Author: Pete

You wrote it fast. Now you read it back and wince. It sounds blunt, or it rambles, or the irritation from your last call is sitting right there in the second line. The client is waiting, and you have twelve other things open.

Here is the fix. To make this email sound more professional, paste your draft into ChatGPT, Claude or Gemini and give it three things: who will read it, what must not change (dates, prices, promises), and how long it may be. Then ask for the rewrite plus a short list of what it changed. That last part is the trick. You check the list in ten seconds instead of comparing two emails line by line, and nothing slips in that you did not mean.

Want one practical AI lesson like this every week?

The prompt that makes an email sound more professional

“Make this more professional” on its own is a coin toss. The AI has to guess what is wrong, so it changes everything: your words, the length, sometimes the facts. Often the email comes back longer and stiffer than what you wrote. Both OpenAI and Anthropic say the same thing in their own guides: be specific about what you want, and you get a better answer (OpenAI, Anthropic).

So tell it exactly what to do. Copy this once and keep it somewhere you can grab it:

Rewrite the email below so it sounds professional and calm, but still like a real person wrote it.

Reader: [who it goes to, e.g. "a client I have worked with for two years"]
Goal: [what you need from them, e.g. "approve the new delivery date"]

Rules:
- Keep every name, date, number, price and promise exactly as written.
- Do not add new facts, apologies or commitments.
- Do not make my request softer or stronger than it is.
- Put the main point or request in the first two lines.
- Keep it about the same length or shorter.
- Plain words. No "I hope this email finds you well".

After the email, give me a numbered list of what you changed and why.

Email:
[paste your draft]

Fill in the two brackets at the top. They do most of the work. “A client who has complained twice” and “my manager” need different emails, and the AI can only know that if you tell it.

What “professional” actually means in a work email

Professional does not mean formal. It does not mean longer either. When people ask for a more professional email, they usually want five things:

• The point first. What you need, and by when, sits in the opening lines.
• Exact details. “Thursday 22 October, before 12:00” instead of “soon”.
• No filler. Cut “just wanted to”, “sorry to bother you” and the third “thanks in advance”.
• Calm words. No capitals for emphasis, no sarcasm, no “as I said before”.
• A clear next step. The reader knows what to do after the last line.

Here is what that looks like. A typical rushed draft:

hey, so the designs are gonna be late again, the printer messed up our files and we're waiting on them. probably end of next week?? sorry!! let me know if thats a problem

And what the prompt above gives back, more or less:

Hi Sarah,

The designs will be later than planned. Our printer had a problem with the files, and we are waiting for them to resend. Our best estimate is the end of next week.

Does that date work for you? If it causes a problem on your side, tell me and we will look at what we can move.

Best,
Tom

Look at what stayed the same: the delay, the reason, the rough date. Look at what changed: the point is up front, the panic is gone, and there is a real question at the end. Notice also that “probably” became “our best estimate”. The AI did not turn it into a promise. That is exactly what the “do not make it stronger” rule is for. If it had written “the designs will be ready Friday”, the change list would show it and you would catch it.

Rewrite an angry email in a professional tone with ChatGPT

This is the one most people need and least like to admit. Someone dropped the ball, it cost you an evening, and your draft shows it.

Do this:

1. Write the angry version first. In a notes app, not in your mail app. Then there is no Send button anywhere near it.
2. Paste it with this prompt:

I am frustrated and this draft shows it. Rewrite it in a calm, professional tone.

Keep: the facts of what went wrong, the impact on me, and what I need from them now.
Remove: blame, sarcasm, capitals, exclamation marks and anything about their character.
Keep it firm. Do not apologise for raising the problem.

Draft:
[paste]

3. Read it once and ask: would I be fine if this got forwarded to their boss? If yes, it is ready.
4. Give yourself a way back. In Gmail you can set Undo Send to 30 seconds under Settings (Gmail Help). Small thing, but it saves you on exactly this kind of email.

The “keep it firm” line matters. Without it, AI rewrites of angry emails can swing too far the other way, into soft, apologetic mush where your actual complaint disappears. Professional is calm and clear. It is not giving in.

Make a long email shorter and clearer

The other common problem: you explained everything, in the order it happened, and the request is in paragraph four. Nobody reads paragraph four.

Make this email shorter and clearer, under 120 words.

- Start with what I need from the reader and by when.
- Keep only the background they need to act.
- If there are three or more items, use a short bullet list.
- Keep every date, number and name exactly as written.
- List anything you cut, so I can check it was safe to cut.

Email:
[paste]

Ask for a word count. “Shorter” without a number can come back only a little shorter. And the “list what you cut” line is your safety net: if the AI dropped the one sentence about the budget limit, you will see it right there.

Do the same in Claude: rewrite or write the client reply

Everything above works in Claude too, word for word. Paste the prompt, paste the draft. Where Claude is handy is the second job.

Rewrite your email to sound more professional in Claude

Same prompt as the first one. If you send a lot of client mail, save the rules once instead of pasting them every time. Claude lets you add account-wide instructions under Settings, and on paid plans you can give a Project its own instructions that only apply inside it (Claude Help Center). Make a project called “Client email”, put the rules in there, and from then on you only paste the draft.

Write a professional reply to a client with Claude

Sometimes you have nothing yet. The client wrote, and you are staring at it. Paste their email and give Claude your answer in rough notes:

Here is an email from a client. Write my reply.

My answer in rough notes: [e.g. "yes to the extra page, costs 350 extra, can start Monday, need their logo files first"]

Tone: professional and friendly, short paragraphs, no hype.
Only use facts from my notes. If something is missing, write [CHECK] instead of guessing.
Under 150 words.

Client email:
[paste]

The [CHECK] instruction is the part to keep. It stops the AI from filling a gap with a date or price you never agreed to. If you write a lot of replies like this, our guide to drafting reply emails faster with ChatGPT goes deeper on the reply side, and for moments that repeat (a delay, a payment reminder, a scope change) see client communication templates that sound like you.

Professional email closing lines that do not sound stiff

The last two lines are where a lot of emails go wrong. “Please do not hesitate to contact me should you have any further questions” says nothing, and it takes eleven words to say it. A good ending has two parts: a closing line that tells the reader what happens next, then a short sign-off.

Closing lines that do a job:

• “Can you confirm by Wednesday so we can book the slot?”
• “If Thursday at 10:00 works, I will send the invite.”
• “No reply needed. This is just to keep you in the loop.”
• “Once I have your files, we start the same day.”

Sign-offs, and when they fit:

• Best, works almost everywhere.
• Kind regards, a bit more formal. Good for a first contact or a complaint.
• Thanks, only when you are actually asking for something.
• Speak soon, for someone you already work with often.

Want the AI to handle it? Add one line to any prompt above: “End with one sentence that says the next step, then sign off with ‘Best,’ and my name.” And if every email you send ends the same way, ask it for three different closing lines and pick one.

If you write in Gmail, Gemini can polish a draft without leaving the compose window. Our guide on how to use Gemini in Gmail shows where to find it and how to check what it writes.

The buttons already in your inbox

You might not need a separate chat window at all.

• Gmail: Gemini’s “Help me write” can refine a draft you wrote yourself, with options like Polish, Formalize, Shorten and Elaborate (Gmail Help). Depending on your account and country, it may need a Google AI or Workspace plan.
• Outlook: “Coaching by Copilot” reads your draft and suggests changes to tone, clarity and how the reader is likely to take it, and you can apply them all at once (Microsoft Support). It needs a Microsoft 365 plan that includes Copilot.

These are great for a quick polish. The catch: you cannot tell a single button “keep the price and do not soften my request”. For the emails that matter, a chat with the full prompt gives you more control. Not sure which assistant to use for what? Here is a plain ChatGPT vs Gemini vs Copilot comparison.

The 30-second check before you hit send

The AI did the wording. You still own what it says. Before you send, look at:

• The change list. Anything on it you did not want?
• Names, dates, numbers. Same as in your draft?
• Certainty. Did “might” become “will”, or “around Friday” become “Friday”?
• Your voice. Would you say this out loud to that person? If it sounds like a press release, ask for “more plain, less formal”.

One more thing: privacy. Do not paste anything your company does not allow in outside tools, like client contracts or personal data. In ChatGPT you can stop your conversations from being used to train its models under Settings, Data controls, “Improve the model for everyone” (OpenAI Help Center). If your employer gives you an approved work account, use that one for work email.

Set it up once, use it every day

The real time saver is not one good rewrite. It is never typing the rules again.

• ChatGPT: put your email rules in Custom Instructions under Settings, Personalization. They apply to all your chats, and they are available on every plan (OpenAI Help Center).
• Claude: a project with the rules as its instructions, as described above.
• Anywhere: a pinned note with the prompts from this page.

Add a few sentences from an email you actually sent and liked, with the line “match this voice”. Then the rewrite sounds like you on a good day, not like a template. Our guide to an AI writing assistant for busy professionals shows how to build that one-page voice brief.

Writing to a client in another language? Draft the email in your own language first, then follow our guide on how to use AI to translate a document so the tone and your key terms survive the switch.

A cover letter is the email where tone matters most. If you are applying for a job, see how to use AI to write a cover letter that sounds professional and still sounds like you.

Quick answers

Is it okay to use AI to fix my work emails?

For tone and clarity, yes, as long as you check the result and follow your company’s rules on which tools you may use. You are the one who sends it, so you are the one who signs off on every fact in it.

Does this work on the free plans?

Yes. Pasting a draft and asking for a rewrite works on the free versions of ChatGPT and Claude (ChatGPT plans, Claude plans). The built-in buttons in Gmail and Outlook are the parts that may need a paid plan.

Why does the AI make my emails longer?

Because you did not give it a limit. Add “same length or shorter” or a word count, and tell it to skip pleasantries like “I hope this email finds you well”.

Can I use these prompts in other languages?

Yes. Write the prompt in English or in your own language, and add “reply in [language]”. Check the names and numbers just as carefully.

Pick the email that has been sitting in your drafts the longest. Paste it in with the first prompt on this page. You will have it sent before your coffee goes cold.

Want the next one in your inbox?

Sources

• Prompt engineering best practices for ChatGPT (OpenAI Help Center)
• ChatGPT Custom Instructions (OpenAI Help Center)
• Data controls in ChatGPT (OpenAI Help Center)
• ChatGPT plans and pricing (OpenAI)
• Be clear and direct (Anthropic prompt engineering docs)
• Understanding Claude’s personalization features (Claude Help Center)
• Claude pricing (Anthropic)
• Draft emails with Gemini in Gmail (Gmail Help)
• Send or unsend Gmail messages (Gmail Help)
• Get email coaching with Copilot in Outlook (Microsoft Support)

Work in Outlook? See our guide to the best AI email assistant for Outlook for ways to get the same polish inside your inbox.

Your company just switched on Copilot. Gmail keeps offering to “help me write” with Gemini. And you already pay for ChatGPT, or nearly did last month. Three AI assistants, one workday, and nobody has told you which one to open when a real job lands on your desk.

Here is the short answer. ChatGPT vs Gemini vs Copilot is less about which AI is smartest and more about where your work lives. Pick the one that sits closest to it. If your day runs on Word, Excel, Outlook and Teams, Copilot is the one that can work inside those files without copy-paste. If your day runs on Gmail, Docs and Drive, Gemini is the natural fit. If your work is spread over many apps, or you mostly start from a blank page, ChatGPT is the strongest all-rounder to open in its own window. All three have a free version, and the main personal paid plans all sit close to $20 a month. One rule beats all of this: if your employer has approved one of them for work files, use that one for work files.

Want one practical AI lesson like this every week?

ChatGPT vs Gemini vs Copilot at a glance

Benchmark scores tell you little at a normal desk. Five questions decide it: where it lives, what it costs, what it can see, what happens to what you paste, and what it is best at.

Question	ChatGPT	Gemini	Copilot
Made by	OpenAI	Google	Microsoft
Lives best in	Its own app on web, desktop and phone, with extensions for Excel, Word, PowerPoint and Google Sheets	The Gemini app, Gmail, Docs, Drive and Chrome	Word, Excel, PowerPoint, Outlook, Teams and Windows
Free version	Yes	Yes	Yes, the Copilot app, plus Copilot Chat with an eligible work account
Main personal plan	Plus, $20 a month	Google AI Pro, $19.99 a month	Microsoft 365 Premium, $19.99 a month
Team version	ChatGPT Business and Enterprise	Gemini Enterprise, Business edition from $21 per seat a month	Microsoft 365 Business Standard with Copilot, $23.50 per user a month paid yearly
Best at	Starting from a blank page, thinking through messy problems, data analysis	Very long documents and work that already sits in Gmail or Docs	Your own Office files, emails and meetings, with your company’s permissions

Prices are US list prices from each company’s own pages, checked on 19 October 2026: ChatGPT pricing, Google AI plans, Copilot for individuals and Copilot for business. Your local price and tax will differ, and all three companies change their plans often, so check again before you pay.

ChatGPT: the all-rounder you open in its own window

ChatGPT is the one most people have already tried, and it is at its best when the work starts outside a document. Messy notes from a call. A plan that is still half in your head. A CSV export you need to understand before Friday. You paste it in, explain what you need, and work it out together.

The Plus plan costs $20 a month, billed monthly, according to OpenAI’s help center. On top of the free version, OpenAI’s pricing page lists projects, scheduled tasks, custom GPTs, deeper research and full use of its extensions for Excel, Word, PowerPoint and Google Sheets. There is also a cheaper Go plan with higher limits than Free, which OpenAI says may include ads.

Where it falls short: ChatGPT does not see your inbox or your company drive unless you connect it or paste things in. Every paste is a small time tax, and a small privacy decision. If you want to see what it does with real numbers, our guide on how to use ChatGPT to analyze a spreadsheet for work walks through it step by step, and summarizing meeting notes with ChatGPT is a good first ten-minute win.

Gemini: the one already sitting in Gmail and Docs

If your company runs on Google Workspace, Gemini has a head start: it can help right inside the email or document you have open. No copying a 40-message thread into another tab.

Google’s plans page lists three personal levels:

• Free: the Gemini app, Deep Research and Canvas, with lower limits.
• Google AI Plus ($4.99 a month): higher limits and Gemini inside Gmail and other Google apps.
• Google AI Pro ($19.99 a month): Gemini in Gmail, Docs and more, a 1M token context window, file uploads of up to 1,500 pages, and 5 TB of storage.

That long-document headroom is the practical difference. A full tender, a year of board minutes or a thick policy manual can go in as one upload. Google also notes the personal plans need a personal Google account you manage yourself, so a work account follows whatever your Workspace admin has set up.

Gemini is a strong fit for repeat reporting too. Our Gemini status update guide shows how to turn a week of notes into an update your manager actually reads.

Copilot: the one inside Word, Excel, Outlook and Teams

Copilot confuses people because “Copilot” is really three things:

1. The free Copilot app at copilot.cloud.microsoft, in Windows and in Edge. Good for general questions, writing help and web answers.
2. Copilot in personal Microsoft 365 plans (Personal at $9.99, Family at $12.99, Premium at $19.99 a month). The subscription owner gets Copilot in the Word, Excel, PowerPoint and Outlook apps, with more usage on Premium.
3. Copilot for work. Copilot Chat comes at no extra cost with eligible Microsoft 365 business plans. The full Copilot license adds Copilot inside Teams and the Office apps, grounded in your emails, meetings and files.

Microsoft explains the split in its free vs Microsoft 365 Copilot article. To add to the confusion, Microsoft has now renamed “Microsoft 365 Copilot” to “Microsoft Copilot”, so you will see both names for a while.

The big strength of the work version is context. Microsoft says it respects your identity, permissions and sensitivity labels, so it can pull an answer from the emails, files and meetings you already have access to, instead of from whatever you paste in. The weakness is the flip side: outside the Microsoft world it has much less to work with, and the features you get depend heavily on your license. Ask IT which one you have before you judge it.

Most of your Copilot time spent in Excel goes on formulas? How to use AI to write Excel formulas shows a prompt that works in Copilot, ChatGPT or Gemini, plus a quick way to check the answer.

What each one really costs per month

Before you buy anything, check what you already pay for. If you have Microsoft 365 or extra Google storage, you may already have more AI than you think.

• $0: all three have a free version that is fine for testing and light, non-confidential work.
• Under $10: Google AI Plus at $4.99, Microsoft 365 Personal at $9.99, and ChatGPT Go.
• Around $20: ChatGPT Plus at $20, Google AI Pro at $19.99, Microsoft 365 Premium at $19.99. Note that the Microsoft and Google plans also include Office apps or cloud storage, so you are not only paying for the AI.
• For a team: ChatGPT Business is priced per seat, Gemini Enterprise Business starts at $21 per seat, and Microsoft 365 Business Standard with Copilot is $23.50 per user a month on a yearly plan.

Privacy: what happens to what you paste

This is the part that can get you in trouble at work. The personal and work versions of each tool follow different rules.

• ChatGPT, personal plans: new chats can be used to train OpenAI’s models unless you turn off “Improve the model for everyone” under Settings, then Data controls. Temporary chats are not used for training. OpenAI says content in Business, Enterprise and Edu workspaces is not used for training by default (OpenAI data controls).
• Gemini, personal account: while “Keep Activity” is on, Google uses your chats to improve its AI, and some are read by human reviewers. Google’s own advice is not to enter anything confidential you would not want a reviewer to see. Temporary chats are not used to train its models (Gemini Apps Privacy Hub).
• Copilot with a work account: Microsoft’s enterprise data protection says prompts, responses and the work data Copilot reads are not used to train its foundation models (Microsoft Learn).

A simple rule that keeps you employed: client data, contracts and anything about colleagues only go into the work version your company approved. Everything else, test wherever you like.

The 20-minute test: one real task, three assistants

Reviews tell you which tool is “best”. Only your own work tells you which one is best for you. Here is a test you can run today on the free versions.

1. Pick one task from last week that took you more than 30 minutes. Turning a long email thread into a status update works well. Remove names and anything confidential first.
2. Write one prompt (use the one below) and paste the same prompt and material into ChatGPT, Gemini and Copilot.
3. Score each answer on three things: how much you had to fix, how much copying it took to get your material in and the result out, and whether it invented anything.
4. Repeat with a second task from a different part of your week, for example a spreadsheet question.
5. Pick a winner per task type. Having one main assistant and one built into your office apps is completely normal.

You are helping me with a work task.
Task: Turn the email thread below into a status update for my manager.
Context: My manager cares about deadlines, budget and risks. The project is [one line about the project].
Format: At most 5 bullet points, then one line with the decision I need from them.
Rules: Do not invent dates, numbers or names. If something is missing, ask me before you write.
Thread:
"""
[paste the thread here]
"""

Watch the copying score closely. A slightly better answer that costs you three extra pastes every time is often the worse deal over a month.

Do you need different prompts for ChatGPT, Claude and Gemini?

Mostly, no. The companies’ own guides give almost the same advice. OpenAI’s prompt best practices say be clear and specific, give context and refine step by step. Google’s Gemini prompt guide for Workspace builds a prompt from persona, task, context and format. Anthropic’s Claude prompting docs cover clarity, examples, roles and structuring your input with XML tags. The prompt above follows all three.

What changes is the setup around the prompt, not the prompt itself:

• ChatGPT: put your standing context (role, clients, tone) in a project, so you stop retyping it.
• Claude: wrap pasted material in simple tags, like <thread> and </thread>, so it never mixes up your instructions and your material.
• Gemini and Copilot inside your apps: skip the paste. Open the email or document and ask about “this thread” or “this document”.

If you want a reusable brief that works in any of them, the voice brief in our AI writing assistant desk setup is a good one to copy.

ChatGPT vs Claude vs Gemini vs Perplexity vs Copilot: where the other two fit

Two more names keep coming up in this comparison, and they do slightly different jobs.

• Claude (Anthropic) is a strong choice when your work is long documents and careful writing. Anthropic lists a context window of up to 1M tokens, depending on the model, and the Pro plan costs $17 a month billed yearly or $20 monthly. It also includes Claude Code and Claude in Microsoft 365 (Claude pricing). Our guide to analyzing long PDF files with Claude shows that in practice.
• Perplexity is built for research. It answers from the web with citations in every answer, so you can click through and check. Pro starts at $17 a month billed yearly (Perplexity pricing). Use it when the job is “find out”, not “write this”.

You do not need five subscriptions. Choose one main assistant by where your work lives, then add a second only when a job keeps coming up that your main one handles badly. If you are curious about the agent side, our comparison of OpenClaw vs Claude Code picks up where chat assistants stop.

If the job you keep coming back to is writing small scripts, our Claude vs Gemini for coding guide puts the two side by side on one real task, from a Google Sheets script to a folder of CSV exports.

Torn between ChatGPT and Claude in particular? Our ChatGPT vs Claude on Reddit guide sums up what people who use both every day say, and gives you one prompt to test them side by side.

If you also use X and need answers about what happened today, our ChatGPT vs Gemini vs Grok guide shows where Grok fits, what the new X Subscription includes, and a 15-minute fresh-news test.

Curious how Perplexity holds up next to ChatGPT and Claude on a real research job? Our ChatGPT vs Claude vs Perplexity comparison walks through it with a 20-minute demo and current prices.

Making slide visuals or social images with your assistant? Our best AI tool for image generation guide compares ChatGPT Images and Gemini with specialists like Ideogram, Midjourney and Adobe Firefly, with a 20-minute test to find your own winner.

So which one should you pick?

• Your company runs Microsoft 365 and has Copilot switched on: use Copilot for work files. Add a personal tool only for non-confidential work.
• Your company runs Google Workspace: start with Gemini, because it already sits in your mail and documents.
• You are a freelancer or use a mix of tools: ChatGPT Plus is the safest all-rounder. Choose Claude instead if most of your day is long documents.
• You are not sure yet: run the 20-minute test on the free versions before you pay for anything.

Once you have picked, the real time savings come from doing the same few jobs with it every week. If you want an agent that clicks through desktop apps for you as the next step, see our guide to the best AI agent for desktop workflows.

Quick answers

Is Copilot just ChatGPT inside Microsoft?

No. Copilot is Microsoft’s own product. Microsoft lists “LLM model choice” for its business Copilot, and its documentation mentions Anthropic models running inside Copilot, so it is not tied to one model.

Can I use the free versions for work?

For non-confidential work, yes. For client or company data, use the work version your employer approved. Free personal accounts may use your chats to improve the models unless you change the settings above.

Which one is best for Excel?

If your files live in Excel and you have Copilot through Microsoft 365, it works inside the sheet. ChatGPT Plus also has an Excel extension and good data analysis if you prefer to work in its own window.

Do I need all three?

No. One main assistant, plus at most the one built into the apps your company already uses, covers almost every desk job.

Want a new hands-on AI workflow every week?

Sources

• ChatGPT plans and pricing (OpenAI)
• What is ChatGPT Plus? (OpenAI Help Center)
• Data controls in ChatGPT (OpenAI Help Center)
• Prompt engineering best practices for ChatGPT (OpenAI Help Center)
• Google AI plans (Google)
• Gemini Apps Privacy Hub (Google)
• Google Workspace with Gemini prompt guide (Google)
• Copilot plans for individuals and Copilot plans for business (Microsoft)
• Microsoft Copilot (free) vs Copilot in Microsoft 365 (Microsoft Support)
• Enterprise data protection in Microsoft Copilot (Microsoft Learn)
• Claude pricing and prompt engineering overview (Anthropic)
• Perplexity pricing (Perplexity)

If your inbox is Outlook, our guide to the best AI email assistant for Outlook shows where Copilot fits and what a free setup can do.

You keep seeing OpenClaw and Claude Code in the same threads, both called “AI agents”, both run on your own computer, and you only have time to learn one properly this month. Which one actually saves you hours at work?

Here is the short answer. They do different jobs. OpenClaw is a free, open-source assistant you run on your own machine and message from WhatsApp, Telegram, Slack or Discord, so it can handle reminders, chat requests and scheduled jobs around the clock. Claude Code is Anthropic’s agent for working inside a folder of code: it reads the project, edits files, runs commands and explains or fixes bugs. If your work lives in chat apps and repeat tasks, start with OpenClaw. If your work lives in scripts, spreadsheet macros, websites or any codebase, start with Claude Code. Many people end up using both, and they can even run together.

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

OpenClaw vs Claude Code at a glance

Most comparisons list features. For someone who is not a full-time developer, five questions decide it: where you talk to it, what it costs, what it can touch, how much setup it needs, and what happens when you are not at your desk.

Question	OpenClaw	Claude Code
What it is	A self-hosted “gateway” that connects your chat apps to an AI agent	A coding agent that works inside a project folder
Where you talk to it	WhatsApp, Telegram, Slack, Discord, iMessage, Teams and more, plus a browser dashboard	Terminal, VS Code or Cursor, JetBrains, the Claude desktop app, or claude.ai/code in the browser
Which AI model	Your choice: Claude, GPT, Gemini or a local model	Claude models
Cost of the tool	Free and MIT licensed; you pay your model provider	Included in paid Claude plans (Pro from $17 a month billed yearly, $20 monthly) or billed per use through an API account
Best at	Always-on help you reach from your phone: reminders, scheduled checks, chat requests	Building and fixing things made of code: scripts, sites, automations, data tools
Setup	One install command, a guided onboarding, an API key, then connect a chat app	One install command or the desktop app, then sign in with your Claude account

Details come from the OpenClaw documentation, the Claude Code overview and the Claude pricing page, checked in October 2026. Both tools ship updates every week or two, so check the pages again before you commit a budget.

What OpenClaw actually does

The OpenClaw docs describe it as “a self-hosted gateway that connects your favorite chat apps” to AI agents. In plain words: you run one small program on your laptop or a cheap server, link it to a chat app, and from then on you message your assistant like you message a colleague.

What that gets you at work:

• You reach it from anywhere. Send “summarize the PDF I just forwarded” from Telegram on the train, and the answer comes back in the same chat.
• It works on a schedule. OpenClaw has a built-in scheduler, so it can post a Monday-morning summary in your Slack or remind you of a deadline without you asking each time (see the automations docs).
• It remembers and learns procedures. It keeps sessions and memory, and “skills” teach it a repeatable routine it loads when needed.
• You pick the brain. It works with Claude, OpenAI, Google models or a model that runs on your own hardware, so you are not tied to one company.

The trade-off is that you are the IT department. It needs Node.js on your machine (the docs list Node 24.16 or newer), an API key from your model provider, and a computer that stays on if you want it available all day. The OpenClaw Foundation that develops it says there is no paid tier and no hosted version, so nobody runs it for you.

What Claude Code actually does

Anthropic calls Claude Code “an agentic coding tool that reads your codebase, edits files, runs commands, and integrates with your development tools.” You open it inside a folder, describe what you want in normal sentences, and it plans the work, changes the files and checks the result.

You do not need to be a software engineer to get value from it. Typical work jobs it handles well:

• Writing a small script that renames 400 files, merges CSV exports or cleans a messy spreadsheet.
• Fixing the contact form on your WordPress site, or explaining why a Google Apps Script stopped running.
• Turning a manual weekly report into a script you run with one command.
• Reading a project someone else built and telling you what each part does.

It runs in the terminal, inside VS Code, Cursor or JetBrains, in the Claude desktop app and in the browser at claude.ai/code. Its usage comes out of the same pool as your normal Claude chats, according to the Claude pricing FAQ, and it is not part of the free plan.

Weighing Claude Code against Google’s tools instead? Our Claude vs Gemini for coding comparison shows which one fits work in files on your laptop and which fits work inside Google Sheets and Apps Script.

Deciding between Claude Code and OpenAI’s Codex instead? Our guide on ChatGPT vs Claude for coding explains what each one does on your files and which plan you need.

Can Claude explain this code and suggest fixes?

Yes, and this is where most non-developers start. The Claude Code docs say you can paste an error message or describe the symptom, and it traces the issue through the project, finds the cause and proposes a fix. The difference with pasting code into a normal chat window is that Claude Code can open the other files involved, run the script and confirm the fix works before it tells you it is done.

A prompt that works well when you inherit something you do not understand:

Explain what this project does in plain English, as if I am not a developer.
Then look at the error below, tell me the most likely cause in two sentences,
and suggest a fix. Do not change any files until I say "go".

Error:
[paste the error message here]

The last line matters. Asking it to explain first and wait for your “go” keeps you in control of what changes. When you do say go, ask it to run the script again and show you the output, so you see the fix with your own eyes.

Cost: what each one really costs you per month

OpenClaw costs nothing to install. What you pay is the AI model behind it, billed by your provider per amount of text processed. A light personal assistant that answers a few dozen messages a day stays cheap; one that reads long documents all day or runs many scheduled jobs costs more. Set a monthly spending limit in your provider’s dashboard before you connect it, so a busy week never surprises you.

Claude Code comes with a Claude Pro plan ($17 a month billed yearly or $20 monthly) or a Max plan from $100 a month for heavier use, per the Claude pricing page. Every plan has usage limits that reset in five-hour windows, and your chats and coding share them. For most people who build a few small tools a week, Pro is the place to start.

If you already pay for Claude Pro, Claude Code is the cheapest way to try an agent this week: nothing extra to buy.

Safety: what each one can touch

Both tools can run commands on your computer, so treat them like a new hire with keys to the office.

OpenClaw ships with careful defaults. The OpenClaw security docs explain that the gateway only listens on your own machine by default, unknown people who message it get a pairing code instead of an answer, and group chats are limited to an allowlist. They also say it is built for one trusted person or a team that trusts each other, not for sharing one assistant with strangers, and that an agent with message access can send across your connected chats by default. Run openclaw security audit after setup and fix what it flags.

Claude Code works inside the folder you open it in. In its default Manual mode it asks before it edits a file or runs a shell command, as the Claude Code permissions docs show, and you can loosen or tighten that in its settings. The simple habit: open it in a copy of your project, or in a folder backed up with Git, so any change can be undone.

One rule for both: never give an agent a password or bank login it does not need for the task in front of it.

Scheduling and working while you are away

This used to be the clearest difference, and it is narrowing. OpenClaw was built to run all the time and act on a schedule or an incoming message. Claude Code now has scheduled “routines” that run in Anthropic’s cloud, desktop scheduled tasks, and “Channels” that push messages from Telegram, Discord or iMessage into a session, as listed in its overview.

The difference that stays: Claude Code’s scheduled work is still centred on a code project, while OpenClaw is centred on you and your chats. “Every Friday, check last week’s failed form submissions and fix the script” is a Claude Code job. “Every Friday at 4, send me my open follow-ups on WhatsApp” is an OpenClaw job.

Using OpenClaw and Claude Code together

You do not have to choose forever. The OpenClaw docs describe two ways to run Claude Code from inside OpenClaw: as a CLI backend that handles normal agent turns, or as an external coding session you start and track from chat.

In practice that means you can message your OpenClaw assistant from your phone, “the export script broke again, fix it”, and Claude Code does the coding part while OpenClaw reports back in the chat. That setup is for the second month, not the first day. Learn one tool well first.

We work this way at RobotsHelp too: a small team of AI agents picks up tasks from a ticket board and reports back, and the lesson that carried over is simple. Give each agent one clear job, and keep a human check before anything goes out.

Which one should you start with?

Start with OpenClaw if most of your day happens in chat apps and email, you want an assistant you can text from your phone, you want reminders and scheduled summaries, or you want to use models from more than one company. The docs promise a basic install in minutes; plan an hour or two to connect a chat app, lock down who can message it and run the security audit, and keep a computer running.

Start with Claude Code if you already pay for Claude, you have a script, spreadsheet tool, website or small app that needs building or fixing, or you keep pasting code into chat windows and copying answers back. You can be useful with it within the hour.

A quick test: write down the last three things you wished an AI could do for you. If they sound like “remind me”, “tell me when” or “send me”, that is OpenClaw. If they sound like “build”, “fix” or “automate this file”, that is Claude Code.

Not ready for either? A no-code tool may be the better first step. Our guide on how to build an AI agent without code gets you a working agent this week without a terminal, and the n8n vs Zapier comparison helps if your jobs are mostly “when this happens in app A, do that in app B”. If you want an agent that clicks through desktop apps for you, see the best AI agent for desktop workflows.

Your first 30 minutes with each

OpenClaw

1. Install it with the one-line command from the OpenClaw docs (there is a separate command for Windows PowerShell).
2. Run the onboarding it starts for you and pick Quick start. Have your model provider’s API key ready.
3. Open the browser dashboard and send a test message.
4. Connect one chat app. The docs say Telegram is the fastest to set up.
5. Give it one real job: “Every weekday at 8, send me the three most important things in my notes file.”

On a Windows PC? Our guide on how to install OpenClaw on Windows compares the four routes (the Windows Hub app, PowerShell, WSL2 and the backup desktop app) and walks through each one step by step.

On a Mac? Here is how to install OpenClaw on a Mac, with the menu bar app for macOS Tahoe 26.2 or later and the one-line Terminal installer for older Macs.

Claude Code

1. Install it with the command on the Claude Code overview, or download the Claude desktop app and open the Code tab.
2. Make a new folder and put a copy of one real file in it, for example last month’s messy CSV export.
3. Open Claude Code in that folder and sign in with your Claude account.
4. Ask: “Look at this file, tell me what is messy about it, then write a script that cleans it. Explain before you change anything.”
5. Run the script, check the result, and save the prompt that worked for next month.

If you mostly need to get through long documents rather than build tools, regular Claude chat may be all you need: see how to use Claude to summarize a long report.

Quick answers

Is OpenClaw better than Claude Code?

Neither is better in general. OpenClaw is better as an always-on assistant you reach through chat. Claude Code is better at reading, writing and fixing code. Pick by the job, not the hype.

Is OpenClaw free?

The software is free and MIT licensed. You still pay for the AI model it uses, unless you run a local model on your own hardware.

Do I need to know how to code for Claude Code?

No, but you need to be willing to open a folder, read what it proposes and say yes or no. Ask it to explain every step in plain English and you learn as you go.

Can OpenClaw use Claude models?

Yes. Anthropic is one of the providers OpenClaw supports, and it can also hand coding work to Claude Code itself.

Comparing OpenClaw to Claude for everyday desk work instead of coding? Our OpenClaw vs Claude Cowork guide covers the non-coding side: spreadsheets, reports and chat-app assistants, with the same cost and safety checks.

Want one hands-on AI lesson like this each week?

You have three clients, and every one of them wants “some automation”. One lives in Gmail and HubSpot, one wants AI to sort support tickets, one sends 2,000 leads a month. You do not want to learn two tools badly. You want to pick the right one and build.

Here is the short answer. Pick Zapier when the client should own and edit their automations in their own account, the workflows are fairly simple, and you hand the work over at the end. Pick n8n when you run the workflows for clients as an ongoing service, the workflows have many steps or run often, and you want to pay per run instead of per step. The rest of this page shows you how to make that call for each client in about ten minutes.

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

n8n vs Zapier for agency workflows at a glance

Most comparisons stop at “Zapier is easier, n8n is cheaper”. For an agency the real differences sit somewhere else: who owns the account, how you are billed when one client gets busy, and how you hand the work over.

Question	Zapier	n8n
What you pay for	Tasks: each successful action step	Executions: one full run of a workflow, however many steps
Free option	Free plan, 100 tasks a month, two-step Zaps	Self-hosted Community edition (you run the server)
Cheapest paid entry	Pro from $19.99 a month (annual, 750 tasks)	Cloud Starter from 20€ a month (billed annually)
Working as a team	Team plan: 25 seats, shared Zaps and shared app connections	Unlimited users on every Cloud plan, work grouped in projects
Client edits it themselves	Easy: it lives in their own Zapier account	Only in their own n8n, never in your instance
Moving work to a client	Copy assets to another account, then reconnect apps	Export workflow JSON, import it, add credentials again

Prices and limits are from the Zapier pricing page and the n8n pricing page as of October 2026. Both change often, so check them on the day you quote a client.

Tasks vs executions: what one busy client costs you

This is the difference that hits an agency’s margin first, so it is worth doing the sum once.

Zapier counts a task every time an action step finishes successfully. Triggers do not count, and neither do its built-in helpers such as Filter, Formatter, Paths and Delay. That is all on the Zapier pricing page.

n8n counts one execution per full run of a workflow. Ten steps or two, it is still one execution, as the n8n pricing FAQ explains.

Take a normal lead intake workflow for a client:

1. Trigger: new form entry.
2. ChatGPT writes a two-line summary and a score.
3. Create or update the contact in the CRM.
4. Post the summary in the client’s Slack channel.
5. Send the lead a short “thanks, we got it” email.

That is one trigger and four actions. In Zapier, 300 leads a month is 300 × 4 = 1,200 tasks, so you need the 1,500 task tier of Pro ($39 a month on annual billing). In n8n, the same 300 leads are 300 executions.

Now picture the same agency with five clients like this one. Zapier counts every step of every run, so the task number grows with each step you add. n8n only grows with the number of runs. That is why agencies with long workflows lean toward n8n, and agencies with short, simple ones often never feel the difference.

One more Zapier detail that matters when a client has a launch week: if you go over your limit, pay-per-task billing can keep the Zaps running at a higher per-task rate (2.5 times the base rate on monthly plans, 1.25 times on annual), and usage stops at 3 times your subscription. With it switched off, the Zaps simply stop until next month. Decide which one you want before the busy week, not during it.

Who owns the account? Decide this before you choose the tool

Ask every client one question: “When we are done, do you want to change these automations yourself?” The answer usually picks the tool for you.

Model 1: the client owns it (build and hand over)

The client pays for their own account, you build inside it, and they keep it when the project ends. Zapier is the comfortable fit here. Its step-by-step editor is easy for a non-technical client to follow, and if you built a Zap in your own account first, Zapier’s Copy to another account feature moves Zaps, Tables, Forms and Agents across. Know what comes with it: app connections must be authenticated again, copied Zaps arrive switched off, run history stays behind, and webhook URLs change. Put “reconnect apps and turn Zaps on” on your handover checklist.

You can also export Zaps as a JSON file and import them in another account (Settings, then Security and data), as described in Zapier’s import and export guide. Import is not available during a free trial.

Model 2: you run it for them (managed service)

You keep the workflows on your side, the client gets the results (the Slack alert, the updated CRM, the draft email), and you charge a monthly fee for running and fixing them. This is where n8n shines, because one busy client does not multiply your bill by the number of steps.

If the client later wants their own copy, n8n saves workflows as JSON. You download it from the workflow menu and they import it into their instance, as shown in n8n’s export and import docs. The file holds credential names and IDs, but not the passwords or keys themselves, so the client connects their own accounts after import. Read the file before you send it: n8n warns that HTTP Request nodes imported from cURL can carry authentication headers.

The n8n license rule every agency needs to know

Self-hosted n8n is free under the Sustainable Use License. For agencies, the line is clear once you read n8n’s license FAQ:

• Allowed: building, running and maintaining workflows for clients on your own instance, and charging for setup, maintenance, training and consulting. There is no limit on the number of clients.
• Allowed: clients connecting their own accounts (their OAuth login or API key) to workflows you built, as long as they cannot see or change the workflow itself.
• Allowed: installing and managing n8n on a client’s own server, as long as you are not also hosting it for them.
• Not allowed without a commercial license: letting clients build or edit workflows in your instance, whether through the n8n editor, your own screen, the API, MCP or an AI agent acting for them. Also not allowed: hosting n8n as a service for others, or white-labelling it.

In plain words: clients get outputs, not access. If a client wants to tinker with their own automations, give them their own n8n (Cloud or on their server), or use Zapier in their own account. The n8n Help Center says the same and points to license@n8n.io when you are unsure.

Keeping clients apart: folders, projects and shared connections

Mixing up two clients’ Gmail connections is the kind of mistake you only make once. Here is what each tool gives you to prevent it.

Zapier. Pro has one seat. The Team plan adds up to 25 seats, a shared workspace, folder permissions and shared app connections, so a colleague can use a client’s connection without ever seeing the password. Separate governed workspaces with their own apps and permissions are an Enterprise feature (all from the Zapier plan table). On Team, a folder per client with clear connection names (“Acme Gmail”, “Acme HubSpot”) is the practical way to stay tidy.

n8n. n8n groups workflows and credentials in projects, and each user can have a different role per project. That fits “one project per client” nicely, but count them: the pricing page lists 1 shared project on Starter, 3 on Pro, 6 on Business and unlimited on Enterprise. The free self-hosted Community edition has no projects and no workflow or credential sharing at all, according to n8n’s edition comparison. On Cloud plans you can share a credential with a project so teammates can use it without seeing the details (credential sharing).

So a solo freelancer with eight clients can live fine on one Community instance with good naming. A three-person agency that wants teammates to share client logins safely needs a paid plan on either side.

Adding ChatGPT to client workflows in each tool

Most agency work now has an AI step in the middle: summarise the email, score the lead, draft the reply. Both tools can do it, with different strings attached.

In Zapier you have two routes. The ChatGPT (OpenAI) app needs an OpenAI API account with prepaid billing; a ChatGPT Plus subscription does not include API access (Zapier’s ChatGPT setup guide). Or you use AI by Zapier, which works without your own key but is only on paid plans and is priced per AI step by model tier. If you build for clients, decide up front whose OpenAI account pays for the tokens, and put it in the proposal.

In n8n the OpenAI node uses an API key credential (n8n OpenAI credentials). n8n Cloud plans also list “AI models without API keys”. The AI step does not change your bill on the n8n side: it is still one execution per run. The model provider’s token bill still comes on top.

Whichever tool you use, give the AI step one tight job and a fixed output. A prompt like this works for the lead intake example:

You are helping a small agency sort new leads for the client "{client name}".
Read the form entry below. Reply with exactly three lines:
SCORE: hot, warm or cold
WHY: one sentence, based only on what the lead wrote
NEXT: the one thing the client should do today

Rules: do not guess a budget or company size that is not in the text.
If the entry is spam or empty, reply SCORE: cold and WHY: spam or empty.

Form entry:
{form fields}

Fixed lines like SCORE and NEXT are easy to map into a CRM field or a Slack message in both tools. For a full scoring build, see our guide to an AI agent for lead qualification in n8n and Zapier.

Want to see an AI step working inside a Zap before you pitch it to a client? Our guide to Zapier AI automation walks through two builds, a Slack summary and an AI follow-up email in Gmail.

Zapier free plan limitations for ChatGPT powered workflows

Agencies often test a client idea on a Free Zapier account first. That is a fine way to learn, but know where the walls are before you promise anything:

• Two steps only. Free Zaps are one trigger and one action. “New email, then ChatGPT” works, but you cannot then send the ChatGPT answer anywhere in the same Zap. Most useful AI workflows need at least three steps.
• 100 tasks a month, shared across all your Zaps.
• 15-minute polling. Fine for a daily digest, too slow for a “reply to a hot lead within five minutes” promise.
• No Filters, Paths or Formatter, so you cannot skip spam or clean up a messy date before it reaches ChatGPT.
• No AI by Zapier and no webhooks. The ChatGPT (OpenAI) app itself is not a premium app, so Free users can add it (Zapier support confirms this in the Zapier Community), but you still pay OpenAI for API usage.
• No autoreplay or custom error settings, so a failed run just fails.

The plan details come from the Zapier feature table. Free is good for proving one idea to yourself. For client work, budget for Pro from day one. If you want ideas that do fit inside Free, our list of Zapier free plan automation examples keeps every build to two steps.

Build the same client workflow in both and compare

If you are still unsure, spend one afternoon building the lead intake workflow above in both tools with a dummy form and a test Slack channel. Then score them on what matters to your agency, not on what a review site says:

1. Time to first working run. Note it honestly, including the time you spent looking for a setting.
2. Could the client follow it? Show the screen to someone non-technical and ask them to explain what happens. If they cannot, plan on managing it for them.
3. What happens on bad input? Send an empty form and a spam entry. Does the workflow stop cleanly or post junk in Slack?
4. Cost at the client’s real volume. Steps times runs for Zapier, runs for n8n, plus your AI tokens.
5. Handover. Try moving it to a second account. That is the job you will do at the end of every project.

Use the free trials for this test: n8n Cloud’s trial gives you 1,000 executions, and Zapier gives new accounts a 14-day Pro trial, both without a credit card according to their pricing pages. For more n8n practice, the n8n CRM automation guide walks through four client-style workflows, and the n8n invoice processing workflow is a good test of a longer, multi-step build.

A quick decision checklist per client

Run each new client through these five questions. Two or more answers in one column usually settle it.

• Will the client edit the automations themselves after you leave? Yes: Zapier in their own account. No: n8n is open.
• Do the workflows have five or more steps, or run hundreds of times a month? Yes: n8n usually costs less.
• Are you selling a monthly “we run your automations” retainer? Yes: n8n on your side, within the license rules above.
• Does the client use niche apps? Check both app directories before you quote. Zapier lists 9,000+ apps. n8n can also call an app’s API directly with its HTTP Request node, which takes more setup.
• Does the client’s data need to stay on their own servers or in the EU? n8n can be self-hosted, and n8n Cloud data is stored in Frankfurt, Germany, per its pricing FAQ.

And you do not have to pick one tool for the whole agency. Plenty of client lists split naturally: the hands-on clients on Zapier, the “just make it work” clients on your n8n.

Weighing Make or Lindy too, or building an AI agent instead of a fixed workflow? See how to choose a no code AI agent builder for the kind of work your clients need.

If a client asks for an AI assistant they can message, rather than a fixed workflow, see OpenClaw vs Claude Code before you promise either one.

Mistakes that cost agencies time

• Building in your own account with your own logins, then “handing over”. Every connection breaks the day you disconnect. Build with the client’s own app accounts from the start, or plan the reconnect step.
• Quoting a flat monthly fee without counting tasks. One client’s busy month can eat your margin on a per-task plan. Do the steps times runs sum per client.
• Letting a client into your n8n editor “just to look”. If they can edit, you are outside the free license. Give them screenshots, a run log, or their own instance.
• No owner for errors. Pick who gets the failure email for each client workflow, and check it weekly.
• Vague connection names. “Gmail 3” is how a client email goes out from the wrong account. Name every connection after the client.

FAQ

Is n8n really free for an agency?

The self-hosted Community edition is free, also for client work, as long as clients only receive outputs and cannot build or edit workflows in your instance. You still pay for the server and your own time to keep it updated. n8n Cloud is a paid subscription with its own terms.

Which one is easier to teach to a client?

For most non-technical clients, Zapier. Its editor reads like a list of “when this happens, do that” steps, and they can log in to their own account and switch a Zap off without calling you. n8n gives you more control, and that control shows up as more settings on the screen.

Can I move a client from Zapier to n8n later?

Plan on rebuilding. Zapier exports Zaps as JSON in its own format, and n8n imports its own workflow JSON, so the files do not swap across. The good news: the logic you mapped out (trigger, AI step, actions) carries over, and that is the hard part.

What is the cheapest way to start testing both?

Use both free trials at once on one dummy workflow, as in the comparison above. You will learn more in that afternoon than from any feature table, this one included.

If you want help building your first client-ready workflows step by step, the Launchpad is where we start.

New to n8n and want to try it before you pick a side? Our n8n tutorial for beginners walks you through your first working workflow step by step.

Sources

• Zapier: Pricing, plan feature table and task usage
• Zapier Help: Copy assets to another account
• Zapier Help: Import and export Zap workflows
• Zapier Help: How to get started with ChatGPT (OpenAI) on Zapier
• Zapier Community: ChatGPT on a Free account
• n8n: Plans and pricing, executions FAQ
• n8n Docs: Community license FAQ (agency examples)
• n8n Help Center: Which license do I need for my use case?
• n8n Docs: Compare editions
• n8n Docs: Organize work in projects
• n8n Docs: Share credentials securely
• n8n Docs: Export and import workflows
• n8n Docs: OpenAI credentials

It’s the end of the month. Somewhere in your inbox are 20 supplier invoices, a few of them forwarded twice, one of them a phone photo of a paper receipt, and you need all of them in one sheet before your bookkeeper asks again. Here’s the short answer. An n8n workflow for automatic invoice processing watches Gmail for PDF attachments, lets an AI model (Gemini works well) read each invoice into fixed fields, writes one row per invoice to Google Sheets, labels the email so it’s never processed twice, and asks you before anything big gets marked as approved. You build it once in about an afternoon, test it on ten old invoices, and from then on your job shrinks to checking rows instead of typing them.

Below is the full build, node by node, plus what to do with scanned invoices and how to keep it cheap and safe.

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

What this workflow does (and what it should never do)

Look at the invoice workflows people share in n8n’s template library and the same proven shape keeps coming back: poll Gmail for emails with PDF attachments, pull the text out of the PDF, have an AI model return the invoice as structured fields, check for duplicates, append the result to a Google Sheet, and optionally route it to someone for approval (n8n template: invoice data from Gmail to Sheets, n8n template: supplier invoices with duplicate checks). We’ll use that shape and make it a bit more forgiving for a small business: it handles scans, it marks what it isn’t sure about, and it never touches money.

That last part matters. This workflow logs and flags. It does not pay, it does not email suppliers, and it does not change anything in your accounting software. You still decide what gets paid. The time you save is the typing and the hunting, which is usually the part nobody enjoys anyway.

Here is the whole thing in one view:

1. Gmail Trigger: picks up new emails with a PDF attached.
2. Extract From File: pulls the text out of the PDF.
3. Information Extractor (with a Gemini chat model): turns that text into fixed fields.
4. Edit Fields: adds a unique invoice key and a status.
5. Google Sheets: writes or updates one row per invoice.
6. Gmail: labels the email as done.
7. Optional: an approval email to you for anything above your limit.

Before you build: three things to set up

1. A sheet with fixed columns

Create a Google Sheet called “Invoices” with these headers in row 1: invoice_key, supplier_name, invoice_number, invoice_date, due_date, currency, subtotal, tax_amount, total, vat_number, notes, status, email_subject. Fixed columns are what make the rest work. The AI fills the same boxes every time, so you can sort, filter and total them later.

2. A Gmail label called invoice-logged

The workflow adds this label to every email it has processed. Your trigger then skips anything with that label, which is your first protection against doing the same invoice twice.

3. Your credentials in n8n

You need a Google credential for Gmail and Sheets, and a Gemini API key. n8n’s docs walk you through the Gemini key: you create it on the API key page in Google AI Studio and paste it into a “Google Gemini(PaLM) API” credential (n8n Docs: Google Gemini credentials). On n8n Cloud you can also pick “Use Gateway credits” instead of your own key. If Gmail throws a 401 “unauthorized_client” error later, the usual fix is enabling the Gmail API in your Google Cloud project (n8n Docs: Gmail Trigger common issues).

Build the invoice workflow in n8n, node by node

Step 1: Gmail Trigger that only catches invoices

Add a Gmail Trigger with the event “Message Received”. Set the poll time to something calm, like every 15 minutes. Invoices are rarely urgent to the minute.

Under filters, put this in Search (these are normal Gmail search operators, so you can test the exact line in the Gmail search bar first, see Google: Refine searches in Gmail):

has:attachment filename:pdf -label:invoice-logged (invoice OR factuur OR rechnung OR facture OR bill)

Then change two settings that trip almost everyone:

• Read Status. The default is “Unread emails only” (n8n Docs: Gmail Trigger). If you glance at an invoice on your phone before the next poll, it’s read, and the workflow never sees it. Set it to unread and read emails, and let the label do the skipping.
• Simplify off, Download Attachments on. Turn Simplify off, then under Options switch on Download Attachments. The first attachment lands in a binary field called attachment_0, the second in attachment_1, and so on (that naming comes straight from the node’s code, see n8n on GitHub: Gmail Trigger node).

Max Emails per Poll defaults to 10 (maximum 50). Extra emails wait for the next poll, so a busy month-end just takes a few more rounds.

Tip: if you can, ask suppliers to send invoices to a separate address like invoices@yourcompany or set a Gmail filter that forwards them to one inbox. A narrow inbox beats a clever search line.

Step 2: Extract From File (PDF)

Add Extract From File and choose the operation “Extract From PDF”. Set Input Binary Field to attachment_0 instead of the default data (n8n Docs: Extract From File). Run it once on a test email. For an invoice made in accounting software you’ll see the invoice text in the output, messy but complete.

If it comes back empty or nearly empty, that PDF is a scan or a photo. Skip ahead to the scanned invoice section below, then come back.

Step 3: Information Extractor with a Gemini chat model

This is the AI step. Add the Information Extractor node and attach a Google Gemini Chat Model to it. In Text, use {{ $json.text }}, which is exactly the example n8n gives for input coming from Extract From PDF (n8n Docs: Information Extractor).

For Schema Type, choose “Generate From JSON Example” and paste this. n8n only looks at the field names and types, not the values:

{
  "supplier_name": "Brightline Print Ltd",
  "invoice_number": "INV-2041",
  "invoice_date": "2026-10-02",
  "due_date": "2026-11-01",
  "currency": "EUR",
  "subtotal": 400.00,
  "tax_amount": 84.00,
  "total": 484.00,
  "vat_number": "NL000000000B01",
  "notes": ""
}

One detail from the docs worth knowing: with a JSON example, n8n treats every field as mandatory. Real invoices often miss a due date or a VAT number, so tell the model what to do when a field isn’t there. Under Options, open System Prompt Template and use something like this (n8n adds its own formatting instructions on top):

You read supplier invoices for a small business.
Copy every value exactly as it is printed. Never guess or fill in a value that is not on the invoice.
Dates: write them as YYYY-MM-DD.
Amounts: plain numbers with a dot as decimal separator, no currency symbol, no thousands separator. So "1.234,50" becomes 1234.50.
Currency: the three-letter code (EUR, GBP, USD).
If a field is not on the invoice, return an empty string (or 0 for an amount) and say which field is missing in "notes".
If subtotal plus tax does not equal the total, write "totals do not add up" in "notes".
If this is not an invoice at all (a quote, a newsletter, a delivery note), write "not an invoice" in "notes".

The number rule is the one that saves you the most cleanup. European invoices write one thousand two hundred thirty-four euros and fifty cents as 1.234,50, while a sheet expects 1234.50.

Step 4: Edit Fields to add a key and a status

Add an Edit Fields (Set) node. Drag in all the extracted fields, then add three of your own:

• invoice_key: drag in supplier_name, type a |, then drag in invoice_number. So “Brightline Print Ltd|INV-2041”. This is how you recognise the same invoice when it arrives twice.
• status: the fixed text to check. Every row starts there. You change it to “approved” or “paid” yourself.
• email_subject: drag in the subject from the Gmail Trigger, so you can find the original email in seconds.

Step 5: Google Sheets, Append or Update Row

Add a Google Sheets node with the operation Append or Update Row. It updates an existing row or adds a new one if no matching row is found. Pick your Invoices sheet, choose Map Each Column Manually, and set Column to Match On to invoice_key (n8n Docs: Google Sheets sheet operations). Now a forwarded copy of the same invoice lands on the same row instead of creating a second one. That’s your second layer against double entries, and the one that catches the supplier who sends the same invoice again with “reminder” in the subject.

Step 6: Gmail, add the label

Add a Gmail node, Resource Message, Operation Add Label. For Message ID, drag in the id from the Gmail Trigger, and pick your invoice-logged label (n8n Docs: Gmail message operations). Put this node last, so an email only gets the label once its row is actually in the sheet. If something fails halfway, the email stays unlabelled and gets picked up again.

Step 7 (optional): an approval email for big invoices

Add an If node after the sheet: is total greater than your limit, say 1000? On the true branch, add a Gmail node with the operation Send and Wait for Approval, sent to yourself, with the supplier, number, total and due date in the message. Set Type of Approval to “Approve and Disapprove”. The workflow pauses until you click. After the click, a second Google Sheets node (Update Row, matched on invoice_key) sets the status to “approved” or “rejected”. n8n describes this operation as the right fit for simple approvals and points to the Wait node for anything more complex.

Still no payments. “Approved” in a sheet just means you looked at it. Paying stays in your bank or accounting tool, done by a person.

What about scanned invoices and phone photos?

A scanned PDF is basically a picture of a page, so there may be no text inside for Extract From File to pull out. This is where Gemini earns its place. Google’s documentation says Gemini processes PDFs with native vision, so it understands text, tables and images on the page, not just a hidden text layer, for files up to 50 MB or 1000 pages (Google AI for Developers: Document understanding).

In n8n you use the Google Gemini node with the operation Document > Analyze Document (n8n Docs: Google Gemini node):

1. After Extract From File, add an If node with a Number condition: {{ $json.text.trim().length }} is less than 50. That’s your “this is a scan” test.
2. On the true branch, add Google Gemini, Analyze Document. Set Input Type to Binary File(s) and the field name to attachment_0.
3. As the text input, write: “Transcribe all text on this invoice exactly as printed, including every amount, date and number. Keep the table rows on separate lines.”
4. Raise the maximum length under Options. In the node’s code the default is 300 tokens, which can cut off an invoice with many line items. 1500 is a safer start.
5. Send the result into the same Information Extractor, so both routes end up in identical columns.

Google’s own tips for better results are simple: rotate pages the right way up and avoid blurry pages. If you photograph paper receipts, do it flat, in good light, and save the result as a PDF (use your phone’s scan-to-PDF option if it has one), so the filename:pdf filter in your trigger actually catches it.

Can Gemini automate invoice processing for a small business?

Yes, with one honest correction. Gemini is the part that reads the invoice. It doesn’t watch your inbox, it doesn’t own a spreadsheet, and it doesn’t remember which invoices it already saw. That’s the job of the workflow around it, which is why n8n and Gemini work well together: n8n does the plumbing, Gemini does the reading.

For a small business, three things decide whether it’s a good fit:

• Your volume. If you get five invoices a month, a manual routine is fine. Drag the PDFs into a chat and ask for the fields, like in our guide on extracting data from a PDF into a spreadsheet with Claude. Once invoices arrive every week from different suppliers, the workflow pays for its afternoon of setup.
• Your data and the free tier. Google’s pricing page lists “Content used to improve our products” under the Gemini API free tier (Gemini API pricing). Invoices carry supplier names, bank details and amounts. For real business invoices, use the paid tier, and check the current rates on that page before you start, because they change.
• Your accountant’s needs. Ask which fields they actually want before you build the sheet. Adding a column later is easy, re-reading 200 old invoices is not.

Test it on ten real invoices before you trust it

Don’t switch it on and walk away. Pick ten invoices from last month: a few from software, at least one scan, one in a foreign currency, one with a credit note or discount. Forward them to the inbox the trigger watches, run the workflow, and compare every row against the PDF, column by column.

What usually needs fixing after the first test:

• Supplier names that vary. “Brightline Print Ltd” and “Brightline Print” break your duplicate key. Add a line to the system prompt: “Use the legal company name as printed at the top of the invoice.”
• Totals versus amounts due. Some invoices show a total and a lower “amount due” after a deposit. Decide which one you want and say so in the prompt.
• Non-invoices. Quotes and order confirmations slip through the search. Filter rows with “not an invoice” in notes, or add an If node that stops them before the sheet.

Then test the duplicate guard on purpose: forward the same invoice twice and check you still have one row. Once ten in a row come out right, activate the workflow. Keep checking the “to check” rows for the first month. You’re no longer typing, you’re reviewing, and that takes a few minutes instead of an evening.

Once the data is in, you can ask questions of it at month end, like spend per supplier or what’s due next week. Our guide on analyzing a spreadsheet with ChatGPT shows how to do that without formulas.

What it costs to run

n8n itself can be free: the self-hosted Community edition is free to use indefinitely (n8n Docs: Compare editions). If you’d rather not run a server, n8n Cloud has a 14-day trial with Pro features and a limit of 1000 executions, then paid plans (n8n Docs: Try free then choose a plan). Every run of the workflow counts as an execution, so testing on a few dozen invoices sits well inside that trial.

The AI part is billed per token by Google. A one-page invoice is small: Google’s docs count each PDF page as 258 tokens, and for Gemini 3 models you aren’t charged for tokens from the text already embedded in a PDF. Check the pricing page for today’s numbers rather than trusting any figure in a blog post, including this one.

Prefer Zapier or Make? The same seven steps work there too. Our Zapier free plan automation examples and the Make.com client onboarding build show how those tools think about triggers and steps. And if your CRM is already running on n8n, the n8n CRM automation guide uses the same Gmail and Sheets credentials, so you’ve done half the setup.

Four mistakes that make invoice automation fail

1. No label, no key. Without the invoice-logged label and the invoice_key match, a re-sent invoice becomes a second row and sooner or later a second payment.
2. Letting it run unread-only. You read the email on your phone, the trigger skips it, and the invoice quietly goes missing from the sheet.
3. Trusting it from day one. AI reads invoices well, not perfectly. The “to check” status is there for a reason.
4. Letting the workflow pay or reply. Keep money and supplier contact with a person. An automation that logs wrong is a small fix. One that pays wrong is a phone call to your bank.

FAQ

Do I need to know how to code to build this?

No. Every step above is a node you configure with dropdowns and drag-and-drop. The only things that look like code are two short expressions in double curly brackets, which you can copy as they are, and the JSON example, which is just a list of field names.

Can n8n read invoices that aren’t PDFs?

Extract From File also handles formats like CSV, XLSX and plain text files. Word files aren’t on its list, so ask those suppliers for a PDF. For photos, the Google Gemini node has an Analyze Image operation that works the same way as Analyze Document. Most suppliers send PDFs, so start there and add other routes only when you actually receive them.

What if two suppliers use the same invoice number?

That’s why the key combines supplier name and invoice number. “INV-1001” from two different suppliers gives two different keys, so two rows.

Build it this week

Make the sheet and the label tonight. Tomorrow, build steps 1 to 6 and run them on ten old invoices. Add the scan route and the approval email only when you need them. The first month-end you spend checking rows instead of typing them is when you’ll know it was worth it.

Want a new practical AI workflow like this every week?

Never built anything in n8n before? Start with our n8n tutorial for beginners, then come back and set up the invoice flow.

Still deciding between building this yourself and paying for a ready-made tool? Our guide to AI invoice processing software compares the main types of tool and shows how to test one on your own invoices before you pay.

Sources

• n8n workflow template: Extract invoice data from Gmail PDF attachments to Google Sheets
• n8n workflow template: File supplier invoices from Gmail to Google Sheets with Gemini
• n8n Docs: Gmail Trigger
• n8n Docs: Gmail Trigger common issues
• n8n on GitHub: Gmail Trigger node source
• n8n Docs: Extract From File
• n8n Docs: Information Extractor
• n8n Docs: Google Sheets sheet operations
• n8n Docs: Gmail message operations
• n8n Docs: Google Gemini node
• n8n Docs: Google Gemini credentials
• n8n Docs: Compare editions
• n8n Docs: Try free then choose a plan
• Google AI for Developers: Document understanding
• Google AI for Developers: Gemini API pricing
• Gmail Help: Refine searches in Gmail

You’ve seen the demos. Someone on YouTube has an AI agent “running their whole business” while they sip coffee. You don’t want that. You want one thing to stop eating your morning: the inbox, the lead form, the same five questions every week. Here’s the honest answer. To build an AI agent without code, you pick one boring job you already do by hand, write it down the way you’d brief an intern, give the agent a trigger, two or three tools and one rule for when it must stop and ask you, then test it on real examples before it touches anything live. In Zapier, Make or n8n that is an afternoon, not a project.

Imagine next Monday. You open Gmail and the 60 emails from the weekend are already sorted: four marked “Today”, a handful of drafts waiting for a quick read, newsletters filed out of sight. You didn’t write a line of code. You wrote a good brief.

Want more practical AI workflows like this one, free? Join Launchpad.

What an AI agent actually is (and when you don’t need one)

Strip away the hype and an agent is a language model that can use tools and decides for itself which ones to call. Anthropic, the company behind Claude, draws a useful line: workflows follow a path you set in advance, while agents direct their own process and tool use (Anthropic: Building effective agents). n8n describes its AI Agent node the same way: you connect a chat model and one or more tools, and the agent decides which tools to call to finish the task (n8n Docs: AI Agent node).

That freedom is the point, and also the risk. Anthropic recommends finding the simplest solution possible, which sometimes means not building an agent at all, and warns that agents cost more and can stack up errors. Make’s documentation gives a practical rule of thumb you can use today (Make Help: Introduction to Make AI Agent):

• Same input, same output, every time? Use a normal automation. Copying form entries into a sheet doesn’t need judgment.
• Fixed steps, but one step writes or summarizes text? Use an automation with a single AI step.
• The job needs a judgment call on messy input? That’s agent territory. Sorting emails, screening leads, categorizing tickets.

Make adds one more line worth taping to your screen: give your agent tasks you would trust an intern with, and keep it away from sensitive data, high-stakes money decisions and strict legal work.

The five parts of every no-code AI agent

Every builder uses slightly different words, but your agent always needs the same five things. Get these clear on paper and the tool part is mostly clicking.

1. Trigger: what wakes it up. A new email, a new form entry, a schedule, a message in Slack.
2. Instructions: the job description. What it does, in what order, and what “done” looks like. Zapier calls this the prompt, Make calls it instructions.
3. Tools: the actions it may take. Add a label, create a draft, look up a row, post a message. Fewer is better.
4. Knowledge: reference material it can check, such as your price list or an FAQ document. Optional for a first agent.
5. A human checkpoint: the moment it must stop and ask you. Zapier has a per-tool switch for this; elsewhere you build it with a draft or an approval step.

Step 1: Pick one job, not “my inbox”

The first agent fails when the job is too big. “Handle my email” isn’t a job. “Read each new email, put it in one of four labels, and draft a reply when someone asks for a meeting time” is a job. A good first job ticks all four boxes:

• You do it at least a few times a day, so you’ll notice the time back.
• You can explain the decision in a few sentences.
• A mistake is cheap and easy to undo (a wrong label, not a wrong invoice).
• You can check the result in under a minute.

Email triage ticks every box, which is why we use it as the demo below. If your pain is somewhere else, the same steps work for leads, bookings or support, and we’ve written those builds out in full further down.

Step 2: Write the brief before you open any tool

This is the step people skip, and it’s the one that decides whether your agent works. Open a blank document and fill in this template. It takes ten minutes and you’ll paste it straight into the builder later.

ROLE: You are my [job] assistant. You work on [what comes in].

GOAL: For each [item], decide [the decision] and then [the action].

OPTIONS (pick exactly one):
1. [Option A]: when [clear rule]. Example: [real example].
2. [Option B]: when [clear rule]. Example: [real example].
3. [Option C]: when [clear rule]. Example: [real example].
If none fits, choose [safe default] and say why in one sentence.

ACTIONS YOU MAY TAKE: [tool 1], [tool 2]. Nothing else.

NEVER: send anything to anyone outside [me / my team]. Never delete.
If you are unsure, stop and leave it for me.

OUTPUT: [the fields you want back, e.g. category, one-line summary].

The examples matter more than the rules. OpenAI gives the same advice for its own assistants: when the AI has to apply classifications, include short examples of right and wrong outputs (OpenAI Help Center: Creating and editing GPTs). Use three real emails, leads or tickets from last week, with names removed.

How to build a no-code AI agent for email triage

Here’s the demo, built with AI by Zapier inside a normal Zap. One note before you start: Zapier is moving its standalone Agents product (agents.zapier.com) into the AI by Zapier step in the Zap editor. It hasn’t set a shutdown date, but a new build belongs in the Zap editor (Zapier Help: Migrating from Agents to AI by Zapier).

Before you build: make the labels

In Gmail, create four labels: AI/Today, AI/This week, AI/FYI and AI/Admin. The agent only files mail into labels you made yourself, so you can always see and undo what it did.

Build it, step by step

1. Trigger: create a new Zap and choose Gmail, New Email. If you get a lot of newsletters, choose New Email Matching Search instead and filter them out there, so you don’t pay for the agent to read them. Both triggers are in Zapier’s Gmail integration (Zapier: Gmail integrations).
2. Add the AI step: add an AI by Zapier action step and paste your brief into the Prompt field. Use Add input field to pass in the sender, subject and body from the trigger.
3. Pick the model tier: only the Advanced and Premium tiers can use tools. On Zapier’s Free plan those tiers are locked, so plan on a paid Zapier plan for this build and check the current pricing first (Zapier Help: Add tools to your AI by Zapier step). Sorting emails is not hard reasoning, so try Advanced first.
4. Add the label tool: under Tools and Knowledge, click + Add tool, choose App action, Gmail, Add Label to Email. Map the message ID from the trigger as a set field, and set the label field to Let agent choose from list with only your four AI labels in it. Now it physically can’t invent a fifth label.
5. Add the draft tool: add Gmail Create Draft Reply. Map the thread ID from the trigger and let the agent determine the body. A draft sits in your Drafts folder until you press send.
6. Optional ping: add Slack Send Direct Message to yourself, and tell the agent to use it only for AI/Today emails.
7. Set the output fields: category, urgency and a one-line summary. Zapier returns them as clean fields you can use in later steps, for example to log every decision in a Google Sheet.

Notice what’s missing: no Send Email, no Reply to Email, no Delete Email. They exist in the Gmail integration, and that’s exactly why you leave them out for now. If you later add a tool that sends or changes something important, switch on Require approval before running for that tool. The run then pauses and Zapier emails you a link to approve or reject it. That switch is off by default, so you have to turn it on yourself.

Here’s a filled-in brief you can paste and adjust:

ROLE: You are my email triage assistant. You read each new email in my work inbox.

GOAL: Put every email in exactly one label, then draft a reply only where described below.

LABELS:
1. AI/Today: a client, my manager or a colleague needs an answer or decision from me today, or there is a deadline within 48 hours. Example: "Can you confirm the quote by tomorrow 10:00?"
2. AI/This week: a real person needs something from me, but not today. Example: "When you have a moment, could you review the draft?"
3. AI/Admin: invoices, receipts, contracts, HR or system notices. Example: "Your invoice INV-2207 is attached."
4. AI/FYI: newsletters, updates, cc's where nobody asks me anything.
If unsure between two labels, choose the more urgent one and say why in the summary.

DRAFTS: Only when someone asks for a meeting or call time. Draft a short, friendly reply that proposes two time slots and asks which works. Sign it with my first name. Never promise anything else.

NEVER: send, forward or delete any email. Never draft replies to invoices, contracts or anything about money.

OUTPUT: label, urgency (high/normal/low), one-sentence summary of what the sender wants.

Preview, then test, then publish

Click Preview and Zapier shows two things: the output fields and a Tools used list with every tool the agent called and whether it worked. Read that list. If it labelled a newsletter as Today, fix the instruction or add an example, then refresh. When the preview looks right, run Test Zap end to end. Zapier notes that full-Zap tests currently use tasks just like live runs, so don’t test 50 times for fun.

Two cost guardrails are built in. Each run has a task limit (75 by default, adjustable up to 500), and if a run hits it, the Zap pauses and asks you instead of looping. And every run lands in Zap history with the tools called, the model tier and the tasks used, so you can see exactly what one email costs you before the bill arrives.

The same agent in Make or n8n

The five parts don’t change. Only the buttons do.

• Make: agents live inside scenarios. Add your Gmail trigger, then the Make AI Agent (New) > Run an agent module, write your instructions, hover over its plus icon to add Gmail modules as tools, and test it with the built-in Chat before it goes live (Make Help: Create your first AI agent). The new agent app is in open beta and runs on all plans through Make’s own AI provider; connecting your own OpenAI, Gemini or Claude account needs a paid plan. If you’ve already built a scenario in Make, for example our client onboarding automation, you’ll recognise the canvas.
• n8n: drop in an AI Agent node, connect a chat model and at least one tool sub-node (n8n requires at least one), and put your trigger in front of it. It’s a more technical canvas than Zapier, but you still don’t write code.

Which no-code builder should you start with?

Don’t spend a week comparing. Start where your work already lives:

• Already pay for Zapier? Use AI by Zapier. Your connected apps are already there.
• Already use Make, or like seeing every step on a canvas? Use Make’s AI Agent app.
• Want the most control and don’t mind a steeper first hour? Try n8n.
• Thinking of a custom GPT in ChatGPT instead? Check first. OpenAI plans to retire custom GPTs in favour of plugins, and new GPTs can’t be created on personal Free, Go, Plus or Pro accounts (OpenAI Help Center). A GPT also waits for you to chat with it; it doesn’t wake up when an email arrives.

Still torn between them? Our guide to picking a no code AI agent builder lays out each tool’s catch side by side and gives you a 30-minute test to run before you pay.

Test it like a new hire, not like a demo

Anthropic recommends extensive testing in a safe environment, with guardrails, before you trust an agent’s autonomy. In practice, for a desk agent, that looks like this:

1. Week one is labels and drafts only. Nothing leaves your account.
2. Check ten decisions a day. At the end of the day, open each AI label and ask: would I have filed this here? Write down every miss.
3. Turn each miss into an example. “Emails from our accountant about deadlines go in Today, not Admin.” One concrete example fixes more than a paragraph of rules.
4. Only widen the job when the misses stop. If it gets the labels right for a full week, then think about the next tool. Not before.

Five mistakes that break a first agent

• Three jobs in one brief. Triage, scheduling and CRM updates in one prompt means none of them works well. One agent, one job.
• Vague categories. “Important” means nothing to a model. “A client asks me for something with a deadline” does.
• Send buttons on day one. Drafts first. Always.
• A giant knowledge dump. Zapier warns that long instructions can hit token limits and that bigger reference material belongs in a knowledge source, not the prompt (Zapier Help: Build an agent). Keep the brief to one screen.
• Never looking at the run history. Every builder shows what the agent did and why. Five minutes a week there tells you more than any dashboard.

Where to go after your first agent

Once triage runs quietly, the same five parts carry over to bigger jobs. We’ve written these out step by step:

• Score and route new leads: AI agent for lead qualification in n8n and Zapier.
• Let an agent book appointments: AI agent for appointment booking automation.
• Let an agent answer the phone: pick an AI voice agent platform and build your first call agent.
• Answer common customer questions: automate customer support with AI agents.
• No agent needed, just a reliable follow-up: automate follow-up emails without code.
• Write the replies faster yourself: use ChatGPT to draft reply emails faster.
• An agent that works on your desktop instead of in the cloud: best AI agent for desktop workflows.
• Thinking about a self-hosted assistant you message from chat, or an agent that works inside your code: OpenClaw vs Claude Code.
• Weighing a chat-app assistant against Claude’s desk agent for non-coding work: OpenClaw vs Claude Cowork.
• Ready to build a real agent in n8n, with memory, Gmail and Calendar tools and an approval step: n8n AI agent workflow, step by step.

FAQ

Do I really not need to code?

For the build above, no. You connect accounts, pick actions from lists and write instructions in plain English. The skill that matters is writing a clear brief with real examples.

Can I build an AI agent for free?

It depends on the tool. Make’s new agent app is available on all plans through Make’s AI provider while it’s in open beta, so pricing may change. On Zapier, tool use needs the Advanced or Premium model tier, which is locked on the Free plan. Check the current plan page before you build.

Is it safe to let an agent read my work email?

Treat it like giving an intern access: only what the job needs, nothing that sends or deletes, and a look at the history every week. Check your company’s policy on connecting work accounts to outside tools before you start, and keep agents away from sensitive data and legal or financial decisions, as Make itself advises.

What’s the difference between an AI agent and a normal automation?

A normal automation follows the exact path you drew. An agent gets a goal and a set of tools and decides which tool to use for each item. If you can draw every path in advance, you don’t need an agent, and the automation will be cheaper and more predictable.

Build yours this week

Pick the one job that annoyed you most last week. Write the brief tonight, using three real examples. Tomorrow, build it with labels and drafts only, and check ten of its decisions before you go home. That’s your first agent.

Want a new practical AI workflow like this every week? Join Launchpad.

Sources

• Anthropic: Building effective agents
• Zapier Help Center: Add tools to your AI by Zapier step
• Zapier Help Center: Migrating from Agents to AI by Zapier
• Zapier Help Center: Build an agent in Zapier Agents
• Zapier blog: Zapier Agents is now AI by Zapier
• Zapier: Gmail integrations
• Make Help Center: Introduction to Make AI Agent (New)
• Make Help Center: Create your first AI agent
• n8n Docs: AI Agent node
• OpenAI Help Center: Creating and editing GPTs

Want to learn this properly? See our AI agents course guide with free options and a 4 week plan.

It’s Monday, 4:40 pm. Someone forwards you a spreadsheet with 2,000 rows and one line on top: “Can you tell me what’s going on here before tomorrow’s meeting?” You open it, scroll, and your eyes glaze over somewhere around column K. You can use ChatGPT to analyze a spreadsheet for work in about twenty minutes, but not by uploading it and typing “analyze this”. What works: clean the file in two minutes, make ChatGPT describe the data back to you first, ask the one question your meeting needs answered, make it show the numbers behind every insight, then turn it into a five-line summary for the person who asked.

Imagine it’s Tuesday morning. You walk in with three findings you have checked yourself, one simple chart, and the exact rows behind the number everyone will ask about. When your manager says “where does that 18% come from?”, you know.

Want more practical AI workflows like this one, free? Join Launchpad.

Short answer

Attach the spreadsheet (XLSX, XLS, CSV or TSV) to a ChatGPT chat. Before you ask for any insight, ask it to list the columns, the number of rows, the date range and anything that looks broken, such as blanks, duplicates or numbers stored as text. Then ask a specific business question, tell it which columns to use, and ask for a small table of the numbers behind each finding. Check three of those numbers yourself in Excel or Google Sheets. Only then ask for the summary, written for one named reader.

Why the check? For data work, ChatGPT often writes and runs Python code to do the calculations, and OpenAI tells you to review that code, the outputs and the assumptions before you rely on the result (OpenAI Help Center: Data analysis with ChatGPT). It can also sound confident when it’s wrong, which is why OpenAI calls it a first draft, not a final source (OpenAI Help Center: Does ChatGPT tell the truth?). A wrong total in a meeting costs you more than the twenty minutes you saved.

Before you upload: two minutes that save you twenty

OpenAI’s own advice for spreadsheets is simple: descriptive column headers in the first row, one row per record, plain-language column names. Avoid several unrelated tables on one sheet, empty rows or columns that split the data, and values that only exist inside images (OpenAI: Data analysis with ChatGPT). In practice that means:

• Delete the decoration. Title rows, merged header cells, a totals row at the bottom, a notes box in the corner. ChatGPT may read a totals row as one more customer.
• Rename cryptic headers. “Amt_EUR_ex” becomes “Amount EUR excl. VAT”. You’ll get better answers and you won’t have to explain it later.
• Remove what you don’t need. If the question is about regions and months, the column with customer phone numbers doesn’t have to come along.
• Check you’re allowed to upload it. On ChatGPT Business, Enterprise and Edu, OpenAI says it does not use workspace content to train its models by default. On a personal account, turn off “Improve the model for everyone” under Settings, Data controls (OpenAI: Data controls in ChatGPT). Your company’s rules still come first.

Working in Google Sheets? Download a copy (File, Download, Microsoft Excel or CSV) and upload that (Google: Export spreadsheets). A CSV only holds the sheet you had open, so pick the right tab first. If your account has connectors switched on, ChatGPT can also attach files straight from Google Drive, OneDrive or SharePoint.

Step 1: make ChatGPT describe the file back to you

This is the step most people skip. It’s also where you catch the problems that make every later answer wrong: a date column read as text, two “Region” spellings, 40 empty rows in the middle. I do it every single time, even with files I made myself.

I'm uploading a spreadsheet. Don't analyze it yet.
First tell me:
1. How many rows and columns it has, and which sheets.
2. Each column name, what kind of data is in it (date, number, text, category), and one example value.
3. The date range covered.
4. Anything that looks wrong or messy: empty cells, duplicates, numbers stored as text, odd spellings of the same value, outliers.
Then ask me up to 3 questions about anything you're unsure of.

Read the answer like a colleague’s handover. If it says 1,940 rows and you know there should be 2,000, stop and find out why before you go on. If it guesses what a column means, correct it in one line: “Status 3 means cancelled, leave those out.”

Step 2: ask the question your meeting needs, not “analyze this”

“Analyze this spreadsheet” gets you a tour of averages nobody asked for. A good question has a decision behind it. Give ChatGPT the context, the columns and the comparison you care about.

Context: this is our [monthly sales export / support ticket log / project budget]. I need to explain [what changed this quarter / why costs are over budget / where tickets pile up] to [my manager] tomorrow.

Use these columns: [Date], [Region], [Product], [Amount EUR excl. VAT].
Compare [Q3 with Q2].

Give me the 3 findings that matter most for that question. For each finding:
- one sentence in plain English
- the exact numbers behind it, in a small table
- how you calculated it (filters, groupings, which rows you left out)
If something can't be answered from this data, say so instead of guessing.

That last line matters. Without it, a chatbot tends to fill gaps with something that sounds reasonable. With it, you get “the file has no cost column, so I can’t calculate margin”, which is far more useful.

Step 3: make it show its work, then check three numbers

When ChatGPT runs Python for an analysis, you can open the code and the outputs. You don’t have to read Python to use this. Ask for the method in words instead:

For finding 1, explain step by step in plain words how you got the number: which rows you filtered, which column you summed or averaged, and how many rows went into it. Then give me one filter I can set in Excel to see the same rows myself.

Now take two minutes in the spreadsheet itself. Set that filter, look at the sum in the status bar at the bottom, and compare. Do this for three numbers: the biggest one, the most surprising one and the one your manager will ask about. If they match, you can trust the rest a lot more. If one doesn’t, ask ChatGPT why, and you’ll often find the totals row or the duplicate it warned you about in Step 1.

How to use ChatGPT to review a spreadsheet and summarize insights

Once the numbers hold up, the summary is the easy part. The trick is to write it for one person, with one length, and one thing you need from them.

Using only the findings we checked above, write a summary for [my manager, who has 2 minutes before the meeting].
- 5 lines maximum
- start with the single most important change, with its number
- then 2 supporting findings, each with its number
- then 1 open question or risk the data can't answer
- end with the decision or next step we need from them
Plain English, no jargon, no numbers that aren't in our checked tables.

Swap the reader and you get a different, equally useful summary: “my team, so they know what to focus on next week” or “a client, without internal cost figures”. Same data, same checked numbers, a different message.

Step 4: one chart, if it helps

ChatGPT can turn uploaded data into charts. Bar, line, pie and scatter charts can be switched to an interactive version, other chart types come back as an image (OpenAI: Data analysis with ChatGPT). Ask for exactly one chart that shows your main finding:

Make one bar chart of [Amount] by [Region], [Q2] next to [Q3]. Sort from largest to smallest change. Title it with the finding in plain words, not the column names. Label the bars with the values.

If the first chart isn’t what you meant, say so directly. OpenAI notes it may pick a chart type or method that doesn’t match your intent on the first try, and naming the chart type, grouping or column fixes most of that.

When the spreadsheet is big, messy or full of formulas

• Very large files. Spreadsheets and CSVs have a size limit of roughly 50 MB, depending on how wide each row is (OpenAI: Uploading files to ChatGPT). Above that, filter to the period or region you need and upload that slice.
• Lots of formulas. OpenAI says the number and complexity of formulas can affect processing, even when the file is small enough. Paste the sheet as values into a new file (Paste special, Values) and upload that copy.
• Several tabs. Tell ChatGPT which sheet to use, or ask it to inspect one sheet at a time. If an answer seems incomplete, that is OpenAI’s own suggestion too.
• A table inside a PDF or screenshot. ChatGPT may not read exact values reliably from scans or images of tables. Get the original spreadsheet if you can. If you only have the PDF, pull the table out first, for example with Claude: extract data from a PDF into a spreadsheet.
• A recurring report. If you do this every month, save your Step 1 and Step 2 prompts in a ChatGPT Project with last month’s file, so next month starts from the same questions.

If what you really need is a formula you can keep in the sheet, not a one-off answer, see how to use AI to write Excel formulas. It shows the prompt to use and how to check the result before you trust it.

Common mistakes

• Skipping the describe-back step. Every answer after that is built on a guess about your columns.
• Asking for “insights”. You get generic observations. Ask for the answer to a question with a decision behind it.
• Trusting a number you didn’t check. Three spot-checks take two minutes. Do them.
• Uploading the whole CRM export. Personal data you don’t need for the question shouldn’t leave your systems.
• A summary for nobody. “Summarize this” gives you a shorter spreadsheet in words. “Brief my manager on what changed” gives you something you can send.

FAQ

Can I do this on the free ChatGPT plan?

Yes. Spreadsheet uploads are available on Free and paid plans, with plan-specific limits. Free users can upload 3 files per day, so one well-prepared file beats three attempts (OpenAI: Uploading files to ChatGPT).

Can ChatGPT pull live data from the web into my analysis?

Not inside the analysis itself. OpenAI says the Python environment used for data analysis can’t make web requests or API calls, so upload or connect any extra data you want included.

I have a sales pipeline export. Is there a ready-made prompt?

Yes, we wrote a full weekly review kit for exactly that: ChatGPT prompts for sales pipeline management.

Related guides

• Survey results instead of a sales sheet: how to use AI to analyze survey data, including open text answers.
• Documents instead of data: how to get ChatGPT to summarize long documents for work.
• Turning your findings into a written report: how to get ChatGPT to help write a report for work.
• Weekly updates from your numbers: Gemini for automated status updates.
• After the meeting: use ChatGPT to summarize meeting notes.
• Not sure ChatGPT is the right tool for your setup: ChatGPT vs Gemini vs Copilot, and when Gemini or Copilot is the better pick for spreadsheets.
• When the job turns into a script instead of a one-off analysis: Claude vs Gemini for coding, for automating a sheet or merging a folder of CSV exports.
• Presenting what you found in the data: how to use AI to make a presentation from your findings.

Try it on the spreadsheet you’ve been avoiding

Pick the file that’s been sitting in your downloads folder since last week. Run only the Step 1 prompt and read what comes back. Often that alone tells you more than ten minutes of scrolling: how much data there really is, what’s broken, and which question is worth asking.

Want a new practical AI workflow like this every week? Join Launchpad.

Sources

• OpenAI Help Center: Data analysis with ChatGPT
• OpenAI Help Center: Uploading files and audio to ChatGPT
• OpenAI Help Center: Data controls in ChatGPT
• OpenAI Help Center: Does ChatGPT tell the truth?
• Google Docs Editors Help: Export spreadsheets

Working with documents as well as spreadsheets? Read our guide to AI document analysis tools.

Your calendar is full, your inbox is louder than your to-do list, and the blank page for a simple update still feels heavy. An AI writing assistant for busy professionals is not a magic ghostwriter. It is a desk habit: one tool you brief once, then reuse for emails, short docs, and status notes so you stop starting from zero every time.

Imagine you open ChatGPT, Claude, Gemini, or Copilot on Monday with a short “how I write at work” note already saved. By lunch you have a reply, a one-pager outline, and a cleaner status update, and none of them sound like a robot wrote them for a stranger.

Want more practical AI workflows like this one, free? Join Launchpad.

Short answer

Pick one AI writing assistant you already have access to at work (or a personal plan your company allows). Save a short voice brief: your role, who you write for, words you never use, and a sample email that sounds like you. For every writing job, give it context, audience, length, and the outcome you need, then edit the draft yourself before you send. Use the same assistant for a week so your prompts get sharper. Switch tools only when a task truly fits another product better, for example drafting inside Google Docs with Gemini, or rewriting inside Word with Microsoft 365 Copilot.

Google documents how Gemini can help you write and rewrite inside Docs when your Workspace admin has turned it on (Google Docs Help: Write and edit with Gemini in Docs). Microsoft explains Copilot as an AI assistant that works across Microsoft 365 apps you already use at work (Microsoft Learn: What is Microsoft Copilot?). The method below works with either of those, and with ChatGPT or Claude in a browser.

What an AI writing assistant is (and is not) at a busy desk

At work, “AI writing assistant” usually means a chat or side panel that turns rough notes into clearer prose. It can close the gap between “I know what I mean” and “I have a draft I can send.” It does not replace your judgment on facts, tone with a client, or what your manager actually asked for.

Busy professionals get stuck in three places:

• Starting cold on a reply that should take five minutes.
• Rewriting the same update every week with slightly different numbers.
• Switching tools mid-task and losing the thread of how you like to sound.

A good setup fixes those. It will not invent your strategy, your pricing, or your missing data. If those are unclear, the assistant will guess, and guessing is how weak emails leave your desk.

Choose the assistant that matches where you already write

You do not need five tools. You need one default and one in-document helper.

1. Default chat tool (ChatGPT, Claude, or similar): best when the writing starts outside a document. Paste notes, ask for a structure, then move the draft into email or Docs.
2. In-document helper (Gemini in Google Docs, Copilot in Word or Outlook): best when the file already lives in Workspace or Microsoft 365 and you want rewrite, shorten, or tone changes without leaving the page.

If your company only allows one, use that one. Policy beats preference. If you can choose, pick the product that sits next to the app you open most. A beautiful prompt in the wrong window still costs you a copy-paste tax.

Still choosing between the three big names? Our ChatGPT vs Gemini vs Copilot comparison matches each one to the apps your work already lives in, with real monthly costs side by side.

For a deeper single-task playbook once your assistant is set up, see how to get ChatGPT to help write a report for work. This page stays one level higher: the assistant itself as a daily desk tool.

Torn between ChatGPT and Claude as your default writing tool? Our ChatGPT vs Claude for writing comparison matches each one to common work writing jobs, with a short test you can run on your own drafts this week.

Build a one-page voice brief (do this once)

Most weak AI drafts fail because the tool does not know you. Spend ten minutes on a brief you can paste at the top of every writing chat.

You are my AI writing assistant at work.

About me:
- Role: [e.g. account manager / ops lead / consultant]
- Company type: [e.g. B2B SaaS / agency / internal team]
- I usually write to: [clients / managers / cross-functional peers]
- Goal of most writing: clear next step, not clever wording

My voice:
- Prefer short sentences and plain words
- Never use: [list 5-10 words or phrases you hate]
- Soften bad news, but do not hide it
- Sign-off style: [e.g. Thanks, FirstName]

Sample of my real writing (keep this tone):
"""
[paste one short email you already sent and like]
"""

Rules:
- Do not invent facts, dates, prices, or names
- If something is missing, ask me up to 3 questions before drafting
- Return drafts I can edit, not final “perfect” copy

Save that block in a note app or a pinned ChatGPT/Claude project. Reuse it. Update it when your role or audience changes. The brief is the difference between “generic helpful AI” and “assistant that sounds like you on a good day.”

The four-part ask that beats “write this for me”

Every request should include context, audience, length, and outcome. Soft asks invite soft drafts.

Using my voice brief above, help with this writing task.

Context: [what happened / what the reader already knows]
Audience: [who reads this, and what they care about]
Length: [e.g. 120 words / 6 bullets / one page]
Outcome: [what I want them to do or understand]

Raw notes:
"""
[paste messy bullets, meeting scraps, or a rough draft]
"""

Return:
1) A draft in my voice
2) A shorter alternative
3) Three lines I should double-check for accuracy before sending

That last line matters. It trains you to verify, and it trains the model to surface risky spots instead of burying them in fluent prose.

A simple weekly writing stack for busy professionals

Use the same assistant for a light weekly rhythm. Consistency beats novelty.

1. Monday: Clear two sticky replies. Paste the thread and your voice brief. Ask for a draft reply plus a one-sentence version for Slack or chat.
2. Midweek: Turn project notes into a status update. Ask for bottom line first, then three bullets of progress, risks, and asks.
3. Before a client call: Ask for a short agenda email and three questions you should confirm in the call.
4. Friday: Turn the week’s scraps into a short wrap for your manager. Keep it under 150 words.

If proposals are a big part of your work, pair this stack with a focused guide like ChatGPT for proposal writing for freelancers. If client tone is the hard part, use ChatGPT for client communication templates that sound like you. Those pages go deep on one deliverable. Your assistant setup is the shared foundation.

Worked example: messy notes to a sendable update

Suppose your notes look like this:

- vendor still late on API keys
- design signed off Tuesday
- launch moved from 22 Oct to 29 Oct (waiting on keys)
- need legal to confirm DPA by Friday
- customer success wants a 15-min walkthrough next week

A useful assistant output is not a novel. It is something like:

• Bottom line: Launch moves to 29 Oct because vendor API keys are still outstanding; design is signed off.
• Ask: Legal to confirm the DPA by Friday; Customer Success to book a 15-minute walkthrough next week.
• Risk: If keys slip again, the new date moves with them.

You still check the dates and names. The assistant saved the blank-page time. You kept the truth check.

Draft an email follow-up after a meeting (same assistant)

Meeting follow-ups are where busy people lose hours. Use the same voice brief and a tight prompt:

Draft a follow-up email after today’s meeting.

Attendees: [names and roles]
My goal for this email: [confirm decisions / assign owners / book next step]
Tone: clear and warm, no fluff

Notes:
"""
[paste decisions, open questions, and owners]
"""

Format:
- Subject line options (3)
- Email body under 180 words
- Bullet the action items with owners and dates if stated
- End with one clear next step
Do not invent owners or dates. Mark gaps as “TBD”.

Send only after you match every owner and date back to what was actually said. If you need faster day-to-day replies beyond follow-ups, the desk guide on drafting reply emails faster with ChatGPT goes deeper on that loop.

Use the in-document helper without losing your voice

When you write inside Google Docs, Gemini can help you rewrite selections, change tone, or expand a thin section. Microsoft 365 Copilot can do similar work inside Word and related apps when your license includes it. Treat both like a second editor, not like the author of record.

Practical rules that keep quality high:

• Highlight one section at a time. Whole-document “make this better” often flattens your point.
• Ask for two options: shorter, and clearer. Pick pieces from each.
• Never accept a rewritten number or quote without checking the source.
• Keep your voice brief nearby so chat tools and in-doc tools stay aligned.

Already wrote the email and it just reads too blunt or too casual? Our guide on making an email sound more professional has copy-paste prompts for ChatGPT, Claude, Gmail and Outlook that fix the tone without changing what you meant.

Privacy and workplace rules

Writing assistants often see customer names, contracts, and internal plans. Before you paste:

• Follow your company’s AI and data policy first.
• Prefer the work account and retention settings your admin chose.
• Redact tokens, passwords, and private links. The draft rarely needs the secret itself.
• If a document is too sensitive for any cloud assistant, outline offline and only ask for help on structure or wording of non-sensitive parts.

This is not fear. It is the same care you already use with shared drives and forwardable email.

Common failure modes (and the fix)

• Sounds nothing like you. Fix: paste a real sample email into the voice brief and ban your least favorite phrases.
• Invented details. Fix: keep “do not invent facts” and require a “lines to double-check” section.
• Too long for the channel. Fix: set a hard word limit in the ask.
• Tool hopping every day. Fix: one default for a full week before you judge it.
• Sending the first draft. Fix: a two-minute human pass for names, dates, and tone with that specific reader.

The same voice brief and four-part ask also work for longer one-off documents. If you are hiring, here is how to use AI to write a job description that still sounds like your team.

Review season works the same way: here is how to use AI to write a performance review that stays specific and fair.

The same voice brief helps with outside announcements too: here is how to use AI to write a press release that reads like news, not hype.

Writing to the same readers every week or month? The voice brief carries over there as well: here is how to use AI to write a newsletter that still sounds like you.

Selling products online? The voice brief works for listings too: here is how to use AI to write product descriptions that stay accurate and specific.

The same voice brief works for LinkedIn. If you want to post about your work without it reading like a template, see how to use AI to write LinkedIn posts that sound like you.

The voice brief also works for longer pieces. For a full article, see how to use AI to write a blog post step by step, from outline to final edit.

What to do next

Today: write the voice brief and save it. Tomorrow: run one sticky email and one status update through the four-part ask. By Friday, you will know whether your default assistant is earning its place on the desk.

When you want more desk workflows in the same practical tone, join Launchpad below. Same idea throughout: AI that helps you finish the writing already on your plate.

One of the most useful places to try this is your own LinkedIn page. Our guide on how to use AI to write a LinkedIn profile uses the same fact-sheet idea for your headline and About section.

Your Slack sidebar looks calm. Underneath, three threads are waiting for you: a client question buried under emoji, a decision that never got a clear owner, and a “quick ping” that is already two days old. The best ChatGPT prompt for action items from a Slack thread is not “summarize this.” It is a prompt that asks for owners, deadlines, open questions, and what is already done, so you leave the thread with a short list you can act on.

Imagine you open Monday with six threads still unread. Fifteen minutes later you have a table: who owes what, by when, and which two replies you should send first. Nothing important is still hiding in a side comment.

Want more practical AI workflows like this one, free? Join Launchpad.

Short answer

Copy the full Slack thread (or connect Slack inside ChatGPT if your workspace allows it). Paste it into ChatGPT with a prompt that forces a structured action list: task, owner, deadline or “none stated”, status (open / done / blocked), and the exact message that proves it. Ask ChatGPT to skip resolved jokes and FYIs, and to flag anything that sounds like a decision but has no owner. Then check two of those proof lines back in Slack before you send follow-ups.

OpenAI documents that ChatGPT can search messages, threads, and channels you already have permission to access when the Slack app is connected, and that available actions depend on your plan, region, and admin settings (OpenAI Help Center: Using Slack in ChatGPT). Whether you paste the thread by hand or use the connector, the same prompt shape works.

Why “summarize this Slack thread” is not enough

A summary tells you what people talked about. At work you usually need something narrower: what still needs a human, who that human is, and whether the clock is already running. Threads mix jokes, reactions, side questions, and one real ask. A free-form summary often keeps the chatty parts and softens the one line that creates work for you.

Action items hide in three places that summaries miss:

• A reply that says “can you send that by Friday?” without @mentioning anyone.
• A reaction or “+1” that looks like agreement, but never names who will do the work.
• A later reply that already closed the ask, so the early request is no longer live.

OpenAI’s own Slack triage guidance is aimed at a ranked queue of next steps, not a paragraph recap: find direct asks, check whether they are still live, and draft replies only after that pass (ChatGPT use case: Prioritize Slack action items). That is the same job this desk method is built for.

Best ChatGPT prompt for action items from a Slack thread

Use this as your default. Replace the bracketed bits. Keep the output format. Soft prompts like “pull out anything useful” invite fluff. Hard columns keep ChatGPT honest.

You are helping me turn one Slack thread into a work list I can trust.

Context:
- My name: [Your name]
- My role: [e.g. project lead / account manager]
- Channel or DM: [name]
- Thread topic in one line: [topic]
- Time window that matters: [e.g. last 48 hours / whole thread]

Here is the full thread, oldest first:
"""
[paste the full thread, including display names and timestamps if you have them]
"""

Do this:
1) List every concrete action item still open for anyone in the thread.
2) Skip jokes, emoji-only replies, pure FYIs, and asks that a later reply already closed.
3) Before you mark something open, check later messages for a done signal (“sent”, “shipped”, “closed”, “done”, a link that answers the ask).

Return a markdown table with columns:
| # | Action | Owner (name or “unassigned”) | Deadline (date or “none stated”) | Status (open / blocked / done) | Proof (quote the exact line) | Suggested next step for me |

After the table, add three short sections:
A) Decisions made (only if clearly agreed)
B) Open questions with no owner
C) Risks if we do nothing this week

Rules:
- Do not invent owners, dates, or tools that are not in the thread.
- If the owner is unclear, write “unassigned” and say why in Suggested next step.
- Prefer my action items first, then others.
- Keep quotes short and exact.

That prompt is the “best” version for most desk days because it does three things a casual ask does not: it forces proof quotes, it separates open from done, and it refuses invented owners. Those three habits matter more than clever wording.

How to paste a Slack thread cleanly

Bad paste in, weak list out. Spend one minute on the copy step.

1. Open the thread in Slack, not only the channel preview.
2. Expand “older replies” so you are not missing the close of the ask.
3. Copy from the first message through the latest reply. Include names. Timestamps help when people say “tomorrow”.
4. If the thread is huge, paste the last 30-40 messages first, then add: “Earlier context: [2-3 lines on how this started].”
5. Strip huge pasted code blocks or log dumps unless the action depends on them. Point ChatGPT at the part that creates work.

If your company has connected Slack inside ChatGPT, you can ask ChatGPT to read the thread in place instead of pasting. Still give it the same table format and the same “do not invent owners” rule. Connected access does not replace a clear output shape.

A shorter prompt when you only need your own next steps

Sometimes you are not facilitating the whole channel. You only need what you owe. Use this shorter version after a long standup thread or a client channel that moved overnight.

From this Slack thread, list only action items for me ([Your name]).

For each item give:
- What I need to do (one sentence)
- Who is waiting on it
- Deadline if stated, else “none stated”
- Exact quote that creates the ask
- One draft reply I could send (max 3 sentences, plain language)

Ignore items already answered later in the thread.
Do not invent work that was only discussed as an idea.

Keep this prompt for morning triage. Use the full table prompt when you are about to update a project board, write minutes, or hand the thread to someone else.

How to turn the table into real work in ten minutes

ChatGPT’s list is a draft. Your job is the last mile.

1. Spot-check two proof quotes in Slack. If either quote is wrong or out of context, ask ChatGPT to rebuild the table from a cleaner paste.
2. Assign the “unassigned” rows. Write one short Slack reply that names the owner. Unowned work is how threads die quietly.
3. Move open items into the place your team already trusts: Asana, Linear, a Google Doc, or a simple checklist in the channel. Do not leave them only inside ChatGPT history.
4. Send at most two follow-ups from the drafts. Prefer one clear ask over a long recap of the whole thread.
5. Mark done items in Slack with a short “Done: …” reply when that helps the next reader. Future-you will thank present-you.

If those items land on a day that is already full, the brain dump in how to use AI to plan your day helps you fit them around your meetings.

Worked example (short thread)

Suppose a thread looks like this (names abbreviated):

Alex: Client wants the revised proposal by Thursday EOD.
Sam: I can update pricing if someone confirms the discount.
You: I’ll confirm discount with finance today.
Alex: Also, can we add the case study link?
Jordan: Case study link is already in the footer of v2.
Sam: Pricing tab updated - waiting on discount confirmation.

A good ChatGPT run should land close to this:

• You: confirm discount with finance today (open).
• Sam: finish pricing once discount is confirmed (blocked on you).
• Alex’s case study ask: done (Jordan already answered).
• Overall delivery: revised proposal Thursday EOD, owner still partly unassigned until you name who uploads the final file.

Notice what a vague “summary” would miss: the case study is already closed, Sam is blocked on you, and “Thursday EOD” is a deadline without a single final owner for the upload. That last gap is exactly why the prompt asks for “unassigned” instead of guessing.

How to use ChatGPT when Slack is connected (optional)

If your workspace admin has enabled the ChatGPT Slack app, you can point ChatGPT at a channel or thread without manual copy-paste. OpenAI’s help article covers search across messages you can already access, and notes that some actions (reminders, uploads, profile updates) need extra admin approval (Using Slack in ChatGPT).

A safe connected prompt looks like this:

In Slack, open the thread in #[channel] that starts with “[first line or topic]” from the last [N] days.

Build the same action-item table as usual:
Action | Owner | Deadline | Status | Proof quote | Suggested next step for me ([name]).

Read the latest replies before calling anything open.
Do not post to Slack. Do not invent owners.
Draft replies only after I approve the table.

Two rules keep this safe at work: never ask ChatGPT to post for you until you have read the table, and never paste confidential threads into a personal ChatGPT account if your company forbids it. Use the account and data controls your workplace already approved.

Common failure modes (and the fix)

• Invented owners. Fix: keep the “Do not invent owners” line. If ChatGPT still guesses, add “If unsure, write unassigned.”
• Old asks treated as live. Fix: paste the full thread tail, or say “A later reply can close an earlier ask.”
• Every message becomes a task. Fix: tell it to skip FYIs and ideas with no ask.
• No deadlines, so nothing feels urgent. Fix: ask for “none stated” explicitly, then you add real dates in your tracker.
• Proof quotes are paraphrased. Fix: require “quote the exact line” and reject the run if quotes look rewritten.

Privacy and what not to paste

Slack threads often hold customer names, salaries, access links, or legal chatter. Before you paste into ChatGPT:

• Follow your company’s AI and data policy first.
• Prefer a work ChatGPT workspace with the retention settings your admin chose.
• Redact tokens, passwords, and private URLs. The action item rarely needs the secret itself.
• If the thread is too sensitive to paste, extract only the lines that create tasks, or use an approved connected app under IT’s rules.

OpenAI’s Slack help page also reminds you that Slack is a third-party service and that data may be handled under Slack’s own terms when the app is enabled. Read your org’s Slack and ChatGPT policies before you connect anything new.

A simple weekly habit

Once a week, pick your three noisiest channels. For each, run the full action-item prompt on any thread that still has an unread badge or an @you mention. Move open rows into your tracker the same day. The goal is not a perfect archive. The goal is fewer surprises on Friday afternoon.

If you lead a team, share the table prompt in your channel topic or a pinned message. When everyone uses the same columns, handoffs get shorter, and “I thought you owned that” happens less often.

What to do next

Take one real Slack thread waiting on you right now. Paste it under the full prompt above. Check two proof quotes. Assign every “unassigned” row with one short reply. That single loop is the whole skill.

When you want more desk workflows like this, join Launchpad below. Same tone, same practical focus: AI that helps you finish the work already on your plate.

It’s Tuesday afternoon. Someone has just dropped a 70-page document in your lap: a new policy, a tender, a vendor proposal, a strategy deck turned into a PDF. You’re expected to “have a look before Thursday”, next to everything else on your plate. ChatGPT can summarize long documents for work in about fifteen minutes, but not with one “summarize this” prompt. What works: upload the file, ask for a map of the document first, summarize it section by section, then write the final summary for the one person who will read it, and check every claim against the page it came from.

Imagine it’s Thursday morning. You walk into the meeting having read five lines you trust, with page numbers next to the two points that actually affect your team. When someone asks “what does it say about the deadline?”, you open page 41 and read it out.

Want more practical AI workflows like this one, free? Join Launchpad.

Short answer

Attach the document to a ChatGPT chat (DOCX, PDF or TXT). Before you ask for any summary, ask ChatGPT to list the document’s sections with page numbers, so you know it has seen the whole file. Then summarize it one section at a time, with a page reference behind every point. Only after that, ask for the final summary, written for a specific reader: your manager, your team, a client or yourself. Finish with a check where ChatGPT quotes the sentence behind each point, and read three of those quotes yourself in the original.

Why the extra steps? ChatGPT can sound confident even when it’s wrong, and OpenAI itself tells you to treat it as a first draft and to verify quotes and data (OpenAI Help Center: Does ChatGPT tell the truth?). A summary of a long document is exactly where that bites, because you’re asking it for the one thing you can’t easily check: what it left out.

Why one big “summarize this” falls short on long documents

File size is rarely the problem. ChatGPT accepts documents up to 512 MB, and text files are capped at 2 million tokens per file, which is far more than a normal work document (OpenAI Help Center: Uploading files to ChatGPT). The problem is attention. A 2023 study led by Stanford researchers found that language models use information at the start and end of a long input best, and do noticeably worse with what sits in the middle (Liu et al., “Lost in the Middle”, 2023).

In practice that means a one-shot summary of a 70-page file often sounds complete and still skips the clause on page 38 that matters to you. Section-by-section summaries keep each piece small enough to get proper attention, and the map tells you if anything was skipped.

Before you upload: three quick checks

• Is it a file ChatGPT can read? DOCX, PDF and TXT work. A Google Doc can’t be uploaded directly: export it as PDF or DOCX first (File, Download). Slides work as PPTX.
• Is the text real text? Try to select a sentence in the PDF. If you can’t, it’s a scan, and the summary will be thin or wrong. On most plans ChatGPT reads the digital text in a document, not the images inside it, so charts and scanned pages can be missed (only ChatGPT Enterprise reads images in PDFs, according to the same OpenAI file-upload page). Run the scan through your PDF tool’s text recognition first, or paste the key table in as text.
• Are you allowed to upload it? On ChatGPT Business and Enterprise, OpenAI says your workspace data isn’t used to train its models by default. On a personal Free, Plus or Pro account it is, unless you switch off “Improve the model for everyone” under Settings, Data Controls (OpenAI Help Center: data controls). Either way, follow your company’s rules for confidential files. If in doubt, ask before you upload.

One more practical note: on the Free plan you can upload 3 files a day (same OpenAI file-upload page), so start with the document that matters most.

Step 1: ask for a map before a summary

This takes one minute and saves you from the most common failure: a summary of the first half of the file.

I've attached a document I need to understand for work. Don't summarize it yet.

1. Tell me what kind of document it is, who wrote it and the date, if it says.
2. List every section and subsection in order, with the page numbers.
3. Tell me the last sentence of the document, so I know you've seen the end.
4. Flag any pages that look like images, scans or tables you couldn't read.

Check the last sentence and the page count against the file. If they don’t match, the file didn’t load fully. Re-upload it, or split it (see below).

Step 2: summarize it section by section

Now work through the map in chunks of a few sections at a time. Ask for page references every time; they’re what make the summary checkable later.

Summarize sections 1 to 3 only.

For each section give me:
- 2 to 4 bullet points with the main points, each with its page number
- any numbers, dates, deadlines or names, exactly as written
- anything that is a requirement, an obligation or a change from how things work now

If a section says nothing that matters for someone in [your role], say so in one line. Don't fill gaps with general knowledge: if it's not in the document, leave it out.

Type “next” and it continues with the following sections. For a 70-page document that’s usually four or five rounds. Skim each round as it comes in. This is where you catch “wait, that sounds wrong” while the section is still fresh.

Step 3: write the summary for the person who will read it

A good work summary depends on who reads it and what they have to do next. Same document, very different summaries. Pick the one you need and paste it in the same chat, so ChatGPT builds on the section notes rather than starting over.

For your manager (a brief they read in 30 seconds):

Using only the section notes above, write a brief for my manager.
- Line 1: what this document is and why it landed on our desk.
- Then the 3 points that matter most for our team, with page numbers.
- Then what we need to decide or do, and by when.
- Then anything that's unclear or that I should check with someone.
Max 150 words. Plain language, no jargon from the document unless it's needed.

For your team (actions, not a book report):

From the section notes, list everything in this document that changes what my team does day to day. Use a table: what changes, from when, who it affects, page number. Then one line: what stays the same.

For a client or someone outside the company: ask for a short plain-language version with no internal names, and read it twice before it leaves your outbox. A summary sent outside is your words, not ChatGPT’s.

For yourself (to really understand it): ask ChatGPT to explain the three hardest sections in simple terms and then ask you three questions about them. If you can’t answer, reread those pages.

Step 4: check the summary against the document

This is the step that turns a nice-sounding summary into one you can defend in a meeting.

Go through the summary point by point. For each point, quote the exact sentence from the document that supports it, with the page number. If a point is not directly supported by a sentence in the document, mark it NOT FOUND.

Then tell me: what's in the document that a reader in my role would care about, but isn't in the summary?

Open the original and read three of those quotes yourself, at least one from the middle of the document. If a quote isn’t there, or says something slightly different, fix that point and ask ChatGPT to recheck the rest. Any number in the summary gets checked against the page, every time.

Summarize a long email thread before a meeting

A 40-message thread is a long document too, just a messy one. The same method works with two changes. First, get the thread into a file: most mail apps can print or save a conversation as a PDF. Second, ask for a timeline before a summary:

This is an email thread I need to be ready for in a meeting tomorrow.
1. Give me a timeline: date, who wrote, what they said or asked, in one line each.
2. What has been decided, and by whom?
3. What is still open, and who is waiting on whom?
4. What will I probably be asked in the meeting, and what does the thread say about it?

Point 4 is the one that saves you. Check the decisions against the actual emails, because threads are full of “I think we agreed” that nobody agreed to. If what you have is notes from the meeting itself rather than an email thread, our guide on how to summarize meeting notes with ChatGPT is the better fit.

When the document is too long, or there are several

• The map doesn’t reach the end. Split the PDF into parts (most PDF tools can do this), upload part 1, do steps 1 and 2, then part 2 in the same chat. Ask for the final summary only when all parts are done.
• Several related documents. Put them in a ChatGPT Project so you can return to them all week. OpenAI’s limit is 5 files per project on Free, 25 on Go and Plus, and 40 on Business and Enterprise. Map each file separately first, then ask for one summary across them.
• Two versions of the same document. Don’t ask for two summaries and compare them yourself. Upload both and ask: “What changed between version A and version B? List every difference with the page numbers in both.”

If the long document is a contract, lease or policy, the summary steps change a little. See how to use AI for legal documents for prompts that pull out key clauses and prepare your questions.

Common mistakes

• One prompt, whole document. You get a smooth summary of the parts the model paid attention to.
• No page numbers. Without them you can’t check anything, so you end up trusting it blindly or rereading everything.
• Uploading a scan. If you can’t select the text, ChatGPT probably can’t read it properly either.
• A summary for nobody. “Summarize this” gives you a shorter document. “Brief my manager on what we need to decide” gives you something useful.
• Letting it fill gaps. Tell it to leave out anything that isn’t in the document. General knowledge sneaks in otherwise.
• Skipping your own read of the key pages. The summary tells you where to look. It doesn’t replace looking.

Related guides

• Prefer Claude for big PDFs: Claude AI analyze long PDF files.
• A long report into a brief with owners and deadlines: summarize a long report and extract action items with Claude.
• Contracts and legal documents: Claude AI key points from a PDF.
• Reviewing contracts every week: how to pick AI contract review software.
• Writing your own report afterwards: how to get ChatGPT to help write a report for work.
• Replying once you’ve read it: use ChatGPT to draft reply emails faster.
• The document is in another language: how to use AI to translate a document and check it before anyone relies on it.

Try it on the document on your desk

Pick the longest thing you’re supposed to read this week. Run the map prompt first and look at what comes back. Often the map alone pays off: you see in one minute which three sections you actually need to read, and which forty pages you can leave to the summary.

Want a new practical AI workflow like this every week? Join Launchpad.