is hardware. On Apple silicon the chip and the graphics share one pool of memory, so the whole model has to fit in it, next to macOS and OpenClaw.

OpenClaw is honest about this. Its managed local setup checks your memory, graphics memory and free disk before it downloads anything. The smallest model recipe has an 8GiB memory floor, larger ones need more, and the docs add that these floors “do not guarantee fit or speed” (OpenClaw docs: Local models). The same page warns that a model that answers a short test prompt can still fail a real agent task, because the agent sends a long prompt with tools and history.

A simple rule of thumb: take the model’s download size, then leave plenty of room on top for macOS, OpenClaw and the conversation. Here are real download sizes from the Ollama library, a popular way to run models locally:

• Llama 3.1 8B: 4.9GB (Ollama: llama3.1)
• Qwen3 14B: 9.3GB, and Qwen3 32B: 20GB (Ollama: qwen3)
• gpt-oss 20B: 14GB (Ollama: gpt-oss)
• Llama 3.1 70B: 43GB

Read that list next to the memory options and the picture gets clear. A 16GB Mac mini can hold a small 8B model, but that leaves little room for anything else. 32GB fits the 14B and 20B class with space to work. A 70B model at 43GB is only realistic on an M5 Pro with 64GB. Small local models are fine for simple jobs like sorting or short summaries. For an agent that plans, browses and uses tools, most people still get better results from a cloud model.

Speed is the second half. When a local model writes its answer, it reads through its memory again and again, so memory bandwidth matters more than the number of CPU cores. That is why the M5 Pro’s 307GB/s, about double the base M6, is the real reason to pay for it. Hands-on buyer guides for OpenClaw land on the same advice: buy the Pro for local models, not for cloud ones (Bitdoze: Best Mac mini for OpenClaw).

Storage: 256GB or 512GB?

For a cloud-model agent, 256GB is enough. OpenClaw, its logs and a browser profile are small next to that.

Go for 512GB or more when any of these apply:

• You run local models. Each one is several GB, and most people try three or four before they settle.
• You want a macOS virtual machine to keep the agent apart from your own Mac. OpenClaw’s docs count about 60GB of free disk per VM (OpenClaw docs: macOS VMs).
• You use Docker or keep big project folders on the machine.

If you are on a budget, the base 256GB plus an external SSD over Thunderbolt is a fair middle way. Keep the operating system and OpenClaw on the internal drive, and store model files and archives on the external one.

Which Mac mini should you buy? Three honest picks

This is what I tell friends who ask me which one to get.

1. “My agent uses Claude or ChatGPT, and I just want it always on.”
Base Mac mini M6, 16GB, 256GB. Don’t let anyone talk you into more. You get an agent that answers your messages day and night, reads your inbox for you and runs your weekly jobs, and the box itself is the cheapest part of the setup.

2. “I’ll run browser jobs, maybe Docker, and I want to try a small local model.”
Mac mini M6 with 24GB or 32GB and 512GB. This is the safe middle. It gives OpenClaw room to run next to other tools, and a small model for simple background work fits.

3. “The model has to run here. Client data stays in the office.”
Mac mini M5 Pro with 48GB or 64GB and at least 1TB. This is where local models start to do real agent work at a speed you’ll accept. Before you spend this much, test the exact model you plan to use on a real task from your week, as OpenClaw’s own docs advise.

Already have an older Mac mini? Try it first. An Apple silicon Mac mini with an M1 or newer can run the OpenClaw Gateway with a cloud model. Check which macOS it runs: below macOS 26.2 you use the Terminal install instead of the menu bar app. Our guide to installing OpenClaw on a Mac covers both routes.

Before you buy: a one-week test that costs nothing

The cheapest way to pick the right specs is to find out what you actually use. Install OpenClaw on the Mac or Windows laptop you already have (here is the Windows guide), connect a cloud model, and give it three real jobs for a week. A morning summary of your inbox, a weekly report, a research task.

At the end of that week you know two things. Whether you want it running when your laptop is closed (then a Mac mini makes sense), and whether a cloud model does the job (then the base model is all you need). Not sure OpenClaw is the right agent at all? Compare it first in OpenClaw vs Claude Code and OpenClaw vs Claude Cowork.

Once the Mac mini is on your desk, follow our OpenClaw Mac mini setup guide: a separate user account for the agent, no sleep, automatic login and remote access, step by step.

Quick answers

Is a Mac mini with 8GB enough for OpenClaw?

For a cloud model, it can work, because OpenClaw lists no memory minimum and the model runs elsewhere. It gets tight once a browser and other apps are open, and it leaves no room for a local model. The current Mac mini starts at 16GB anyway.

Do I need the M5 Pro for OpenClaw?

Only if you want to run large AI models on the machine itself. For a cloud-model agent, the extra power mostly sits idle.

Can I add more memory later?

No. You choose the memory when you order, so decide based on whether you’ll want local models in the next few years. Storage you can extend with an external drive.

Mac mini or a cheap server (VPS)?

OpenClaw’s docs list both as good options. A VPS is cheaper to start with. A Mac mini gives you iMessage, a home internet address that websites block less often for browser jobs, and hardware you own (OpenClaw docs: macOS VMs).

Do I need a monitor?

Only for the first setup. After that you reach the Mac mini from your laptop or phone. The setup guide linked above shows how.

Pick the specs for the jobs you have today, not for the agent you might build in three years. If you want help building those jobs step by step, the free RobotsHelp Launchpad on our home page is where we start.

Get one hands-on AI workflow every week.

Sources

• Install: system requirements (OpenClaw docs)
• macOS app: requirements (OpenClaw docs)
• Local models: hardware floor (OpenClaw docs)
• macOS VMs: always-on options and disk space (OpenClaw docs)
• Mac mini technical specifications (Apple)
• Buy Mac mini: prices and configurations (Apple US store)
• Model download sizes (Ollama library)
• Best Mac mini for OpenClaw (Bitdoze)

Your inbox is in Gmail, your files are in Drive, and everyone keeps telling you to “just use Gemini”. You open it, type “help me with my work”, and get three friendly paragraphs that help with nothing. Sound familiar?

Here is the short answer. To start with Gemini, go to gemini.google.com and sign in with your Google account. Then give it one real task from today, with four things in the message: who you are, what you need, the background, and what the result should look like. Add your own file with Add files, or type @ to let it read your Gmail or Drive. Correct the first answer in the same chat instead of starting over. When a task comes back every week, save your instructions as a skill. That is the whole beginner loop. This Gemini tutorial for beginners walks you through it step by step, with prompts you can copy.

Want one practical AI lesson like this every week?

Step 1: Sign in with the right Google account

Gemini is Google’s AI assistant. It runs in any normal browser at gemini.google.com, and there is a Gemini app for phones too. You can try a few things without signing in, but file uploads, saved chats and the connection to Gmail and Drive only work once you are signed in (Gemini Apps Help: Use Gemini Apps).

Which account you pick matters more than with most AI tools, because Gemini can read your Google data:

• Personal Google account (Gmail address): works straight away and is free to start.
• Work account on Google Workspace: works only if your company’s edition includes it and your admin allows it. You also have to be 18 or over (Gemini Apps Help: What you need to sign in).

My advice: if your employer gives you Gemini on your work account, use that one for work. Your files are already there, and the company has decided what is allowed. If not, learn on your personal account with practice material, and keep client documents out of it until you have checked the privacy setting in Step 6.

Step 2: Write your first work prompt (four parts)

A prompt is simply your message. Google’s own guide for writing instructions names four parts: the persona (what role to play), the task (what to do), the context (the background) and the format (how the result should look). You do not need all four every time, but a few of them make a big difference (Gemini Apps Help: Tips for writing instructions).

Here it is with a real task. Swap in your own details:

You are an experienced office manager who writes short, friendly emails.

Task: write an email to the whole team about the office move.

Context: we move to the new building on Monday 9 November.
Everyone packs their own desk on Friday 6 November before 3pm.
Boxes and labels are in the kitchen from Wednesday.
IT moves the screens; people only take laptops home.

Format: under 150 words, a clear subject line,
the three dates as a short list, one friendly line at the end.

Now compare that with “write an email about the move”. The long version gives Gemini the dates, the rules and the tone. The short one makes it guess all three, and guesses read like a template you would never send.

Once the draft is close, a second prompt can polish it. The ones that work best are in make this email sound more professional, and they work the same in Gemini.

Step 3: Give Gemini your own material

This is where it gets useful. An answer from your document beats an answer from general knowledge every time. There are two ways in.

Upload a file

Click Add files in the text box, then Upload files, and pick a document, spreadsheet or image. According to Google’s help page you can add up to 10 files in one prompt, each up to 100 MB (videos can be bigger), and you can ask Gemini to make a chart from an uploaded spreadsheet (Gemini Apps Help: Upload and analyze files).

A good first test: upload a report you already know well, so you can judge the answer.

Read the attached report.
1. Give me the 5 most important points in plain English, one line each.
2. For each point, say which section it comes from.
3. List anything that needs a decision from me, with the deadline if there is one.

Asking where each point comes from is the best beginner habit. You can check it in seconds, and you quickly learn how far you can trust it on your kind of documents.

Let it read Gmail and Drive with @

This is the part that makes Gemini different from ChatGPT or Claude for a Google user. Type @ in the text box and pick Gmail or Drive. The first time, Gemini asks your permission to connect. After that it can find and summarize emails and documents for you (Gemini Apps Help: Connect Google Workspace).

Try one of these:

@Gmail Summarize the emails from Sarah about the Henderson project
from the last two weeks. What did we agree, and what is still open?

@Drive Find my notes from the Q3 planning meeting and list
every action item with the person who owns it.

Three things Google’s help page warns about, and they match what beginners run into:

• Use keywords, not links. Don’t paste the URL of an email or document. Describe it with words from the content, like a name or project.
• Check the sources. Gemini can pick an older email when a newer one exists. Open the sources listed under the answer before you act on it.
• It can’t connect? Keep Activity has to be on, and in Gmail the setting “Smart features in other Google products” has to be switched on. On a work account, your admin decides.

For long PDFs and reports, our guides go deeper. They use Claude and ChatGPT, but the prompts carry over to Gemini: summarize long documents for work and summarize meeting notes.

Step 4: Correct it, don’t start over

The first answer is a draft. Most beginners either take it as it is, or open a new chat and start again from zero. Both throw away the work Gemini already did.

Reply in the same chat, short and direct, the way you would to a colleague:

• “Shorter. Cut the second paragraph.”
• “Warmer. This client has been with us for five years.”
• “The deadline is the 14th, not the 12th. Fix it everywhere.”
• “Put this in a table: task, owner, deadline.”

Gemini can also turn a finished answer into a file. Ask “put this in a Google Doc” or “make this a spreadsheet”, and it creates a Doc, Sheet, PDF or Word file you can download or save to Drive (Gemini Apps Help: Generate files from your chats). That saves the copy and paste at the end.

Step 5: Save repeat work as a skill (not a Gem)

If you paste the same background into Gemini every Monday, save it once. Until now that was called a Gem. Google is replacing Gems with skills: on personal accounts Gems go away from November 2026 and are turned into skills automatically, and work accounts follow in 2027 (Gemini Apps Help: The move from Gems to skills). So if you are starting now, start with skills.

A skill is a set of saved instructions. You call it by typing / and its name in any chat, and Gemini can also use it by itself when your request fits. Skills work without a paid plan, but you need to be 18 or over and have Keep Activity on (Gemini Apps Help: Create and manage skills).

The easiest way to make one is to ask for it in a chat:

Create a skill called weekly-client-update based on these instructions:

When I paste my rough notes, turn them into a client update with
three headings: Done this week / Next week / Needs your input.
Max 150 words. Plain English, no jargon.
Use the client's first name, never "Dear client".
If my notes mention a risk or a delay, put it first under "Needs your input".

Next Friday you type /weekly-client-update, paste your notes, and the update comes back in your format. You can find, edit or switch off your skills under Customize, then Skills, in the sidebar. If your updates come from a Google Sheet, our guide to Gemini for automated status updates shows the whole Friday routine.

Step 6: Check one privacy setting before real work

On a personal account, while Keep Activity is on, Google can use your chats to improve its services, and some chats are read by human reviewers (Gemini Apps Privacy Hub). The catch for beginners: the Gmail and Drive connection and skills need Keep Activity on.

So here is the simple rule I use:

• For something sensitive: use a Temporary chat (next to New chat). It doesn’t show up in your history, but it also can’t use skills or connected apps (Gemini Apps Help).
• For everyday work: leave names, salaries and client numbers out of the prompt, or use the work account your employer has approved.

Five beginner tasks to try in Gemini this week

It sticks when you use it on Tuesday’s real work, not on test questions. Pick two:

1. Catch up after a day off. “@Gmail What came in from my manager and my clients since Friday? List what needs an answer from me.”
2. The email you keep postponing. Paste the email you got, say in one sentence what you want to answer, ask for a reply under 100 words.
3. Meeting notes into actions. Upload or paste your notes and ask for decisions, action items with owners, and open questions.
4. A spreadsheet you don’t understand. Upload it and ask: “What does this sheet track, what stands out, and make one chart of the most important trend.”
5. Prep for a call. “@Drive Find the last proposal we sent to [client]. Give me the three points they are most likely to ask about.”

Free or paid: what a beginner actually needs

The free plan is enough for everything in this tutorial: chat, file uploads, the @ connection to your Google apps and skills. Google lists the paid plans (Google AI Plus, Pro and Ultra) with higher usage limits, stronger models and Gemini built into Gmail and Docs themselves (Gemini subscriptions).

Pay only after you have hit the free limits several times on real work. Not because a new model launched. If you are torn between paid plans, read ChatGPT vs Gemini Pro or Claude vs Gemini Pro first.

Once you are comfortable with text, try images too: how to use Gemini to edit photos walks through prompts that work.

Beginner mistakes that cost the most time

• Vague first prompts. “Help me with this report” gets you a lecture. Say who it is for and what you need from it.
• Trusting a summary of the wrong email. With @Gmail, always open the source once. It takes five seconds.
• A new chat for every correction. Stay in the same chat, so Gemini keeps your context.
• Building ten skills on day one. Make one for the task you really repeat. Add the next one when you catch yourself retyping again.
• Sending without reading. Google itself says Gemini can make mistakes. You are still the one who hits send.

Most of your work lives in email? Here is how to use Gemini in Gmail to summarize long threads and draft replies right inside your inbox.

Gemini, ChatGPT or Claude for a beginner?

If your day already runs on Gmail, Docs and Drive, Gemini is the easiest place to start, because your material is one @ away. If your files come from everywhere, ChatGPT or Claude may feel smoother. The basics are the same in all three, so nothing you learn here is wasted. Our ChatGPT tutorial for beginners and Claude tutorial for beginners follow the same steps, so you can compare for yourself.

If your main job is learning from notes and PDFs, not inbox triage, start with Claude vs Gemini for studying before you lock in a daily tool.

Pick one tool, use it for the same three jobs every week for a month, and the hours come back. If you want a place to build that habit step by step, the free RobotsHelp Launchpad on our home page is where we start.

If you are enrolled in higher education, you may also qualify to get Gemini for students free for a year instead of starting on the paid plan.

Once those six steps feel natural, you can run the same habits inside Gmail, Docs, Sheets and Meet. Our guide on how to use Gemini for work shows a desk system for the full workday, including prompts you can save and reuse.

FAQ

Is Gemini free?

Yes, there is a free plan that covers chat, file uploads and the connection to Gmail and Drive. Paid Google AI plans add higher limits and stronger models (Gemini subscriptions).

Can I use Gemini with my work Google account?

Only if your company’s Google Workspace edition includes it and your admin allows it. If it doesn’t work, ask your IT team before you move work files to a personal account.

What happened to Gems?

Google is turning Gems into skills. On personal accounts that starts in November 2026, and your existing Gems are converted automatically. New users can simply start with skills.

Can Gemini read my emails?

Only when you connect it. Type @ and pick Gmail, and Gemini asks for your permission first. You can disconnect it again in the Connected Apps settings.

Get one hands-on AI workflow every week.

Sources

• Use Gemini Apps (Gemini Apps Help)
• What you need to sign in to Gemini Apps (Gemini Apps Help)
• Tips for writing Gem instructions (Gemini Apps Help)
• Upload and analyze files in Gemini Apps (Gemini Apps Help)
• Connect the Google Workspace app to Gemini Apps (Gemini Apps Help)
• Create and manage skills (Gemini Apps Help)
• About the transition from Gems to skills (Gemini Apps Help)
• Gemini Apps Privacy Hub (Google)
• Gemini subscriptions (Google)

You built a workflow in n8n, it works, and now everyone online is talking about “agents”. You drop the AI Agent node on the canvas, connect it, ask it something simple, and it either does nothing useful or calls the wrong tool three times. Frustrating. And very normal for a first agent.

The short answer: an n8n AI agent workflow is a trigger, one AI Agent node, a chat model, a memory, and a small set of tools the agent may choose from. The model decides which tool to call, n8n runs it, and the agent answers with the result. Keep the first one to three tools, write a clear system message, and put human approval on any tool that sends or changes something. Below you build exactly that: a meeting-prep agent that reads your calendar and email, then drafts a short prep note for you to check.

Short, practical AI lessons for your work week. Join Launchpad free, right above.

What an n8n AI agent workflow actually is

A normal n8n workflow is a fixed recipe. Step 1, then step 2, then step 3, every time. If you followed our n8n tutorial for beginners, you built one of those, with a single AI step in the middle that rewrites some text.

An agent is different. You give it a goal and a toolbox, and it picks the steps itself. n8n describes the AI Agent node like this: connect a chat model and one or more tools, and the agent decides which tools to call to complete a task (n8n docs: AI Agent).

That freedom is the whole point, and also the whole risk. A fixed recipe can’t surprise you. An agent can. So the craft of a good n8n agent is mostly about limits: few tools, clear instructions, and a person approving anything that leaves your account.

Not sure you need an agent at all? Our guide on how to build an AI agent without code helps you decide, and covers the same idea in Zapier and Make.

The five parts of every n8n AI agent

Open any working agent workflow in n8n and you will find the same five parts. Learn these and you can read almost any agent template.

• Trigger: what starts the run. For a chat agent that is the Chat Trigger. It can also be a schedule, a form, or a new email.
• AI Agent node: the “brain” box in the middle. Since n8n 1.82.0 every AI Agent node works as a Tools Agent, so there is no agent type to pick anymore (n8n docs).
• Chat model: the language model that does the thinking, for example the OpenAI or Anthropic chat model sub-node.
• Memory: lets the agent remember earlier messages in the same conversation. Simple Memory is the easy starting point.
• Tools: the actions the agent is allowed to take. Gmail, Google Calendar, Google Sheets, Slack, an HTTP Request and many more. n8n requires at least one tool on an AI Agent node.

On the canvas, the model, memory and tools hang underneath the agent as small “sub-nodes”. That is why it looks different from a normal workflow. Nothing to fear, they are just plugged into the agent instead of lined up after it.

Before you build: pick one job

The fastest way to fail is “an agent that handles my inbox”. Too big, too vague, too many tools. Pick one job you do every week that has a clear finish line.

Our example: meeting prep. Before a client call you open your calendar, search your email for the last messages with that person, and jot down what’s open. Ten minutes, every meeting. The agent does the looking up and writes the note. You read it and walk into the call prepared.

Why this is a good first agent:

• It mostly reads. Reading your calendar and searching your mail can’t hurt anyone.
• It has one write step (a Gmail draft), and you’ll put approval on it.
• You can check the result in seconds. You know your own meetings.

Build the meeting-prep agent, node by node

You need an n8n account (n8n Cloud has a free trial with no credit card, or the free self-hosted Community Edition), plus a Google account with Gmail and Calendar. Plan on about 45 minutes the first time.

Step 1: Add a Chat Trigger

Start a new workflow and choose On chat message as the trigger. This gives you a chat window right inside the editor, so you can talk to your agent while you build. Leave Make Chat Publicly Available off for now. n8n’s own advice is to keep it off while building and only turn it on when you are ready (n8n docs: Chat Trigger).

Step 2: Add the AI Agent node and a chat model

Click the plus after the trigger and add AI Agent. Leave Prompt on “Take from previous node automatically”, so the agent reads whatever you type in the chat.

Under the agent, click Chat Model and pick one, for example OpenAI Chat Model or Anthropic Chat Model, then add your API key as a credential. n8n’s Cloud plans also list “AI models without API keys” as a feature, so check your plan before you go and create a key (n8n pricing).

Step 3: Add Simple Memory

Click Memory under the agent and add Simple Memory. Its Context Window Length setting decides how many earlier messages the agent keeps in mind. The default is fine to start.

Two honest limits: n8n says this memory doesn’t persist between sessions (n8n docs: Tools Agent), and Simple Memory doesn’t work in an active production workflow when n8n runs in queue mode (n8n docs: Simple Memory). For a personal prep assistant, that’s fine. It only needs to remember the current chat.

Step 4: Give it two read tools

Click Tool under the agent.

• Google Calendar tool, operation Event: Get Many. Pick your calendar. For the time range fields, click the small “let the model define this” button so the agent fills them in from your question.
• Gmail tool, operation Message: Get Many. Set a sensible limit (10 is plenty) and let the model fill in the search filter, so it can look up mail from the person you’re meeting.

That button writes a $fromAI() expression for you. You can also type it yourself, with a hint for the model, like {{ $fromAI('search', 'Gmail search query, for example from:name@company.com', 'string') }}. The key isn’t a reference to existing data. It’s a hint the model uses to fill in the right value (n8n docs: $fromAI()).

Give each tool a clear name on the canvas, like “Read my calendar” and “Search my email”. The agent sees those names and descriptions when it decides what to use.

Step 5: Add a write tool, behind human review

Now the one action that leaves a trace: a Gmail tool with operation Draft: Create. It saves the prep note as a draft in your own mailbox. It doesn’t send anything.

Even so, put it behind approval. In the agent’s Tools panel, find the Human review section, pick a channel (the built-in Chat works, so do Slack, Telegram, Gmail and others), and connect the draft tool to that review step. When the agent wants to use it, the workflow pauses and asks you to approve or deny, showing which tool and which values it wants to use (n8n docs: Human-in-the-loop for tools). n8n lists human approval for tool calls on its Cloud plans, including Starter.

Later, if you ever add a tool that sends mail or changes a record, it goes behind the same review step. That’s the habit worth building on agent number one.

Step 6: Write the system message

Open the AI Agent node, go to Options, add System Message, and paste something like this. Change the names to your own.

You are my meeting-prep assistant. I am a freelance consultant.

When I name a meeting or a day:
1. Use "Read my calendar" to find the meeting. If you find more than one match, ask me which one.
2. Use "Search my email" to find the last messages with the attendees. Search by their email address.
3. Write a prep note with: who I'm meeting, what we last discussed, open questions or promises, and 3 questions I could ask.

Rules:
- Only use facts from my calendar and email. If something isn't there, say "not found". Never guess.
- "Save draft" needs my approval. If I deny it, don't try again. Just show the note in the chat.
- Keep the note under 200 words.

Notice the line about a denied approval. n8n recommends telling the agent which tools need approval and how to respond when you say no, so it handles it gracefully instead of trying again (n8n docs).

Step 7: Test it in the chat

Click Open chat and try: “Prep my next meeting tomorrow.” Watch the canvas. Each tool lights up as the agent uses it, and you can click into every step to see exactly what it sent and got back.

When it works, save the workflow and publish it. From now on, one chat message gets you a prep note, instead of ten minutes of clicking between tabs.

Test it like a new colleague, not a demo

A demo is one happy question. Real work is messy. Before you trust it, run these five tests:

1. A meeting that doesn’t exist. It should say “not found”, not invent one.
2. Two meetings with similar names. It should ask which one.
3. A new contact with no email history. The note should say so plainly.
4. Deny the draft. It should show the note in chat and stop.
5. A vague question, like “anything important this week?” See if it stays inside its job.

If something goes wrong, open the AI Agent options and turn on Return Intermediate Steps. You’ll see each tool call the agent made, which makes it much easier to spot where it went off track. The Max Iterations option (10 by default) caps how many rounds the model gets, which stops a confused agent from looping (n8n docs: Tools Agent).

Most fixes are in the system message, not the nodes. Vague answer? Make the instruction more specific. Wrong tool? Rename the tool so its job is obvious.

What an agent workflow costs to run

Two meters run at the same time.

• n8n executions. n8n bills by executions, one full run of a workflow. With a Chat Trigger, every message is an execution, so a chat of 10 messages uses 10 (n8n docs: Chat Trigger). The Cloud trial includes 1,000 executions, and paid Cloud plans start at the Starter plan (n8n pricing). Self-hosting the Community Edition has no execution bill, but you run the server.
• The AI model. If you use your own OpenAI or Anthropic key, every agent step is billed by that provider. One prep note usually means a few model calls (decide, read calendar, search mail, write), not one.

For a personal assistant you use a few times a day, the execution count stays small. It’s a public chatbot on your website that can surprise you, which is one more reason to keep that “public” switch off until you’ve seen the numbers.

When a plain workflow beats an agent

Honest point: plenty of jobs don’t need an agent. If the steps are the same every time, a fixed workflow is cheaper, faster and more predictable.

• Same steps every time? Use a normal workflow with one AI step. Our n8n invoice processing workflow is a good example. Invoices always go Gmail, read PDF, Sheet.
• Need to decide what to look up, or ask a follow-up question? That’s agent work. Meeting prep is like that, because every meeting needs a different search.

A good middle ground: a fixed workflow that runs on a schedule, with an agent inside it for the one step that needs judgment. Our n8n CRM automation guide shows that pattern with an AI summary inside regular workflows.

Four mistakes that break a first n8n agent

1. Too many tools. Ten tools on agent number one means the model has ten ways to get it wrong. Start with two or three.
2. No system message. Without one, the agent guesses what you want. Write down the job, the steps and the rules.
3. Write tools without review. Sending mail, updating a CRM or deleting anything should go behind human review until you trust it fully.
4. Testing only the happy path. Run the five tests above. The wrong answers teach you more than the right ones.

Where to go next

Once this agent runs, you can reuse the same skeleton for other jobs by swapping the tools and the system message:

• Qualify incoming leads with a scoring prompt: AI agent for lead qualification.
• Let an agent offer and book time slots: AI agent for appointment booking.
• Answer customer questions with a clear handoff to a person: automate customer support with AI agents.
• Still choosing a builder? Compare n8n with Zapier, Make and others in no code AI agent builder: how to pick.

And if you want one practical lesson like this at a time, start free with Launchpad on the RobotsHelp home page.

FAQ

Do I need to code to build an AI agent workflow in n8n?

No. Everything above is clicking and filling in fields. You will see expressions like $fromAI(), but the button writes them for you. n8n does let you add code steps later if you want.

Which chat model should I use for an n8n agent?

Use one you already pay for or know. The Tools Agent supports, among others, the OpenAI, Anthropic, Groq and Mistral Cloud chat models (n8n docs). Pick a model that is good at tool calling, and test with your real questions.

Why does my agent remember nothing the next day?

Simple Memory only holds the current conversation, and n8n notes that memory doesn’t persist between sessions. If you need longer memory, n8n offers other memory sub-nodes backed by a database, but for meeting prep you rarely need it. The facts live in your calendar and email anyway.

Can the agent send emails by itself?

Only if you give it a send tool. In this build it can only create a draft, and only after you approve. Keep it that way until you’ve seen it work for a few weeks.

Is n8n better than Zapier for AI agents?

Different, not better. n8n gives you more control over each piece and can run on your own server. Zapier is quicker if your apps are already connected there. Our n8n vs Zapier comparison goes deeper.

Build yours this week

Pick one meeting you have coming up. Build the agent, ask it to prep that meeting, and compare its note with what you would have written. If it got 80 percent there, you just bought back ten minutes before every call. If it didn’t, the fix is almost always one clearer line in the system message.

Sources

• n8n docs: AI Agent node
• n8n docs: Tools Agent (system message, max iterations, supported models)
• n8n docs: Human-in-the-loop for tools
• n8n docs: Use AI for parameters ($fromAI)
• n8n docs: Simple Memory
• n8n docs: Chat Trigger
• n8n docs: Gmail node operations
• n8n docs: Google Calendar node operations
• n8n pricing and plans

New to agents in general? Start with our AI agents course guide and a simple 4 week plan.

You already know Claude can write an email. But on a normal Tuesday you still copy text in, copy answers out, and switch between your inbox, your calendar and six open documents. It feels like a smart tool sitting next to your work instead of inside it.

Here is the short answer. To use Claude for work, stop treating it as a chat box and give it access to where your work already lives. Connect your email, calendar and files once, put each task you repeat into its own Project, use Cowork in the desktop app when you need a finished file from several sources, and run the same three prompts every day: a morning inbox brief, a meeting prep, and an end-of-day handoff. Claude asks before it sends or deletes anything by default, so you stay in charge. The setup takes about an hour, and the steps are below.

Get one practical AI lesson for your work every week.

Brand new to Claude? Do the first 30 minutes in our Claude tutorial for beginners first (sign up, your first prompt, your first file). This page picks up where that one stops: making Claude part of your working week.

Chat, Project or Cowork: pick the right mode for the job

Most people who say “Claude didn’t really help me at work” used one mode for everything. Claude has three ways of working, and each fits a different kind of task.

• Chat is for anything that fits in a few messages: a question, a rewrite, a gut check on a plan, a summary of one document.
• A Project is a workspace with its own files and standing instructions. Every chat inside it starts with that background, so you stop pasting the same context every Monday. Free accounts get up to 5 Projects; paid plans have no cap (Claude pricing).
• Cowork runs in the Claude desktop app on Mac or Windows and is for delegating a deliverable: it reads files, folders and connected apps, works through the steps, and hands you a document, deck or spreadsheet. It needs a paid plan (Claude Cowork getting started guide).

Anthropic’s own rule of thumb is simple: if you want an answer, use chat. If you want a file someone will open, or the task touches more than one file or app, use Cowork (Best practices for getting started with Claude Cowork). Their example: “How do I VLOOKUP something?” is a chat. “Go through my spreadsheets and change all the VLOOKUP to INDEX MATCH” is a Cowork task.

Not sure if Cowork or a tool like OpenClaw fits you better? We compared them in OpenClaw vs Claude Cowork.

Step 1: Connect the apps your work lives in (15 minutes)

This is the step that changes everything, and most people skip it. Without connectors, Claude only knows what you paste. With them, it can look things up in your mail, calendar and files itself, and it shows you which emails or documents it used, with links back to the originals.

If you work in Google (Gmail, Calendar, Drive)

1. In Claude, go to Customize > Connectors.
2. Find Gmail, Google Calendar and Google Drive, click Connect and sign in with your Google account.
3. In a chat, click the + button, hover over Connectors and switch on the ones you need for that conversation.

According to Claude’s Google Workspace connectors help page, Claude can then search and read your email, draft replies, look at your calendar and find free slots, and search, read and create files in Drive. It can also send email and change calendar events, but by default it asks for your approval before each of those actions. On Team and Enterprise plans, an owner has to switch the connectors on for the organization first.

Two limits worth knowing from that same page: Claude sees attachment details in Gmail but not the attachment content, and when it reads a Drive document it takes the text only, not embedded images.

If you work in Microsoft 365 (Outlook, Teams, SharePoint)

The Microsoft 365 connector covers Outlook mail and calendar, Teams chats, SharePoint and OneDrive. It needs a work account, not a personal outlook.com or hotmail.com address, and your IT admin has to approve it once for your company. Sending email or posting in Teams only works if your admin turned on the write tools (Connect to Microsoft 365, Claude Help Center). If you get stuck, that is a two-minute question for IT, not a reason to give up.

Step 2: Turn your three most repeated tasks into Projects (20 minutes)

Look at last week. Which three things did you do more than once? For most office jobs it is something like a weekly client update, a recurring report, and answering the same kind of question from colleagues or customers.

Make one Project for each. Give it a clear name, write instructions the way you would brief a new colleague, and add the files it should always know: a template, a style guide, last month’s version. Here is an example for a weekly client update:

This project is for my Friday update to the client Northwind.
When I paste rough notes, write the update under three headings:
Done this week / Next week / Needs your decision.
Max 150 words. Plain English. Use first names, never "Dear client".
If a deadline moved, put it first under "Needs your decision" with the new date.
Ask me before guessing any number that is not in my notes.

That last line matters. Telling Claude to ask instead of guess is the cheapest way to avoid a wrong number landing in a client’s inbox.

If a task has no written steps yet, write those first. Our guide on how to use AI to write SOPs turns a spoken walkthrough into a procedure you can drop straight into the Project as a file.

The click-by-click setup of a Project is in our beginner tutorial, Step 5.

Step 3: Run a daily Claude routine (10 minutes a day)

Tools stick when they are attached to a moment in your day. These three prompts work with the connectors from Step 1. Copy them, then change the names, channels and time windows to your own.

Morning: the inbox brief

Check my Gmail from the last 16 hours.
List only the emails that need a reply or a decision from me, max 8.
For each one: sender, one line on what they want, and a deadline if they gave one.
Skip newsletters, notifications and anything I'm only in CC on.
Then draft replies for the two most urgent ones. Don't send anything.

You read eight lines instead of forty emails, and two drafts are waiting. Want the mail itself to sort and label automatically, without you asking? That is an automation job, and we show eight of them in email automation examples.

Before a meeting: the one-page prep

I have a call with Acme at 14:00 today (it's in my calendar).
Find our last emails with anyone at Acme and any Drive docs with "Acme" in the name.
Give me one page: where we left off, what they are likely to ask,
open points on our side, and one thing I should push for.

This is close to the meeting prep example Anthropic gives in its Cowork best practices guide: name the sources, the time window and the shape of the answer, and you get a brief instead of a guess.

End of day: the handoff to tomorrow

Here are my rough notes from today: [paste]
Look at my calendar for tomorrow.
Give me a short plan for tomorrow morning: the 3 things to do first,
what to prepare for each meeting, and any reply I promised today but haven't sent.

Ten minutes at 17:00, and tomorrow starts with a plan instead of a scroll through your inbox.

Step 4: Ask for files, not just answers

A lot of work ends in a file: a report, a slide deck, a cleaned-up spreadsheet. Claude can make those directly. File creation is included on every plan, Free too (Claude pricing).

• Want a Google file? Say so by name. “Make a Google Slides deck from these notes” creates a real Slides file you can open and share. If you just ask for “a deck”, you get a file to download instead (Claude Help Center).
• Edit together. With the Docs, Sheets and Slides connectors on, Claude can edit a Google file in a pane next to the chat while you work in it too. Your normal Google sharing rules still apply.
• Several sources, one file. That is Cowork territory: “Read the three PDFs in this folder and the Q3 sheet, then build a two-page summary in our report template.”

Working from long PDFs? These guides go deeper: Claude AI analyze long PDF files, summarize a long report and pull out action items, and extract data from a PDF into a spreadsheet.

How to use Claude for work in your role

The routine above works for almost anyone. Here is where to aim it first, depending on what you do all day.

• Sales and account managers: the meeting prep prompt before every client call, plus a Project with your offer, pricing rules and the three objections you hear most, so follow-up emails come out right the first time.
• Consultants and freelancers: a Project per client with their brief, past deliverables and tone. Ask Claude to turn your call notes into a proposal outline, then write the parts that need your judgment yourself.
• Team leads and managers: the Friday update Project from Step 2, and a Monday prompt that reads last week’s notes and calendar and lists what slipped.
• Operations and admin: Cowork for the boring middle of a task: compare two spreadsheets, clean up a column, rename a folder of files, build the monthly summary.
• Anyone who writes a lot: keep your own writing samples in a Project so drafts sound like you. If you’re still choosing between tools for writing, read ChatGPT vs Claude for writing.

Keep your work data safe

Using Claude for work means client names, numbers and emails pass through it. Four habits keep that sensible.

• Follow your company’s rules first. If your employer offers Claude on a Team or Enterprise plan, use that account for work. Anthropic doesn’t train on chats from those commercial plans by default (Claude pricing).
• Check the training setting on a personal account. On Free, Pro and Max, your chats are only used to improve Claude if you allow it. Look under Settings > Privacy and set it the way you want (Anthropic Privacy Center).
• Know what connectors share. Anthropic says it doesn’t train on raw content from connectors like Gmail or Drive. Text you copy into the chat yourself is treated like any other chat (same Privacy Center page).
• Keep the approval step on. Let Claude draft, and you press send. Reading every email before it goes out costs seconds and saves awkward apologies.

Free or Pro for work?

The Free plan already includes connectors, file creation, memory and up to 5 Projects, which covers Steps 1 to 4 for light use. Pro costs $20 a month (or $17 a month billed yearly) and adds more usage, Projects without the 5-Project cap, Research, scheduled tasks, Claude in Chrome and Microsoft 365, and Cowork (Claude pricing).

My advice: run the daily routine on Free for a week. If you hit the usage limit before lunch most days, or you want Cowork to build files for you, Pro will pay for itself quickly. What happens when you hit the cap is explained in ChatGPT vs Claude usage limits.

Mistakes that make Claude feel useless at work

• Never connecting anything. Then every answer depends on what you remembered to paste.
• One giant chat for everything. Long, mixed chats get messy and use up your limit faster. New topic, new chat. Repeated topic, a Project.
• Vague asks with connectors on. “Find anything important” makes Claude guess. Name the sender, the time window and what counts as important.
• Trusting figures without a check. Claude links the emails and documents it used. Click one or two before a number goes to a client or your boss.

FAQ

Is Claude good for work tasks?

Yes, especially for writing, summarizing, reading long documents and preparing for meetings. It gets much more useful once it can read your own email, calendar and files through connectors, instead of working only from what you paste in.

Can Claude read my work email and calendar?

Yes, through the Gmail and Google Calendar connectors or the Microsoft 365 connector. On company plans, an admin may need to enable them first. Claude only sees what your own account can already see.

Can Claude send emails for me?

It can send, reply and forward through Gmail, but by default it asks for your approval before each one (Claude Help Center). With Microsoft 365, sending only works if your admin enabled write tools.

Do I need Cowork to use Claude for work?

No. Chat plus connectors and Projects covers most daily work. Cowork is worth it when you regularly need finished files built from several sources, and it needs a paid plan and the desktop app.

Claude or ChatGPT for work?

Both work well. Pick the one your company allows and you will actually open every day. Our side-by-side is in ChatGPT vs Claude AI for work.

Studying for a course or exam on the side? Our guide on how to use Claude to study shows how to turn it into a tutor that quizzes you.

Your next step

Do Step 1 today. Connect one app, run the morning inbox brief tomorrow, and see how it feels by Friday. If you’d like a guided path with one practical AI win for your work at a time, join our free Launchpad on the RobotsHelp home page.

Start using Claude for real work this week.

Sources

• Claude pricing and plan comparison
• Claude Help Center: Use Google Workspace connectors
• Claude Help Center: Connect to Microsoft 365
• Claude Cowork product guide: Getting started
• Best practices for getting started with Claude Cowork
• Claude Cowork product guide: Best practices and tips
• Anthropic Privacy Center: Is my data used for model training?

If your work is advisory, AI tools for consultants walks through six client-facing jobs you can start with this week.

You open your inbox at 8:30 and it already feels like the day has been decided for you. Forty new emails, three of them matter, and you will spend the first hour finding out which three.

Here is the short answer. The email automations that actually save time at work fall into eight jobs: sorting incoming mail, flagging what is urgent, drafting replies to the same questions, answering form sign-ups, chasing proposals nobody answered, sending your weekly numbers, pulling invoices into a sheet, and sending a clean recap after a meeting. Each one is the same three parts: something arrives, an AI step reads or writes, and a draft or label lands where you want it. Below are real email automation examples for each job, the tool to build it in, and which one to start with.

Build your first email automation with us, step by step.

What an email automation really is

Every tool in this article uses the same idea, just with different names. Zapier calls it a Zap: a trigger (the event that starts it, like “new email with this label”) followed by one or more actions (the steps it then does). Zapier explains the pieces in its own guide to key Zap concepts. Make calls them scenarios, n8n calls them workflows. Same shape.

The AI part is simply one step in the middle. It reads the email and decides a label, writes a summary, or drafts a reply. You do not need code for any of the examples below. You need a free account in one tool, access to your mailbox, and twenty minutes of patience for the first one.

8 email automation examples you can build this week

1. Sort incoming mail without any AI at all

Trigger: a new email from a known sender or with certain words. Result: it gets a label, skips the inbox, or is forwarded.

Start here, because it is free and already inside your mailbox. Gmail lets you turn any search into a filter that labels, archives, stars or forwards matching mail (Gmail Help: create rules to filter your emails). Outlook does the same with rules (Microsoft Support: manage email with rules). Newsletters into one label, tool notifications into another, invoices into a third.

One honest catch from Google’s own help page: a new filter only affects new messages, not the pile already sitting there. That is fine. You are fixing tomorrow, not yesterday.

Why it matters for the rest of this list: almost every smarter automation below starts with “a new email gets label X”. A clean label is the cheapest trigger you will ever build.

2. Let AI flag what is urgent and give you one line per email

Trigger: a new email in your inbox (or in one label). AI step: read it and answer two questions. Result: a label like “Reply today” plus a one-line summary sent to you or added to a sheet.

Filters only match words. They cannot tell “can we move Thursday?” from “we are cancelling the contract.” An AI step can make that call. In n8n the Gmail Trigger checks your mailbox on a schedule you choose (the docs call it poll mode, see the n8n Gmail Trigger docs), and the Gmail node can then add the label for you. In Zapier the same idea is Gmail as the trigger and the ChatGPT integration as the step in the middle.

The prompt does most of the work. Keep it boring and strict:

You sort work email for a busy consultant.
Read the email below and answer in exactly this format:
URGENCY: reply-today | this-week | fyi
SUMMARY: one sentence, max 20 words, what they want from me
Only choose reply-today if they ask a question with a deadline in the next 2 working days, or a client is unhappy.

EMAIL:
{{subject}}
{{body}}

Start with one label (for example only mail from clients), not the whole inbox. You will see within a week whether the urgency calls match yours, and you can tighten the rule in the prompt before you widen it.

3. Draft replies to the questions you answer every week

Trigger: a new email that looks like a common question (pricing, opening hours, “where is my order”, “can you send the deck”). AI step: write a reply using your own notes. Result: a Gmail or Outlook draft waiting for you, never sent on its own.

This is where AI email automation earns its keep. You paste your standard answers into the prompt as a short fact sheet, and the AI only uses those facts. You open the draft, fix one word, press send. If you want the hands-on version for a support inbox, we built it step by step in automate customer support with AI agents. For one-off replies you write yourself, the prompts in how to use ChatGPT to draft reply emails faster are quicker than any automation.

4. Send a follow-up email after someone fills in a form

Trigger: a new form submission (Typeform, Google Forms, Tally, your website form). AI step (optional): personalise one sentence based on what they wrote. Result: a reply in their inbox within minutes, from your own address.

People search for exactly this, word for word: “make automation send follow up email after form submission.” In Make it is a scenario with your form app as the first module and the Gmail or Microsoft 365 email module as the second. In Zapier it is a two-step Zap, which even the free plan allows. The AI step is where you go from “Thanks for your message” to “Thanks, I saw you need help with onboarding three new clients a month, here is the checklist we mentioned.”

Two rules keep it human. Send from a real person’s address, and put something useful in the first email (a link, an answer, a booking slot), not just “we received your request.” If the form starts a bigger process, like a new client signing up, the full flow is in how to automate client onboarding with Make.com.

5. Chase the proposal nobody answered

Trigger: a sent proposal with no reply after a set number of days. AI step: write a short, friendly nudge that refers to the actual proposal. Result: a draft follow-up, ready when you open your mail.

Often a proposal does not get a no. It gets silence, and the follow-up you meant to send never goes out. This one has its own full guide on our site, so I will not repeat it here: automate proposal follow-up emails with Zapier (with HubSpot, or just a Gmail label), and for a wider sales sequence, automate follow-up emails without code.

6. Send your weekly performance email from Google Sheets and ChatGPT

Trigger: a schedule, for example every Monday at 8:00. AI step: read last week’s numbers from a sheet and write three plain sentences about what changed. Result: one email to your team or your boss, same format every week.

If you send a “how did last week go” email by hand, you know the drill: open the sheet, copy numbers, write the same intro you wrote last Monday. In Zapier or n8n this is a schedule trigger, a “get rows” step from Google Sheets, a ChatGPT step, and a send-email step. If you prefer to stay inside Google, Apps Script’s MailApp service can send an email straight from a sheet (Google Apps Script: MailApp), though that does mean pasting a small script.

The prompt that keeps it honest:

Here are last week's and the week before's numbers as a table.
Write a weekly update email of max 120 words:
1. One sentence: the biggest change, with both numbers.
2. One sentence: anything that went down, with both numbers.
3. One sentence: one question for the team.
Use only numbers that appear in the table. If a number is missing, say "not filled in".

TABLE:
{{rows}}

That last line matters. You do not want an AI rounding your conversion rate up because it sounds better. For the sheet side of this, how to use ChatGPT to analyze a spreadsheet for work shows how to get trustworthy answers from your numbers first.

7. Pull invoices out of email and into a spreadsheet

Trigger: a new email with a PDF attachment from a supplier. AI step: read the invoice number, date, amount and supplier name. Result: a new row in your bookkeeping sheet, plus the file saved to a folder.

This one removes a job that feels like it has to be done by hand: typing numbers from a PDF into a sheet. We wrote the full build in n8n: n8n workflow for automatic invoice processing. Keep a human check on the amount column for the first month.

8. Send a meeting recap email people actually read

Trigger: a new transcript or notes file after a call. AI step: turn it into decisions, owners and dates. Result: a draft recap email to everyone who was in the meeting.

The trick is the structure, not the tool: decisions first, then who does what by when, then open questions. The prompts for that are in turn meeting notes into a report people can act on. Once the prompt works by hand, wrapping it in a trigger is the easy part.

Which email automation to build first

Do not build all eight. Pick the one that hurts most this week. Three quick questions:

• Is your problem “too much noise”? Start with example 1 (filters), then add example 2 on top of one label.
• Is it “I type the same reply again and again”? Example 3. It is often the quickest win, because you already know the right answers.
• Is it “things fall through the cracks”? Examples 4 and 5. A missed follow-up costs more than a messy inbox.

My own rule: build one, live with it for a week, then build the next. Two automations you trust beat ten you keep checking.

Draft first, send later: keep a person in the loop

Every example above ends in a draft or a label, not a sent email. That is on purpose. An AI that drafts a wrong reply costs you ten seconds. An AI that sends a wrong reply to a client costs you a phone call.

The tools support this. In Gmail and Outlook steps you can choose “create draft” instead of “send”. n8n goes further and lets you require a human approval before an AI agent is allowed to use a tool like “send email” (n8n Docs: human-in-the-loop for tools). Only switch a flow to auto-send when you have read enough of its drafts to stop editing them, and even then, only for low-risk mail like a form confirmation.

What these automations cost

• Gmail filters and Outlook rules: free, part of your mailbox.
• Zapier: the Free plan includes two-step Zaps and 100 tasks per month (Zapier pricing). Enough to test example 4 or 5. A flow with an AI step in the middle has three steps, so that needs a paid plan.
• n8n: a self-hosted Community Edition is available for free, and the hosted plans are paid (n8n pricing). If you have never used it, our n8n tutorial for beginners gets your first workflow running.
• The AI step: if you connect your own ChatGPT or OpenAI account, that usage is billed there, separate from the automation tool.

Prices and limits change often, so check the pricing page on the day you sign up. If you want to see what fits inside Zapier’s free tier, we listed working builds in Zapier free plan automation examples.

Mistakes that make email automation backfire

• Loops. An automation that replies to email can end up replying to its own replies. Always exclude your own address and anything with “Re:” from auto-reply triggers.
• Reply-all by accident. Set the “to” field to the original sender only, unless you really mean everyone.
• Sending private data to the wrong place. Before you pass client email into an AI step, check what your company allows. Start with mail that holds no personal or confidential data.
• Building for the whole inbox on day one. One label, one sender type, one week. Then widen.

Run an online shop? The flows look different there. Here is email automation for ecommerce, with the five flows to set up first.

FAQ

Do I need to code to build these email automation examples?

No. Every example works in Zapier, Make or n8n by clicking steps together. The only one with an optional script is the Apps Script route in example 6, and the Zapier or n8n route does the same job without code.

Can AI send emails automatically on my behalf?

Yes, technically. Our advice is to let it create drafts first, read them for a week or two, and only auto-send low-risk messages like form confirmations.

Does this work with Outlook instead of Gmail?

Yes. Zapier, Make and n8n all have Microsoft Outlook or Microsoft 365 email steps, and Outlook rules do the job of Gmail filters in example 1.

What is the easiest email automation to start with?

A Gmail filter or Outlook rule for your newsletters. Five minutes, no new account, and it makes every other automation easier because your mail is already labelled.

Your next step

Pick one example from this list and build it today, even if it is only the filter. If you would like a guided start, our free Launchpad on the RobotsHelp home page walks you through your first practical AI wins for work, at your own pace.

Start your first email automation today.

Using Outlook? Read about the best AI email assistant for Outlook to see how to speed up replies there.

You watched a demo where an AI agent sorted an inbox, booked a meeting and updated a CRM on its own. Then you opened the first builder you found, saw a blank canvas, and closed the tab. Now there are ten tools that all say “no code” and you have no idea which one is meant for you.

Here is the short answer. The best no code AI agent builder is usually the one that already sits next to your apps. Already on Zapier? Use Zapier Agents. Already on Make? Use Make’s AI Agents. Want help inside your inbox and calendar this week with approval before anything goes out? Try Lindy. Want full control, or need to keep data on your own server? Learn n8n, and accept a steeper first hour. Then test your pick on one real, boring job for 30 minutes before you pay for anything.

Want a guided start with AI at work? Join free:

What a no code AI agent builder actually gives you

An AI agent is a language model with instructions and tools. You tell it the job in plain words, you give it a few actions it may take (read an email, look up a row, draft a reply), and it decides which action to take next until the job is done. The n8n documentation puts it simply: connect a chat model and one or more tools, and the agent decides which tools to call to complete a task (n8n docs).

A no code builder wraps that in a screen you can click through. Instead of writing code, you get four things:

• A place to write the instructions: the job description, in your words.
• Connections to your apps: Gmail, Outlook, Slack, Sheets, your CRM.
• A trigger: what wakes the agent up, like a new email, a form, a schedule or a chat message.
• A log: a record of what the agent did, so you can check it.

That last one matters more than any feature list. If you cannot see what your agent did, you cannot trust it with real work.

The five builders most people end up choosing between

There are dozens of tools. For a professional who wants something working this month, five keep coming up. They split into two groups: automation platforms that added agents (Zapier, Make, n8n), and products built around the agent from day one (Lindy, Relevance AI).

Zapier Agents: if your apps are already connected through Zapier

Zapier’s big strength is reach. Its agents can act across the same app directory as its normal automations, which Zapier puts at more than 9,000 apps (Zapier Agents). You describe the job in plain words and choose which actions the agent may take. If you already have Zaps running, your connections are there, which saves you the most annoying hour of any setup.

The catch: usage is counted in “activities” for agents, separate from tasks for Zaps. The free agent tier gives you 400 activities a month, which is fine for testing but not for a busy inbox (Zapier pricing).

Make AI Agents: if you like seeing every step on a canvas

Make lets you give an agent three kinds of tools: single app modules, whole scenarios, and MCP server tools (Make help center). The scenario option is the clever one. You build a workflow you trust, set it to run on demand, define what goes in and what comes out, and the agent can only call that tested workflow. It cannot freelance outside it.

The catch: you need to be comfortable with Make’s way of thinking. Its free plan includes 1,000 credits a month, and most actions use one credit, so a chatty agent burns through that faster than you expect (Make pricing).

n8n: if you want control, or your data has to stay with you

n8n gives you an AI Agent node on a visual canvas, and you choose every piece: the model, the memory, each tool and what happens after the agent answers. It is the only one of these five you can run on your own server, through the free self-hosted Community Edition. Its cloud plans include human approval for tool calls, so the agent can ask before it acts (n8n pricing).

The catch: n8n is the most “low code” of the group. You will not need to write code for a first agent, but you will meet words like JSON and webhook. If that sounds fine, our n8n tutorial for beginners gets you through the first workflow.

Lindy: if the job lives in your inbox and calendar

Lindy feels less like a builder and more like hiring an assistant. You set it up in plain language, and it handles inbox triage, meeting prep, scheduling and drafting replies in your voice. Lindy says anything with outside impact, like sending an email or updating a ticket, waits for your approval (Lindy pricing and FAQ). For a first agent that touches client email, that default is exactly what you want.

The catch: it is billed per user and per credit, and the free credits last 7 days. Plan your test week before you sign up, so the trial does not run out while you are still reading the settings.

Relevance AI: if you are planning several agents that work together

Relevance AI is built for teams of agents, often for sales and operations work: one agent researches a lead, another writes the outreach, a third updates the CRM. That is powerful, and it is also more than most people need for a first agent. Start here only if you already know you want a small team of agents, not one helper (independent comparison, October 2026).

Which one fits you? A quick way to decide

Don’t spend a week reading comparison tables. Answer these three questions, in this order:

1. Where does the job happen? If it is your own inbox and calendar, start with Lindy. If it moves data between business apps, look at Zapier or Make.
2. What do you already pay for? If your company runs on Zapier or Make, stay there. Your connections, your billing and your colleagues’ know-how are already in place.
3. Does the data have to stay in-house? If a client contract or your IT team says yes, n8n self-hosted is the honest answer, even if it takes longer.

Still stuck between two? Pick the one where you can finish the 30-minute test below without help. Ease at the start predicts whether you will keep using it.

The 30-minute test before you pay for anything

Every builder looks good in its own demo. The only demo that counts is your own work. Take one small, repetitive job you do every week, for example sorting incoming client emails into “needs me today”, “can wait” and “just information”.

Paste this as the agent’s instructions, and adjust the details to your work:

Then run it on 20 real emails from last week, not 2. Check three things:

• Did it pick the right label? Count the misses. Two out of twenty is workable. Eight is not.
• Can you see why it chose what it chose? Open the log. If the reasoning is a black box, that is a warning sign.
• Did it stay inside the fence? It should have saved drafts, never sent them.

If the builder passes, you have a real agent. If it fails on setup, try the next one on your list with the same instructions. Same test, fair comparison. For a full walk-through of building this kind of agent step by step, see how to build an AI agent without code.

Keep a person in charge, whatever you pick

The safety rules are the same in every builder, and they matter more than the brand:

• Give the agent only the tools it needs. An email sorter does not need access to your bank or your calendar.
• Start with drafts, not sends. Let it prepare work. You press the button.
• Add an approval step for anything a customer sees or anything that moves money. Lindy and n8n both offer approvals before actions. In Make, you can route a draft to yourself in Slack first, which Make’s own best practices describe (Make help center).
• Write a “when in doubt” rule. “If you are not sure, ask me” saves you from the confident wrong answer.

I run agents like this every day, and the ones that last are never the clever ones. They are the ones with a narrow job and a clear fence.

Do you even need an agent, or is a simple automation enough?

Often a plain automation is enough, and it is cheaper and easier to check. If the steps are the same every time (new form entry, add to sheet, send a thank-you email), you do not need an agent. A fixed workflow with one AI step does the job. Our guide to Zapier free plan automation examples shows plenty of those.

Reach for an agent when the steps change from case to case: triage, research, booking, answering questions. That is where the “decide what to do next” part earns its cost. Good examples are an AI agent for appointment booking and automating customer support with AI agents.

Picked n8n? Our n8n AI agent workflow guide shows the whole build for one real job, a meeting-prep agent, including where to put human approval.

Quick answers

Is there a free no code AI agent builder?

Yes, for testing. Zapier has a free agent tier with 400 activities a month, Make has a free plan with 1,000 credits a month, n8n’s Community Edition is free if you host it yourself, and Lindy gives free trial credits for 7 days. For daily work you will likely move to a paid plan, so check the pricing page on the day you sign up. Plans change often.

Is the AI model included, or do I need my own API key?

It depends on the tool and plan. n8n’s cloud plans list AI models without API keys, while self-hosted n8n asks you to bring your own key. Check this before you compare prices, because a model bill on top of a plan can double the real cost.

Can a no code agent work with my existing automations?

That is the main reason to pick the platform you already use. Make agents can call your existing scenarios as tools, and Zapier agents use the same app connections as your Zaps. If you are weighing those two for client work, our comparison of n8n vs Zapier for agency workflows goes deeper.

What about an agent that works on my desktop instead of in the browser?

That is a different kind of tool, one that clicks through apps on your computer. If that is what you need, read our guide to the best AI agent for desktop workflows.

Your next step

You do not need the perfect builder. You need one agent that takes one boring job off your plate, tested on twenty real cases, with a fence around it. Pick the builder that sits closest to your apps, run the 30-minute test tonight, and decide with real results in front of you.

If you want help taking steps like this, one practical lesson at a time, start with the free RobotsHelp Launchpad.

Sources

• n8n documentation: AI Agent node
• n8n pricing and plan features
• Zapier Agents and Zapier pricing
• Make help center: AI agent best practices and Make AI Agent (New) best practices
• Make pricing
• Lindy pricing and FAQ
• Stepthrough: AI agent builders compared (October 2026)

If you would rather learn the ideas behind these tools, our AI agents course guide lists the best free courses.