Author: Pete

You need feedback by Friday. Maybe it’s the team after a messy tool rollout, maybe it’s clients after a project. You open a blank form, type “Survey” as the title, and stare at it. Writing ten good questions feels like it should take ten minutes. It never does.

Here’s the short answer. To use AI to create a survey, don’t ask for “a customer feedback survey”. Tell it what decision the answers will feed, who will fill it in, and how many minutes they will give you. Ask for one idea per question, neutral wording and answer options that don’t overlap. Then let the AI check its own draft for leading and double-barreled questions, build it in Google Forms or Microsoft Forms, and test it on two people before you send it. ChatGPT, Claude, Gemini and Copilot can all do the drafting in about a minute. The thinking about what you actually need to know stays with you.

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

Why most AI-made surveys get weak answers

Ask any chatbot for “an employee satisfaction survey” and you will usually get a long list of agree/disagree statements that could belong to any company on earth. It looks finished. The answers you get back are hard to act on.

Survey researchers have known the reasons for decades. The Pew Research Center’s guide to writing survey questions lists the classic traps:

• Double-barreled questions. One question that asks about two things, like “How happy are you with our speed and quality?” Nobody knows which half the answer is about.
• Agree/disagree statements. Pew notes that some people tend to agree with whatever statement you put in front of them (acquiescence bias). Offering a choice between two alternatives gives cleaner answers.
• Answer options that overlap or leave people out. Closed questions should cover all reasonable answers, and the options shouldn’t overlap.
• Jargon and double negatives. Simple, concrete words get understood the same way by everyone.
• Question order. Earlier questions can colour how people answer later ones.

A chatbot knows all of this. It just doesn’t apply it unless you ask. So the trick is not a better tool. It is a better brief, and a second pass where the AI checks itself.

Step 1: Write a 60-second brief

Before you open any AI tool, answer four questions in plain words. This is the part only you can do.

1. What decision will these answers feed? “Decide whether we keep the new shared inbox tool or go back to email” is a decision. “Get feedback” is not.
2. Who fills it in? 25 support agents, 40 clients, your newsletter readers. Their context changes the words you use.
3. How much time will they give you? Be honest. A busy client gives you three minutes, not fifteen.
4. What will you do with each answer? If you can’t say what you’d do differently based on a question, that question goes.

That last one is the filter that keeps a survey short. Most bloated surveys are full of “nice to know” questions nobody ever looks at again.

Step 2: Generate the survey with one prompt

Paste your brief into ChatGPT, Claude or Gemini with this prompt. It works the same in all three.

The last line matters more than it looks. When the AI has to say which part of the decision each question serves, the filler questions show themselves. If it writes “general sentiment” next to a question, cut it.

Step 3: Let the AI check its own draft

First drafts still slip. Open a new message in the same chat and run a review pass:

Asking for “the one question I could drop” is a small habit that pays off. It forces a ranking, and you’ll often agree with the cut.

Step 4: Build it in Google Forms or Microsoft Forms

You now have a clean list of questions. Getting them into a real form is the boring part, and both big form tools have AI built in for it.

Google Forms with Gemini

Open a new form and use Help me create a form. Paste your final survey as the prompt and Gemini builds a draft form you can edit before you click Create form. According to Google’s help page for this feature, it needs an eligible Google Workspace or Google AI plan, works on desktop, and can’t yet build forms with multiple sections or edit a form you already have (it replaces the content). So for a longer survey with sections, generate the questions here and add the sections by hand.

Microsoft Forms with Copilot

In Microsoft Forms, start a new form and describe what you want in the Copilot box, or paste your finished questions. Microsoft’s Copilot in Forms guide says Copilot drafts a title, description, questions and answer options, can suggest fixes for leading or double-barreled questions, and needs a Microsoft 365 account and a Copilot license. You pick Keep to move the draft onto the form.

No AI plan in your form tool?

Then the free version of any chatbot plus a normal form builder is plenty. Copy the questions from Step 3 and paste them into Google Forms, Microsoft Forms, Typeform or whatever your company uses. It takes about five minutes for a ten-question survey, and you get to look at every question once more on the way in.

Step 5: Test it on two people before you send it

This is the step almost everyone skips. Pew calls pretesting “an essential step” in designing a questionnaire, and it doesn’t need a research budget. Send the form to two people who look like your real audience and ask them to fill it in while thinking out loud, or to tell you afterwards:

• Which question made you pause?
• Was there a question where none of the answers fit you?
• How long did it take?

Paste their comments back into the chat and ask the AI to fix exactly those spots. Ten minutes of testing saves you from fifty answers to a question everyone read differently.

A before and after example

Say you rolled out a shared inbox tool to a 25-person support team six weeks ago, and you need to decide whether to keep it. Here is what a lazy prompt tends to produce, next to what the brief-based version gives you.

Before (generic prompt):

Question 1 asks about two things in an agree/disagree format. Question 2 is double-barreled. Question 3 leads. Question 4 has overlapping options (is 10 in the first or second group?) and no option for more than 30.

After (brief + review pass):

Five questions, about three minutes, and every answer points at the keep-or-drop decision. That is the whole difference: the AI didn’t get smarter, the brief did.

How long should your survey be?

Shorter than you think. SurveyMonkey analysed roughly 100,000 surveys and found that people spend about 5 minutes on a 10-question survey, and that they spend less time per question the longer a survey gets. In their write-up on survey length they also report that abandon rates rose for surveys taking more than 7 to 8 minutes, and that people tolerated longer surveys for work or school than for customer surveys.

A practical rule from that: for clients or customers, aim for 5 to 8 questions. For your own team, 10 to 12 is usually fine if every question earns its place. Put the time in the intro (“takes about 3 minutes”) and make sure it is true.

After the answers come in: let AI do the first read

The open text answers are where the useful surprises hide, and also where most people give up reading. AI helps here too.

• In Google Forms, Gemini can summarize answers to short-answer and paragraph questions in the Responses tab once a question has 3 to 200 responses, on eligible plans (Google’s help page on summarizing responses).
• In Microsoft Forms, Copilot can summarize results into key findings from the responses view, per the Microsoft guide linked above.
• Anywhere else, export the responses as a CSV and give it to ChatGPT or Claude. Our guide to analyzing a spreadsheet with ChatGPT walks through how to do that and check the numbers.

One prompt that works well for open answers: “Group these answers into the 5 most common themes. For each theme, give the count and two quotes in the respondent’s exact words. Don’t add anything that isn’t in the answers.” Then read the quotes yourself. A summary is a starting point, not the finding.

For the full process, from cleaning the export to a one-page summary you can share, see our guide on how to use AI to analyze survey data.

What not to paste into a chatbot

Drafting questions is safe: there’s nothing private in “what do we ask our clients”. The answers are a different story. Responses can contain names, email addresses, and honest opinions people gave you in trust. Before you paste them anywhere, remove names and contact details, check what your company allows, and use the business version of the tool your workplace has approved. If you promised the survey was anonymous, keep it that way all the way through.

Common mistakes when you create a survey with AI

• Skipping the brief. Without the decision and the audience, you get a generic list of questions.
• Keeping every question the AI suggests. More questions mean less careful answers. Cut hard.
• Only agree/disagree scales. Mix in balanced choices and one or two open questions.
• No “Not sure” or “Other” option. People forced into a wrong box give you wrong data.
• Not testing. Two testers catch the confusing question before fifty people do.
• Trusting the AI summary without reading quotes. Always check the themes against real answers.

FAQ

Can ChatGPT create survey questions?

Yes, and it does it well when you give it the decision, the audience and the time limit. The prompts in Steps 2 and 3 work in ChatGPT, Claude and Gemini without changes. If you’re new to it, our ChatGPT tutorial for beginners gets you set up.

Is there a free AI survey generator?

The free versions of ChatGPT, Claude and Gemini can draft a full survey from a good prompt, and you can paste the result into the free version of Google Forms. The AI features built into Google Forms and Microsoft Forms need a paid plan or license, according to their help pages. Dedicated survey generators mostly wrap the same idea in a template. The quality comes from your brief.

How many questions should a survey have?

As few as your decision needs. For customers, 5 to 8 questions is a good target. SurveyMonkey’s data shows people take about 5 minutes for 10 questions and start dropping off past 7 to 8 minutes.

Can AI analyze survey results too?

Yes. Gemini in Google Forms and Copilot in Microsoft Forms can summarize responses, and any chatbot can group open answers into themes from a CSV export. Check counts and quotes yourself before you share conclusions.

Should I use Gemini or ChatGPT for this?

If your team already lives in Google Workspace, Gemini inside Google Forms saves you the copy and paste. If not, use whichever chatbot you already know. Our Gemini tutorial for beginners shows the basics if you want to try it.

Need right-and-wrong answers instead of opinions? A short scored quiz fits better than a survey when you are checking knowledge. Here is how to make a quiz with AI, with copy-paste prompts for ChatGPT or Claude.

Your next step

Think of the one decision you’ve been putting off because you “need input first”. Write the four-line brief from Step 1, run it through the two prompts, and send the form to two testers today. By tomorrow you can have it out to the people whose answers you need. For more small workflows like this, the kind that take an annoying task off your plate, start with the free Launchpad on our home page.

The candidate is booked for Thursday at two. You have their CV open, a job description you wrote last month, and a blank page where your interview questions should be. So you reach for the same five you always ask: tell me about yourself, your biggest weakness, where you see yourself in five years.

Here’s the short answer. To use AI to create interview questions, don’t ask ChatGPT or Claude for “good interview questions for a project manager”. Paste in the job description, name the three to five skills that really decide success in the role, and ask for one behavioral question and one situational question per skill, each with follow-up probes and a short note on what a strong answer sounds like. Then ask for a simple 1 to 5 scoring guide and a check for anything you shouldn’t ask. That takes about ten minutes. You end up with a question set you can use for every candidate, so you compare people on the same things instead of on gut feeling.

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

Why the usual interview questions tell you so little

Most interviews are a friendly chat. You follow whatever sounds interesting, you like some people more than others, and afterwards you go with the feeling. It feels thorough. It isn’t.

Iris Bohnet, a behavioral economist at Harvard Kennedy School, put it bluntly in Harvard Business Review: unstructured interviews get the highest ratings from hiring managers for how useful they feel, yet dozens of studies found them among the worst predictors of how someone will actually do the job (HBR, How to Take the Bias Out of Interviews).

The fix is not more clever questions. It’s structure. The US Office of Personnel Management, the federal government’s HR agency, sums up the research: interviews that ask every applicant the same predefined questions and follow-ups, and score answers against set benchmarks, show higher validity, better agreement between interviewers and less adverse impact than loose conversations. Questions built on the specific skills the job needs do best (OPM, Structured Interviews).

That is exactly the boring, careful work AI is good at. It won’t decide who to hire. It will do the drafting that most of us skip because it takes an hour.

Step 1: Start from the job, not from a list of “top 50 questions”

Generic question lists are where most AI interview question generators start, and that’s why their output feels generic. Your input decides the quality.

Before you open ChatGPT, write down three things. It takes five minutes:

• The job description. If you don’t have a good one yet, our guide on how to use AI to write a job description gets you there first.
• Three to five must-have skills. Not twelve. The few things that, if missing, mean this hire fails. For a customer support team lead that might be calming angry customers, coaching a struggling agent, and prioritising a queue on a bad day.
• What good looks like in month three. One or two sentences per skill. “Handles escalations without passing them up” is useful. “Great communicator” is not.

That last point is the one people skip, and it’s the one that makes the AI’s scoring guide sharp later.

Step 2: The prompt that writes your interview questions

There are two proven question types, and OPM describes both. Behavioral questions ask what someone did in a real past situation (“Tell me about a time you…”). Situational questions ask what they would do in a realistic scenario from the job. Both work; behavioral ones hold up especially well for complex professional and managerial roles (OPM). Ask the AI for one of each per skill.

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

The “Missing info” line matters more than it looks. Without it, the AI fills gaps with invented details about your team, and you end up asking a candidate about a tool you don’t even use.

Step 3: Ask for a scoring guide before you meet anyone

Questions alone are half the structure. The other half is deciding, in advance, what a 1, a 3 and a 5 sound like. Otherwise the candidate who was charming at the start gets a 5 on everything.

If several people interview, OPM’s description of how structured panels work is a good model: each interviewer scores on their own first, then the panel talks through the big differences. Doing it in that order stops the most senior voice in the room from setting everyone’s score.

Step 4: Check the questions for bias and anything you shouldn’t ask

AI drafts can slip in questions that sound harmless and aren’t. “Where did you grow up?” or “Do you have kids?” have no place in an interview. In the US, the EEOC says pre-employment questions should stick to what’s needed to judge if someone is qualified, and that questions which reveal protected characteristics, such as age, religion, disability or family status, should generally be avoided (EEOC, Prohibited Employment Policies/Practices). Most other countries have similar rules, so check your own.

Run one more pass:

Treat this as a second pair of eyes, not legal advice. If you’re unsure about a rule where you hire, ask your HR person or an employment lawyer.

A before and after example

Say you’re hiring that customer support team lead. Here’s a question most of us would write on the spot:

Before: “How do you handle difficult customers?”

Every candidate answers that one the same way: “I stay calm, I listen, I find a solution.” You learn nothing.

Here’s the kind of question set the prompt above gives you for the skill “calming angry customers”, after one quick round of edits:

After (behavioral): “Tell me about the last time a customer was angry with you or your team. What had gone wrong, what did you say first, and how did it end?”

Probes: “What would you do differently now?” and “What did you tell your team afterwards?”

After (situational): “It’s Friday at 4pm. A customer emails the CEO directly, saying their order is three weeks late and they’re cancelling. The agent who owns the ticket is out sick. Walk me through your next hour.”

Strong answer signs: names a concrete first action, owns the problem without blaming the agent, mentions following up after the fix. Weak answer sign: stays general and never says what they’d actually write or say.

Same ten minutes of your time. A completely different conversation.

How to generate interview questions from a job description

Many people search for an AI interview questions generator based on a job description. You don’t need a special tool for this. The prompt in Step 2 already does it: the job description is the main input, and the “use details from the description” rule makes the situational questions specific to your role.

Two tips make it work better:

• Cut the fluff out of the description first. Company history and benefits don’t help the AI write questions. Paste the responsibilities and requirements only.
• Tell it which requirements matter most. A job description lists ten things. If you don’t rank them, the AI gives every line equal weight and your interview becomes a checklist.

Free versions of ChatGPT, Claude and Gemini all handle this well. If you’re new to these tools, our ChatGPT tutorial for beginners covers the basics in a few minutes, and how to use Claude for work shows the same idea in Claude.

Tailoring questions to one candidate’s CV (carefully)

A fair interview asks everyone the same core questions. You can still add two or three follow-ups about a specific CV, as long as the core set stays the same for all candidates. Ask the AI something like: “Here are my core questions and this candidate’s work history. Suggest three follow-up questions about gaps or claims I should explore. Don’t change the core questions.”

Before you paste a CV anywhere, remove the name, contact details, photo and date of birth. You only need the work history. Also check your settings. In ChatGPT’s personal plans, the “Improve the model for everyone” setting is on by default, which means your chats can be used for training unless you turn it off under Settings, Data controls. Temporary chats aren’t used for training, and business workspaces aren’t used for training by default (OpenAI Help, Data controls in ChatGPT). If your company has rules about candidate data, they win.

Common mistakes to avoid

• Asking for “the best interview questions”. You get the same list as everyone else. Always paste your role and your must-haves.
• Too many questions. In a 45 minute interview, eight to ten questions with probes is plenty. A candidate rushing through twenty gives you twenty shallow answers.
• Skipping the scoring guide. Without it, you’re back to gut feeling with nicer questions.
• Letting AI change the set between candidates. Generate once, edit once, then freeze it for this hiring round.
• Using the output without reading it. Read every question out loud once. If it sounds odd when you say it, rewrite it.
• Letting AI judge the answers. Your notes and your scores decide. AI helps you prepare, not decide who gets the job.

Questions people ask about this

Can I use ChatGPT to create interview questions for free?

Yes. The free versions of ChatGPT, Claude and Gemini can all write a full question set and scoring guide with the prompts above. Paid plans mostly help with longer documents and higher limits.

How many interview questions should I prepare?

Plan two questions per must-have skill, so six to ten for three to five skills, plus probes. For a 45 minute slot that leaves room for introductions and the candidate’s own questions.

Should I use behavioral or situational questions?

Both. Behavioral questions show what someone has actually done. Situational ones help with candidates who haven’t faced that exact situation yet, like a first-time team lead. If you want candidates to give fuller answers, many career services teach the STAR method (situation, task, action, result), and MIT’s career office has a clear guide on it (MIT CAPD, The STAR method). Your probes can gently steer someone through those four parts.

Can AI write the answers I should expect?

It can describe the signs of a strong answer, which is what the scoring guide needs. Don’t use it to write a “perfect answer” you compare word for word. Good candidates answer in their own words, from their own experience.

What about performance reviews and internal promotions?

The same structure works for internal interviews. For the review side of things, see how to use AI to write a performance review.

After the hire, a short new-starter survey tells you whether your interview process worked. Here is how to use AI to create a survey people actually finish.

Once you’ve made the hire, the same approach helps with their first months. See how to use AI for employee onboarding for a 30-60-90 day plan, a first-week checklist and a welcome email.

Your next interview, ten minutes from now

Pick the next role you’re hiring for. Write down the three skills that matter most and what good looks like in month three. Paste them into the Step 2 prompt, then run the scoring and bias checks. You’ll walk into Thursday’s interview with questions that actually tell candidates apart, and a sheet that helps you defend your choice afterwards.

The AI writes the drafts. You decide what the job really needs, and who fits it.

We share one practical workflow like this every week inside Launchpad, our free starting point for using AI at work. You can join from our home page or right here:

The meeting is tomorrow at ten. You booked it, six people accepted, and the invite still says nothing but the title. You know roughly what needs to happen. Writing it down in a way that keeps the hour on track is the part that keeps sliding.

Here’s the short answer. To use AI to create a meeting agenda, don’t ask for “an agenda for our project meeting”. Give it the goal of the meeting, who is coming, how long you have and the decisions you need, then ask it to turn each topic into a question the group has to answer, with a time box and an owner next to every item. ChatGPT, Claude, Gemini or Copilot can do this in under a minute. You spend two more minutes cutting what doesn’t belong. The AI writes the structure. You decide what the meeting is for.

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

Why most agendas don’t help (and what AI should fix)

An agenda on its own doesn’t make a meeting better. Meeting researcher Steven Rogelberg writes in Harvard Business Review that research has found little to no relationship between having an agenda and how attendees rate the meeting. What matters is whether the items are relevant and how the discussion is run. On his own site he adds that an estimated half of agendas are simply recycled from the last meeting.

That’s exactly the trap with AI. Ask a chatbot for “a team meeting agenda” and you get the generic one: welcome, updates, discussion, any other business, next steps. It looks tidy and changes nothing.

So the job you give the AI is different. You want it to do three things a tired organizer skips:

• Turn topics into questions. Rogelberg’s advice is to frame items as questions to be answered, not topics to discuss. “Budget” becomes “Where do we cut 5,000 from Q1 without delaying the launch?”
• Put the important item first. In the same essay he notes that research found early agenda items get a disproportionate share of time and attention, whatever their importance.
• Say what each item is for. Roger Schwarz suggests in Harvard Business Review that every item states whether it’s for sharing information, gathering input or making a decision, with a realistic time estimate and a clear note on how people should prepare.

A model is good at all three. It’s fast at rephrasing, it doesn’t get bored ordering things, and it will happily add time boxes that add up to your slot.

How to use AI to create a meeting agenda in 4 steps

Step 1: Write a 60-second brief

Before you open any AI tool, jot down five lines. This is the part only you can do:

• Goal: what must be true when the meeting ends? (“We’ve picked a supplier.”)
• People: who is coming and what they bring. (“Sara, finance. Tom, ops. Two people from the supplier.”)
• Time: how long, and whether it’s in person or a video call.
• Open questions: the two or three things nobody has answered yet.
• Input you have: an email thread, last meeting’s notes, a short report.

If you can’t fill in the goal line, stop there. Rogelberg’s point is blunt: if you can’t come up with questions to answer, you may not need the meeting at all. Sometimes the best agenda is a short email instead.

Step 2: Paste the brief into this prompt

Open ChatGPT, Claude or Gemini and use this. Replace the brackets with your brief.

The “Missing info” line matters. Without it, a model will often fill gaps with plausible guesses, like an owner who isn’t coming or a deadline you never set. With it, you get a list of what to check instead.

Step 3: Cut and sharpen

Read the draft like an attendee who has three other meetings that day. Then ask the AI for one round of changes:

This is usually where a seven-item draft drops to four, and the meeting gets shorter.

Step 4: Send it with the invite

Put the agenda in the calendar invite, not in a separate document people won’t open. Send it at least a day ahead if anyone needs to prepare. Add the “send before” items as a short note above the agenda so people read them first.

A before and after example

Here’s an example from a small agency planning a client website relaunch. The organizer’s first version, typed in two minutes:

After the brief and the prompt above:

Same meeting, 15 minutes shorter, and every item ends in something you can write down. “Updates” disappeared because they went into the status sheet. “AOB” disappeared too: anything that matters now has its own question and owner.

Using the AI already built into your calendar

If your company runs on Microsoft 365 or Google Workspace, you may not need a separate chatbot.

• Outlook with Copilot: Microsoft’s support page explains that in new Outlook you create the meeting, add a title and attendees, open Copilot in the description (Alt+I) and choose “Auto draft an agenda”, or type your own prompt with the goal and talking points. With a Microsoft 365 Copilot license, “Schedule with Copilot” can also turn an email thread into an invite with a title and draft agenda.
• Google Docs with Gemini: Google’s Workspace Learning Center shows how to draft an agenda with “Help me write” in Docs, and how to type “@” and pick Meeting notes to link a doc to a Calendar event. Availability depends on your Workspace plan and admin settings.

These built-in tools are handy because they already know the title and attendees. Their weakness is the same as any chatbot: if you don’t give them the goal and the open questions, you get the generic agenda. Paste the rules from Step 2 into their prompt box and you get the better version.

A reusable agenda template prompt for recurring meetings

Weekly team meetings are where recycled agendas live. Instead of reusing last week’s list, keep this prompt in a note and run it the afternoon before:

That last line is a small gift to your team. Some weeks the honest answer is that a status message is enough.

If you also want the agenda to start from the previous meeting’s outcome, our guide on summarizing meeting notes with ChatGPT shows how to get clean decisions and actions out of a messy transcript. Paste that summary into the prompt above and the loop closes itself. Only have a recording of the last meeting? Here is how to transcribe audio with AI so you have text to work from.

What to keep out of the chatbot

Agendas often touch sensitive things: a reorganization, a client dispute, salaries. Before you paste background into a tool, check what your company allows. On personal ChatGPT plans, OpenAI’s data controls page says conversations can be used to improve its models unless you switch off “Improve the model for everyone”, and Temporary Chats are not used for training. It also says content from ChatGPT Business and Enterprise workspaces is not used for training by default. When in doubt, describe the situation in general terms (“a budget overrun with a supplier”) instead of pasting names and numbers.

Common mistakes

• Asking for an agenda without a goal. You’ll get a pretty list that steers nowhere.
• Keeping every item the AI suggests. Models like to be thorough. Your attendees don’t. Cut to what needs the group.
• Time boxes that don’t add up. Check the total yourself. A 60-minute plan in a 45-minute slot means the last item, often the decision, gets skipped.
• No owner per item. An item without a lead tends to become a general chat.
• Sending it an hour before. If people need to read something, they need the agenda the day before.

FAQ

Can ChatGPT make a meeting agenda?

Yes. It’s one of the easiest work tasks to hand over, as long as you give it the goal, the people, the length and the open questions. The prompt in Step 2 works in ChatGPT, Claude and Gemini without changes. New to ChatGPT? Start with our ChatGPT tutorial for beginners.

Is there a free AI meeting agenda generator?

The free versions of ChatGPT, Claude and Gemini all handle this task well with a good prompt, so you don’t need a dedicated generator. Dedicated tools mostly add a template around the same idea. The quality comes from your brief, not the tool.

What should a good meeting agenda include?

The goal of the meeting, each item written as a question, its purpose (inform, input or decide), a time box, an owner, and what to read or bring. End with a few minutes to confirm decisions and who does what.

Can AI turn the agenda into a project plan afterwards?

Yes. Once the meeting has made its decisions, paste the agenda plus the outcomes into a chatbot and ask for tasks, owners and dates. Our walkthrough on using AI to create a project plan takes it from there.

Your next step

Open your calendar, find the next meeting you’re running, and spend 60 seconds on the brief from Step 1. Run it through the prompt and send the agenda with the invite today. Your attendees will notice the difference before the meeting even starts. To find room for that prep in a busy day, see how to use AI to plan your day. Once the meeting is under way, an AI note taker for meetings can capture decisions so you stay present instead of typing. For more small workflows like this, the kind that take an annoying task off your plate, start with the free Launchpad on our home page.

Your manager just said “can you put a plan together for this by Friday?” and you nodded. Now you’re staring at a blank spreadsheet, the project lives in five email threads and your head, and you’re not sure where it even starts.

Here’s the short answer. To use AI to create a project plan, don’t type “make me a project plan for a website relaunch”. Give ChatGPT, Claude or Gemini a five-minute brain dump of the goal, the deadline, the people and the limits. Let it ask you questions first. Then have it break the work into deliverables, turn those into a dated task list with owners, and poke holes in it before anyone sees it. The AI does the structuring. You keep the dates honest.

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

Can ChatGPT create a project plan?

Yes, and it’s good at it. A project plan is mostly structure: what has to be delivered, which tasks get you there, in what order, by whom, by when. Turning a messy description into that kind of structure is exactly what language models do well.

What it can’t do is know your project. It doesn’t know that the designer is on holiday in week three, that legal takes ten days to sign anything, or that the client always changes their mind after the first draft. Ask for a plan without that context and you get a neat, confident, generic plan with made-up durations. OpenAI itself explains that language models produce “plausible but false statements” partly because standard training rewards guessing over admitting uncertainty (OpenAI, Why language models hallucinate). A guessed task duration is exactly that: a guess that looks like a fact.

So the deal is the same as with any work task you hand to AI. You bring the facts and the judgment. The AI brings the structure and the speed.

How to use AI to create a project plan in 6 steps

This works with the free version of ChatGPT, Claude or Gemini. Keep all six steps in one chat so the AI remembers what you’ve already agreed on.

Step 1: Brain-dump the project (5 minutes, no AI yet)

Open a note or record a voice memo and answer these in plain words. Messy is fine.

• The goal: what is different when this project is done? “New CRM live and the sales team using it” beats “CRM project”.
• The deadline, and whether it’s fixed (a trade show) or soft (someone’s wish).
• Done means: the two or three things that must be true to call it finished.
• People: who’s involved, how many hours a week each one really has, and any holidays you know about.
• Limits: budget, tools you must use, approvals you need, things that are out of scope.
• What worries you: the part you suspect will go wrong. You usually know.

Paste in anything you already have: the client email, meeting notes, a rough budget. If your notes are a wall of text from a kickoff call, our guide on turning meeting notes into a report with ChatGPT cleans them up first.

Step 2: Let the AI ask you questions before it plans

This is the step that separates a useful plan from a generic one. Paste your brain dump below this prompt:

Answer each question in a line or two. This is usually where the plan gets real: “who signs off on the design?” is a question you’d rather answer now than in week four.

Step 3: Break the work into deliverables

Project managers call this a work breakdown structure. It sounds formal, but it’s simply the full list of things the project has to produce, split into smaller pieces. The Project Management Institute describes a good one as deliverable-oriented and complete: the pieces together should add up to 100% of the work in scope, including the project management itself, and nothing outside it (PMI, Developing and elaborating effective work breakdown structures). That 100% idea is your best check on what the AI gives you.

Read the list slowly. Cross out what doesn’t belong, add what’s missing. This is your scope now, and it’s the thing you’ll point at when someone asks for “just one more small thing” in week five.

Step 4: Turn it into a timeline with owners

Now the plan gets dates. The trick is to make the AI ask for durations instead of inventing them, and to put the result in a table you can paste straight into Excel or Google Sheets.

Those brackets matter. They turn every AI guess into a visible question for you, instead of a quiet assumption buried in row 23. Go through each one and replace it with a real estimate, ideally from the person who’ll do the work. If you want to sort, filter or chart the table afterwards, our guide to analysing a spreadsheet with ChatGPT picks up from here.

Step 5: Stress-test the plan (risks and buffer)

Almost every first plan is too optimistic, and it’s not because you’re bad at planning. In a well-known study, psychology students predicted they’d hand in their thesis in about 34 days. They took about 55 on average, and fewer than half finished even by their own “if everything goes wrong” date (Buehler, Griffin and Ross, 1994, Journal of Personality and Social Psychology). People plan for how the work should go, not how it usually goes.

AI is a useful outsider here because it has no stake in your plan looking good:

You don’t have to accept every point. But if it flags that your designer has 60 hours of work in a 20-hour week, believe it.

Step 6: Make it visible and keep it alive

A plan nobody looks at is a document, not a plan. Two easy ways to share it:

• A spreadsheet: ask for the final schedule “as CSV” and paste it into Excel or Google Sheets. Add a status column people update each week.
• A Gantt chart: ask “Turn this schedule into a Mermaid Gantt chart.” Mermaid is a free text-to-diagram tool, and its Gantt chart syntax is simple enough that AI writes it well. Paste the result into the free Mermaid Live Editor and you get a timeline image you can drop into a slide or an email.
• A flowchart: for the steps, decisions and handoffs inside the plan, see how to make a flowchart with AI. It works in the same free Mermaid editor.

To keep the plan alive, put it in a project space inside your AI tool. ChatGPT Projects keep related chats, files and instructions together so ChatGPT can use the same context each time (OpenAI Help Center), and Claude has Projects too, including on free accounts, with a limit of five (Claude Help Center). Upload the plan once, and every Friday you can paste in what happened and ask “what’s slipping, and what should I tell the team?”

Example: a 6-week CRM switch for a small team

Here’s what this looks like on a real kind of project. Say you run a five-person agency and you’re moving from spreadsheets to a CRM in six weeks. Your brain dump is ten lines: the goal (all leads and clients in the CRM, the team using it daily), the deadline (before the busy season starts), who’s involved (you, an assistant with 10 hours a week, a freelancer for the data import), and your worry (the old spreadsheet is a mess).

In Step 2, the AI asks things like: Who decides which CRM? Is the data clean-up part of this project? Does anyone need training? Do you need to connect email or a website form? Each answer changes the plan.

After Steps 3 and 4, a slice of the schedule looks like this:

The “(5?)” next to the clean-up is the AI telling you it guessed. Your assistant looks at the spreadsheet and says it’s more like eight days, which pushes the import back. Step 5 then flags that the website form has no owner and that the whole plan hangs on the clean-up. That’s two real problems found in week zero instead of week five. And if the CRM is meant to automate follow-ups later, our list of email automation examples is the natural next project.

And the Mermaid version of the same plan starts like this, readable even before you render it:

More ChatGPT prompts for project planning

Once the plan exists, the same chat can take care of the admin around it. A few prompts worth saving:

• Kickoff email: “Write a short kickoff email to the team: the goal, the deadline, who owns what in the first two weeks, and what I need from each person by Friday.”
• Kickoff meeting agenda: “Turn this plan into a 45-minute kickoff agenda with a goal, time boxes and the decisions we need to make.” Our guide on how to use AI to create a meeting agenda walks through it step by step.
• Weekly status update: “Here’s what got done this week [paste]. Update the schedule, mark what’s late, and write a five-line status update for my manager.” For a fully automated version, see our guide to automated status updates with Gemini.
• Scope change: “The client wants to add [request]. Show me what it does to the timeline and draft a polite reply that explains the trade-off.”
• One-page summary: “Summarise this plan on one page for someone who has two minutes: goal, deadline, three biggest risks, and what I need from them.”

Planning a whole new business or side venture rather than a single project? Our guide on how to use AI to write a business plan covers the market, the numbers and the pitch.

Mistakes to avoid

• Trusting dates you didn’t check. Every duration the AI made up is a risk with a nice font. Keep the brackets until a person confirms the number.
• Planning people at 100%. Someone with “40 hours” also has email, meetings and other projects. Tell the AI their real project hours.
• Too much detail too far ahead. Plan the next two or three weeks in detail and keep later phases at the deliverable level. PMI’s guidance on the WBS makes the same point: it can be more detailed on near-term work and more summary on future work, filling in as things become clear (PMI, The WBS: basic principles).
• Pasting things you shouldn’t. Client names, contracts and salaries may be confidential. Check your company’s AI rules, or replace names with roles (“Client A”, “Developer”).
• Skipping the “done means” line. Without it, the AI and your team will each assume a different finish line.

If you’d rather build this habit with Claude in particular, our guide on how to use Claude for work covers the basics, including how to set up a project space.

FAQ

Can ChatGPT create project plans for free?

Yes. The free versions of ChatGPT, Claude and Gemini can all build a work breakdown, a schedule table and a risk list from your notes. What makes the plan good is the brief and the questions in Step 2, not a paid plan.

Can AI make a Gantt chart?

Yes. Ask it to turn your schedule into a Mermaid Gantt chart and paste the result into the Mermaid Live Editor, or ask for a CSV and build the chart in your spreadsheet or project tool.

Should I use AI instead of project management software?

Use them together. AI is great at turning a vague idea into a structured plan and at writing the updates around it. A tool like a shared spreadsheet, Trello or Asana is better at tracking who ticked off what. Many people draft the plan with AI, then import the table into the tool the team already uses.

What’s the best prompt for project planning?

One that makes the AI ask questions first, forbids invented facts, and asks for a fixed table format. The Step 2 and Step 4 prompts above do exactly that.

How detailed should an AI project plan be?

Detailed enough that every task has one owner and fits in a few days. If a task says “build website” and takes 20 days, ask the AI to split it.

If the project starts with a choice, like which vendor, tool or approach to go with, settle it first with a decision matrix made with AI, then plan around the winner.

Your next step

Take the project that’s been sitting in your head all week, spend five minutes on the brain dump from Step 1, and run it through the Step 2 prompt today. You’ll have a first plan, with its weak spots marked, before your next meeting. Then, to turn this week’s tasks into a realistic schedule, here’s how to use AI to plan your day. For more small workflows like this, the kind that take a dreaded task off your plate, start with the free Launchpad on our home page.

Once the plan exists, the daily work starts: status updates, risk checks, stakeholder emails. These ChatGPT prompts for project management cover that part.

Once the plan exists, here is how to use AI for project management day to day, from status updates to risk logs.

Planning a building project? Read project management software for architects for the tools that fit studios.

A new review lands on your Google profile. Two stars. “Waited 40 minutes, nobody told us anything.” Your stomach drops, you start typing a reply, delete it, start again, and twenty minutes later you still have nothing you would want a future customer to read.

Here’s the short answer. To use AI to respond to Google reviews, don’t ask it to “reply to this review”. Give it three things first: the review itself, the facts only you know (what actually happened, what you can offer), and how your business talks. Then let ChatGPT, Claude or Gemini write a short draft, check it against a five-point list, and post it yourself. The AI finds the calm words. You bring the truth and press the button.

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

Can AI respond to Google reviews?

Yes, in two ways. The simple way works for everyone: you copy a review into ChatGPT, Claude or Gemini, get a draft, and paste your final reply into your Business Profile. The second way is newer. Google now lets you connect your Business Profile to the Gemini web app, and Gemini can then fetch your reviews and draft replies for you. Google announced that connection on its official blog in June 2026.

Google has also been testing a “Reply to reviews with AI” button inside Business Profile itself. Search Engine Land reported in March 2026 that the test was limited, showed up unevenly across accounts, and was seen in the U.S., Brazil and India but not widely in Europe. So if you don’t see that button, nothing is wrong with your account.

Whichever route you take, one thing doesn’t change. According to Google’s own help page on managing reviews, your reply is public, it shows as coming from your business, the reviewer gets notified, and Google checks replies against its content policies before they go live. That check usually takes up to 10 minutes, but it can take up to 30 days. A reply is a small public statement from your business. Treat the AI draft like a first draft from a new colleague: useful, never final.

What a good review reply actually needs

Before you write a prompt, know what you’re asking for. A guide by a Google Business Profile Product Expert in Google’s own community boils it down well:

• Positive reviews: short and sincere. Thank them, mention what they liked, no hard sell.
• Negative reviews: no generic copy-paste. Understand what went wrong, apologise if it’s warranted, stay professional, and never get defensive or personal.
• Timing: that guide suggests aiming to reply within 24 to 72 hours.

Notice what’s missing from that list: long paragraphs, marketing slogans, and the phrase “we value your feedback” three times. That’s exactly what AI writes when you give it nothing to work with. So the trick is not a clever prompt. It’s giving the AI the facts it can’t guess.

Step 1: Write a five-line “reply brief” once

This takes about ten minutes and you reuse it for every review. Open a note and fill in:

Business: [name, what you do, city]
Voice: [e.g. friendly and direct, first names, no exclamation marks]
Sign-off: [e.g. "Sanne, owner" or "The team at ..."]
How unhappy customers can reach us: [business email or phone]
Never say: [e.g. promises about refunds, staff names, anything about health]

The sign-off matters more than it looks. Google shows every reply as coming from the business, so a human name at the end makes it feel like a person read the review. The same community guide suggests that when a review mentions a specific employee, that person can reply and add their name or initials.

The “never say” line is your safety rail. AI likes to promise things (“we’ll refund you in full”) because it sounds kind. Only you know what you can actually offer.

Step 2: The prompt for positive reviews

Positive reviews are where AI saves you the most time, because they’re easy to answer and easy to put off. Paste your brief, then this:

You help me reply to Google reviews for my business. Use the brief above for voice and sign-off.

Write a reply to this review:
"[paste the review]"

Rules:
- 2 to 3 sentences, max 60 words.
- Thank them and mention ONE specific thing they wrote about.
- No marketing slogans, no discounts, no links.
- Don't repeat the business name more than once.
- Don't invent details that are not in the review or the brief.
Give me 2 versions: one warm, one shorter.

Asking for two versions is the small move that makes this work. You pick the one that sounds like you, change a word, done. If you have ten unanswered five-star reviews, paste them all in one go and ask for one reply per review, each mentioning something different. Then read every single one before it goes out.

Step 3: The prompt for negative reviews

This is where people get burned, so slow down. Write the facts down first, in your own words, even if they’re messy. Then:

Help me reply to a negative Google review. Use the brief above.

The review:
"[paste the review]"

What actually happened (my notes, not for publishing as-is):
[e.g. Saturday 7pm, two staff out sick, kitchen was 30 min behind, nobody went to the tables to explain]

What I can honestly offer:
[e.g. invite them to email me directly / we now tell tables when the kitchen is behind]

Rules:
- Max 80 words, calm and human, no excuses.
- Acknowledge their experience in plain words.
- Apologise only for what we actually got wrong.
- Mention the one concrete change or offer from my notes, nothing more.
- Invite them to contact us using the channel in the brief.
- Never share any personal details about the customer or our staff.
- Never argue about facts in public.
Then list anything in your draft that I should double-check.

That last line is my favourite part. It makes the AI point at its own weak spots (“you mentioned a change in process; confirm this is true”) instead of hiding them in nice wording.

A habit worth building: if a review makes you angry, write your notes, run the prompt, and don’t post until later that day. The AI reply will be calm either way. You want to be calm too when you read it.

Step 4: Check the draft before you post (the five-point list)

Read the AI’s reply once, slowly, against these five questions:

1. Is every fact true? If the draft says “we’ve retrained the team”, did you? Delete anything you can’t back up.
2. Is anything private in it? Google’s Maps content policy doesn’t allow posting other people’s personal information without consent, such as a full name, last name, or financial or medical details. Your own business email or phone number is allowed. Booking details, order numbers and staff surnames stay out.
3. Does it offer anything in exchange for changing the review? The same policy forbids offering payment, discounts or free goods for posting a review or for revising or removing a negative one. “Email me and I’ll give you 20% off if you update your review” breaks that rule. “Email me so I can make this right” doesn’t.
4. Does it sound like you? Read it out loud. If you’d never say “we sincerely apologise for any inconvenience caused” across the counter, change it.
5. Is it short? Most good replies fit in two to four sentences. A wall of text under a two-star review reads as defensive, even when it isn’t.

This check takes about a minute. It’s the minute that keeps a time-saver from turning into a public mistake.

Step 5: Post it, then keep an eye on it

Posting is the same as always: go to your Business Profile, select Read reviews, then Reply next to the review, paste your final text and confirm. Google’s help page explains you can also edit or delete your reply later from the same place.

Two things people miss. First, Google reviews your reply before it shows, so if it doesn’t appear right away, give it time. If it’s rejected, you’ll be asked to edit it. Second, customers can change their review after reading your reply, and the review date updates when they do. The community guide adds that owners aren’t notified about those edits, and an old reply can then look like it came before the new review. It’s worth a quick look at your recent reviews once a week.

Using Gemini connected to your Business Profile

If you have one business location, the Gemini connection can cut out the copy-paste. Per Google’s help page, here’s what you need:

• You’re the owner or manager of only one verified Business Profile. If you manage several, it isn’t available.
• You’re signed in to the Gemini web app with a personal Google Account linked to that profile. Work or school accounts don’t work for this yet.
• Keep Activity is switched on in Gemini.
• You’re 18 or over, and you use it at gemini.google.com.

Then you go to gemini.google.com, ask something about your Business Profile, and follow the steps to connect it. If Gemini doesn’t pick it up, Google suggests adding @Google Business Profile to your prompt. Once connected, Gemini can fetch your star ratings and reviews, draft replies, and summarise themes in your feedback.

That last one is underrated. Ask it something like:

@Google Business Profile Look at my reviews from the last 3 months. What are the 3 things customers praise most and the 3 things they complain about most? Quote one short example review for each.

Now your reviews aren’t just something to answer. They’re a to-do list for the business. And because Gemini can also post and delete replies through that connection, keep the same habit as above: you read every draft before anything goes live.

New to Gemini? Our Gemini tutorial for beginners walks you through the basics first.

A before-and-after example

Here’s an illustration, a made-up café, to show the difference a brief makes.

The review: “Coffee was great but we waited 40 minutes for two sandwiches and nobody said a word.”

What AI writes with no brief:

Dear valued customer, thank you for taking the time to share your feedback! We sincerely apologise for any inconvenience you may have experienced. At [Café Name], customer satisfaction is our top priority and we strive to provide excellent service at all times. We hope to see you again soon!

Polite. Empty. It could sit under any review of any business, and readers can tell.

What AI writes with the brief and the notes:

You're right, and I'm sorry. Forty minutes for two sandwiches is too long, and someone should have come to your table to explain. We were short two people that Saturday. That's on us, not on you. We now tell every table when the kitchen is running behind. Glad the coffee hit the spot. If you'd like to tell me more, email me at hello@[yourcafe].com. Sanne, owner

Same AI, same two minutes. The difference is the facts you gave it.

What AI should never do with your reviews

Using AI to reply is fine. Using it to fake reviews is a different story, and it’s worth saying plainly because AI makes it so easy.

• Google: its Maps policy says reviews must reflect a genuine experience. Fake engagement, paid reviews and reviews from conflicts of interest (like staff or family) aren’t allowed and get removed.
• U.S. law: the FTC rule on consumer reviews, 16 CFR 465.2, makes it a violation for a business to write or create a review that misrepresents that the reviewer exists, used the product, or what their experience was. The final rule in the Federal Register took effect on October 21, 2024, and it names AI tools as something that makes fake reviews easier to write.

Outside the U.S., check the rules where you do business. This isn’t legal advice, but the safe line is simple: AI helps you answer real customers. It never pretends to be one.

Common mistakes when you use AI for review replies

• Same reply under every review. Google’s policy also lists repetitive content as not allowed on Maps, and to a reader ten identical “Thanks for the kind words!” replies look like nobody’s home. Ask the AI to mention something specific each time.
• Letting it apologise for things you didn’t do. An apology for a mistake you never made reads as an admission. Only apologise for what’s real.
• Pasting customer data into the chat. The review text is public, so that’s fine. Booking systems, phone numbers and order histories don’t belong in your prompt.
• Arguing in the reply. If a review is false or breaks Google’s rules, Google lets you flag it for review. Do that instead of debating it in public.
• Answering only the bad ones. The community guide recommends replying to positive reviews too. With AI drafting, that’s a two-minute job.

If you write a lot of customer messages beyond reviews, the same “brief first, draft second” approach works for email too. See how to use ChatGPT to draft reply emails faster and our ChatGPT client communication templates. And when a reply comes out stiff, the tips in make this email sound more professional work just as well the other way round, to make it sound more human.

Reviews only come from the customers who choose to write one. To hear from the quiet majority too, see how to use AI to create a survey people actually finish.

FAQ

Can AI respond to Google reviews automatically?

Partly. Gemini connected to your Business Profile can draft replies, and Google has tested an AI reply button inside Business Profile. Search Engine Land noted conflicting reports on whether the test’s bulk replies still needed review. My advice: even if a tool can post on its own, read each reply first. It’s public, and it’s your name.

Does Google penalise AI-written review replies?

Google’s help pages don’t say AI replies are a problem, and Google itself offers AI drafting. What it does check is whether your reply follows its content policies. A reply that’s accurate, polite and free of personal data is fine however you drafted it.

Should I reply to every review?

The Google community guide recommends replying to all reviews, positive ones included. With a reply brief and the prompts above, most replies take a couple of minutes.

Which AI is best for review replies?

For drafting, ChatGPT, Claude and Gemini all do a good job once you give them a brief. Gemini has the extra option of connecting to your Business Profile if you manage a single location. If you’re choosing between them, our ChatGPT vs Claude for writing comparison helps.

How fast should I reply?

The Google community guide suggests aiming for 24 to 72 hours. For a negative review, a calm reply the same day beats a perfect one next week.

Your next ten minutes

Open your Business Profile and find the oldest review you never answered. Write the five-line reply brief, run the prompt from Step 2 or Step 3, check it against the five points, and post it. That’s one review done, and a system for every review after it.

If you’d like more workflows like this, small things that give you back real hours at work, start on our home page and join Launchpad. It’s free.

It’s review season. You have six people to write up, a blank form for each, and a calendar that says they are due on Friday. You know who did great work and who struggled. Turning that into fair, specific sentences is the part that eats your evenings.

Here’s the short answer. To use AI to write a performance review, don’t ask it to “write a review for Sam”. Spend ten minutes collecting real evidence first: the goals, three to five concrete examples with results, one or two things to improve, and what you want for next quarter. Remove names and private details, give that to ChatGPT, Claude or Gemini with a clear prompt, then check the draft for vague praise and anything it made up. The AI does the wording. You keep the judgment.

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

Can you use AI to write a performance review?

Yes, for the writing. Not for the verdict.

A chatbot is good at structure, tone and turning messy notes into clear paragraphs. It has no idea what your colleague actually did this year. If you feed it nothing, it fills the gap with friendly filler like “consistently exceeds expectations” or “a true team player”. That reads fine and tells the person nothing.

That matters more than it sounds. When Gallup looked at how reviews land, only 14% of employees strongly agreed their performance reviews inspire them to improve. Generic text is a big part of why. A review full of empty praise feels like a form someone had to fill in, because it was.

So the rule for this whole guide: you bring the facts and the decision, the AI brings the clear sentences. If you are in the EU, there is a legal reason too. The EU AI Act lists AI systems “intended to be used to make decisions” about promotion or termination, or “to monitor and evaluate the performance and behaviour” of workers, as high-risk (Annex III, point 4). Using a chatbot to polish your own written assessment is not the same as letting software rate people, but it is a good reason to keep the rating and the decision firmly with you. This is not legal advice; check your company’s HR policy before you start.

Step 1: Collect the evidence (10 minutes per person)

This is the step that decides whether the review is useful. Open a plain note and answer five questions for each person:

1. What were their goals for this period? Copy them from last review, the job description or your one-to-one notes.
2. Three to five specific things they did, with the result. Not “good with clients” but “took over the Jansen account in March, cut response time on their tickets from two days to same-day”.
3. One or two things to improve, with a real example. “Two of the four monthly reports went out late” beats “needs to work on time management”.
4. What you want from them next period. One or two clear goals.
5. The rating or outcome you have already decided. Decide this yourself before the AI writes anything.

Where do the examples come from? Your one-to-one notes, project updates, emails where someone thanked them, the tickets they closed. If you keep meeting notes, our guide on turning meeting notes into a report with ChatGPT shows how to pull the key points out of a pile of notes quickly.

A small tip from doing this myself: if you can’t think of three concrete results for someone, that is information. Ask them or a colleague before you write, rather than letting the AI paper over the gap.

Step 2: Strip names and private details

Performance notes are personal data. Before you paste anything into a chatbot, replace the name with “the employee” or a first initial, and leave out health, family situations and anything about pay you are not allowed to share.

Then check where your text goes:

• ChatGPT: on Free, Plus and Pro, “Improve the model for everyone” is on by default. You can turn it off under Settings, Data controls, or use a temporary chat, which is not used to train models. Business and Enterprise workspaces are not used for training by default (OpenAI, Data controls in ChatGPT).
• Claude: on Free, Pro and Max you choose whether your chats can be used for training, and you can change it in Privacy Settings. Team and Enterprise plans fall under commercial terms and are not part of that setting (Anthropic, Updates to Consumer Terms).
• Gemini: Google says that even with Keep Activity off, it still uses your chats to respond and to keep its service safe, including with help from human reviewers (Gemini Apps Privacy Hub).

The safest option is the AI tool your company already approved. If there isn’t one, anonymise first and use the privacy settings above.

Step 3: Give the AI a clear prompt

Now paste your notes in with a prompt that tells the AI what it may and may not do. This is the one I use:

You are helping a manager write a performance review.
Use ONLY the notes below. Do not invent results, numbers or examples.
If something is missing, list it under "Missing info" at the end instead of guessing.

Write the review in these sections:
1. Summary (3 sentences)
2. What went well (each point = what they did + the result)
3. Where to grow (each point = specific behaviour + what better looks like)
4. Goals for next period (2 goals, measurable)

Tone: warm, direct, plain English. No buzzwords like "rockstar",
"synergy" or "exceeds expectations". Max 400 words.

The rating is already decided: [rating]. Do not change it or argue with it.

Notes:
[paste your notes from Step 1]

Three parts of this prompt do the heavy lifting. “Use only the notes” stops the AI from inventing achievements. The “Missing info” list shows you where your evidence is thin. And stating the rating stops the text from drifting more positive or more negative than your actual decision.

Which tool? Any of the big three can do this. In my experience, Claude tends to write in a softer, more natural tone, and ChatGPT sticks to a fixed structure well. If you are choosing, our comparison of ChatGPT vs Claude for writing walks through the differences.

Step 4: Check for vague praise and bias

This is the step most people skip, and it is where AI can make your review better than what you would write alone.

Research by Shelley Correll and Caroline Simard at Stanford looked at real performance reviews and found that women were less likely than men to get specific feedback tied to business outcomes. In one company’s reviews, vague praise showed up in 57% of reviews for women against 43% for men (Harvard Business Review, 2016). Their advice is simple: use the same criteria for everyone, tie each piece of feedback to a result, and aim for reviews of similar length and detail.

You can ask the AI to check exactly that. After your draft is done, run a second prompt:

Review this performance review draft. Flag:
- any praise or criticism that is not tied to a specific example or result
- personality words ("abrasive", "too aggressive", "nice") instead of behaviour
- anything that is not in my original notes
Suggest a more specific rewrite for each flag.

If you write several reviews for people in similar roles, put them side by side and ask: “Do these reviews use the same criteria and a similar level of detail? Where is one much vaguer than the others?” That one question catches a lot.

Step 5: Make it sound like you

The person reading this knows how you talk. If the review suddenly sounds like a corporate brochure, they will notice, and they will trust it less.

Read the draft out loud. Swap any sentence you would never say in a meeting. Add one line only you could write: a moment you remember, a thank-you in your own words. Then delete anything from the AI that you can’t back up with an example.

If the tone still feels stiff, a quick pass with “make this sound like a friendly, direct manager talking to a colleague” usually helps. We cover more tone fixes in how to make an email sound more professional, and the same prompts work in reverse to make text sound less formal.

Example: the same review, before and after

Vague (what AI writes with no notes):

“Alex had a strong year and is a valued member of the team. Alex consistently delivers high-quality work and communicates well. Going forward, Alex should continue to develop leadership skills.”

Specific (what AI writes from real notes):

“Alex led the move of our client reporting to a shared dashboard in Q2. Clients now get their figures on the first working day of the month instead of the fifth, and two clients mentioned it in their renewal calls. Next period I’d like Alex to run the Monday planning meeting for two months, so the team gets used to coming to Alex with questions about priorities.”

Same person, same tool. The only difference is the ten minutes of evidence in Step 1.

Prompts for the hard parts

Writing constructive feedback without sounding harsh

Rewrite this feedback so it is honest and kind.
Keep the facts exactly as they are.
Structure: what happened, why it mattered, what better looks like next time.
Feedback: [your blunt version]

Write the blunt version first, in your own words. It is easier for the AI to soften an honest sentence than to sharpen a mushy one.

Turning comments into measurable goals

Turn these development points into 2 goals for the next 6 months.
Each goal: one sentence, a clear result, and how we will both know it's done.
Points: [your notes]

Preparing the conversation, not just the document

The written review is half the job. Gallup’s research also found that people who get feedback weekly rather than once a year are far more likely to say the feedback is meaningful (same Gallup article). The review should hold no surprises. Ask the AI: “Based on this review, what three questions might the employee ask me, and how could I answer them honestly?” Ten minutes with that list makes the meeting calmer for both of you.

Mistakes to avoid

• Letting the AI pick the rating. Decide first, then write.
• Pasting full names, health details or salary info into a personal chatbot account.
• Accepting numbers you didn’t give it. If the draft says “increased sales by 20%” and your notes don’t, delete it.
• Using one template for everyone. Six reviews that read the same tell your team you didn’t really look.
• Skipping the read-aloud. If you wouldn’t say it to their face, don’t put it in writing.

FAQ

Is it OK to use ChatGPT to write a performance review?

For the wording, yes, as long as the facts and the judgment are yours and you follow your company’s rules on AI and personal data. Many companies have an approved AI tool; use that if you have one.

Will my employee be able to tell?

They can tell when a review is generic. They can’t tell, and won’t mind, when it is specific, accurate and sounds like you. The edit in Step 5 is what makes the difference.

Can AI help with a self-review too?

Yes. The same steps work: collect your own results first, then ask the AI to structure them. Be extra careful it doesn’t inflate what you did.

How long does this take?

Plan on about 10 minutes for evidence and 10 to 15 for prompting, checking and editing, per person. The evidence step is the one not to rush.

Reviews and interviews have the same weak spot: vague questions. If you are interviewing for a new role or an internal move, see how to use AI to create interview questions tied to the real job.

Where to go from here

Reviews are one of many writing jobs where AI saves real hours once you give it good input. If you are hiring as well, the same approach works for writing a job description with AI, and our guide to using Claude for work shows more everyday tasks you can hand off.

Imagine next review season with your notes already collected, your prompts saved, and Friday evening free. That’s what we build together at Launchpad. Start on our home page, or sign up right here:

On the other side of the table? If your company asks you to rate yourself first, our guide on how to use AI to write a self evaluation uses the same evidence-first approach for your own review.

You need to hire someone. The work is piling up, you know exactly what you need help with, and yet the job post has been sitting as an empty document for a week. Every version you start sounds either like a wish list for a superhero or like every other ad on the job board.

Here’s the short answer. To use AI to write a job description, don’t ask it to “write a job description for an office manager”. Give it a ten-minute brief first: what this person must get done in their first 90 days, the five things they truly need, the pay range, how and where they work, and how to apply. Then let ChatGPT, Claude or Gemini turn that into a clear draft, cut the wish list, and check the wording before you post it. The AI writes. You decide what the job really is.

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

Can you use AI to write a job description?

Yes, and it’s one of the most sensible things to hand to AI. A job description is structured writing with a clear goal, and most of the effort goes into getting a messy idea into clean sections. That’s exactly what language models are good at.

Two things stay with you, though. First, the AI doesn’t know your job. Ask for “a marketing assistant” and you get the average of thousands of marketing assistant ads, which is the bland text you were trying to avoid. Second, you’re responsible for what you publish. LinkedIn says it plainly about its own AI job description feature: the draft “may produce surprises or mistakes”, you’re still responsible for the final job description, and the law may require you to add things like a salary range or an applicant privacy notice (LinkedIn Help).

So the deal is simple: you bring the facts, the AI brings the structure and the words.

How to use AI to write a job description in 5 steps

Step 1: Write a ten-minute brief (no AI yet)

This is the step everyone skips, and it’s the one that makes the difference. Open a note, or record a voice memo on your walk, and answer these questions in plain words:

• Why are you hiring? What’s falling through the cracks right now?
• What does success look like after 90 days? Three concrete results, not traits. “Invoices go out by the 3rd of every month” beats “detail-oriented”.
• What does a normal week look like? The real tasks, roughly in order of time spent.
• Must-haves: the few skills or experiences they can’t do the job without. Try to stop at five.
• Nice-to-haves: things you can teach or live without.
• The facts: hours, location or remote/hybrid, start date, pay range, who they report to, how to apply.
• Why would someone good say yes? Be honest: flexible hours, a small team, real responsibility, learning.

If you already have process notes or checklists for the work, paste those in too. If you don’t, our guide on using AI to write SOPs from a voice memo uses the same trick, and the SOPs you get out of it double as onboarding material for the new hire.

Step 2: Turn the brief into a draft with one prompt

Open the AI tool you already use. The free version of ChatGPT, Claude or Gemini is fine here: this is a writing job, not a research job. Paste your brief below this prompt:

The “Missing info” line is the important part. It stops the AI from filling gaps with invented perks, and it shows you what you forgot before a candidate has to ask.

Step 3: Cut the wish list

Your first draft will probably still ask for too much. AI loves to add “excellent communication skills” and “experience with a wide range of tools”. Push back with a follow-up:

Be strict here. A long list of requirements doesn’t find you a better person, it just tells good people they’re not good enough. If a requirement is really “years of experience”, ask yourself what those years should prove, and write that instead. “You’ve run payroll for a small team on your own” says more than “3+ years of experience”.

Step 4: Check the wording before anyone sees it

This is where a quick AI pass saves you real trouble. Words in a job ad can push people away without you meaning to.

• Age and other protected traits. In the US, the EEOC says it’s illegal to publish a job ad that shows a preference for, or discourages, applicants based on things like sex, age (40 or older) or disability. Its own example: an ad asking for “recent college graduates” may discourage people over 40 (EEOC). Phrases like “young team” or “digital native” sit in the same risky corner.
• Gendered wording. A well-known set of studies by Gaucher, Friesen and Kay found that job ads with more masculine-coded words (like “competitive” and “dominant”) made those jobs seem less appealing to women, mostly because they expected to belong less (Journal of Personality and Social Psychology, 2011).
• Pay. If you hire in the EU, new pay transparency rules mean you have to tell job seekers the starting salary or pay range in the vacancy notice or before the interview, and you can no longer ask about their pay history (European Commission). The directive also asks for gender-neutral job titles and vacancy notices (Directive (EU) 2023/970). Each EU country writes its own version into national law, so check the rules where you hire.

Then ask the AI to do a review pass:

Treat the answer as a second pair of eyes, not legal advice. If you hire often, or in a regulated field, have someone who knows employment law in your country read your template once.

Step 5: Make it fit where you post it

One description, a few versions. Ask the AI for a short version (around 80 words) for social posts and a message you can send to your network, and keep the full version for the job board or your careers page.

Posting on LinkedIn? It has a built-in “Draft with AI” button when you post a job (LinkedIn Help). It drafts from your job title, company page and profile, which means it knows even less about the real job than your brief does. Paste your own reviewed text instead, or use their draft only as a starting point and fix it with your brief.

Last step: read the final text out loud once. If you’d be slightly embarrassed to say a sentence to a candidate across the table, cut it.

A quick example: from vague to clear

Say you run a small accounting practice and want a part-time office manager. Here’s a typical first line from a lazy prompt, and the version you get from a real brief.

Lazy prompt result: “We are seeking a highly motivated, detail-oriented Office Manager to join our dynamic team and drive operational excellence.”

From a brief: “You’ll keep our six-person office running so the accountants can focus on clients: sending invoices on time, booking client meetings, and making sure nothing gets lost between email and the client file. It’s 24 hours a week, mornings, in our office.”

Same AI, same five minutes. The difference is that the second one had facts to work with. A candidate reading it knows right away if this is their kind of job, and that’s the whole point of a job description.

Mistakes that make AI job descriptions worse

• Prompting with only a job title. You get the average ad. Always give the brief.
• Letting it invent perks. “Competitive salary and great culture” means nothing, and made-up benefits come back to bite you in the interview. Keep the “use only my facts” line in your prompt.
• Posting without a pay range. Beyond the legal side in some places, it saves everyone from interviews that end at the salary question.
• Pasting candidate data into the chat. For writing the ad you don’t need any personal data, so keep it that way. Screening applicants with AI is a separate, more sensitive job.
• Skipping your own read. AI text tends to sound polished and empty. Your read is what makes it sound like your company.

If the drafts keep sounding stiff, it’s worth trying the same brief in a second tool. Our comparison of ChatGPT vs Claude for writing shows where each one tends to sound more natural.

FAQ

Can I use ChatGPT to write a job description for free?

Yes. A job description is a short writing task, and the free versions of ChatGPT, Claude and Gemini all handle it well. What matters is the brief you give it, not the paid plan. New to the tool? Start with our ChatGPT tutorial for beginners.

What’s a good ChatGPT job description prompt?

One that gives the AI your real facts, a fixed structure and a rule not to invent anything. Use the Step 2 prompt above, and add the Step 3 and Step 4 follow-ups before you post.

Is it OK to use AI to write a job description?

Yes, as long as you check the result. You’re responsible for what you publish, including the accuracy of the role and any legal requirements where you hire, like pay information or a privacy notice.

Will candidates notice it was written with AI?

They notice generic text, whoever wrote it. A description built on your brief, with real tasks and a real pay range, reads like a person who knows the job. That’s what candidates respond to.

Can AI write the interview questions too?

Yes. Once your description is final, paste it in and ask for interview questions that test each must-have, plus what a strong answer would include. Same rule: you check them before you use them.

When the new hire’s first review comes around, the same brief-first approach works there too: see how to use AI to write a performance review.

Once the description is posted, the next job is the interview itself. Here is how to use AI to create interview questions that test each must-have, with a scoring guide you agree on before you meet anyone.

When the right person says yes, the next job is their first 90 days. Here’s how to use AI for employee onboarding, from the manager’s brief to the 30, 60 and 90 day check-ins.

Your next step

Take the role you’ve been putting off, spend ten minutes on the brief from Step 1, and run it through the prompt today. You’ll have a draft you can actually post before lunch. For more small workflows like this, the kind that take an annoying task off your plate, start with the free Launchpad on our home page.

A supplier sends over a 14-page agreement. It is due back Friday, the liability clause looks odd, and you are not a lawyer. You have seen ads for AI contract review software and wonder if one of them would save you the afternoon, or just add another subscription.

Here is the short answer. AI contract review software reads a contract, flags clauses that differ from what you normally accept, and suggests edits. If you sign a few contracts a month, a general assistant like Claude or ChatGPT with a good checklist covers most of what you need. If you sign ten or more a month, or several people negotiate for the same company, a dedicated tool with playbooks starts to pay off. Either way, a person still makes the final call on anything with real money or risk attached.

Below: what these tools actually do, the four kinds on the market, six questions that decide which one fits you, and a 30-minute test you can run on your own contract before paying anything.

What AI contract review software actually does

Strip away the marketing and most tools do four jobs:

• Find and label clauses. Payment terms, liability caps, termination, auto-renewal, confidentiality, governing law.
• Compare them to a standard. That standard is called a playbook: your list of “we accept this, we push back on that, we never sign this.” The tool flags every clause that falls outside it.
• Suggest a redline. Tracked changes in Word with replacement wording, so you can accept, edit or reject each one.
• Summarize and store. A short summary for your manager, plus key dates such as renewal deadlines pulled into a list.

What it does not do is know your business. It does not know that this supplier is your only option, or that you agreed something different on the phone last week. That context is still yours.

The four kinds of AI contract review tools

1. A general AI assistant (Claude, ChatGPT, Gemini)

You upload the contract and ask questions. Claude, for example, accepts PDF and DOCX files and reads text and visuals in PDFs of up to 100 pages (Claude Help Center). There is no built-in playbook, so the quality depends on your prompt and your checklist. Cheapest option, and for many freelancers and small business owners, enough. We walk through the full method in getting the key points out of a contract PDF with Claude.

Best for: a handful of contracts a month, one person reviewing, standard agreements like NDAs, service agreements and supplier terms.

2. AI inside Word

Microsoft’s Copilot in Word can summarize a document, answer questions about it and rewrite selected passages, and on shared documents it shows suggested changes before anything is added (Microsoft Support). It needs an eligible Microsoft 365 subscription or Copilot license. It is a general writing assistant, not a contract specialist, so it will not know your negotiating positions unless you tell it.

Best for: teams that already live in Word and Microsoft 365 and want help reading and rewording, not a full review system.

3. Dedicated contract review add-ins

Tools like Spellbook and LegalOn sit inside Word and are built for this one job: review against a playbook, then redline. LegalOn advertises ready-made playbooks for common agreements such as NDAs and MSAs. Spellbook offers a 7-day free trial and prices by the number of people on the license, with the exact price given after a demo. These are built mainly for lawyers and in-house legal teams.

Best for: a legal team, or a business with a regular flow of contracts and a clear set of positions it wants applied every time.

4. Contract management platforms with AI review

Platforms like Juro and Luminance cover the whole life of a contract: templates, review, signing, storage and renewal tracking, with AI across all of it. Juro’s own pricing page says businesses signing around ten or more contracts a month are the ones that see a return (Juro pricing). Expect a demo and a sales process, not a sign-up button.

Best for: companies where sales, HR, procurement and legal all touch contracts and need one place to keep them.

How to choose: six questions that decide it

1. How many contracts a month? Under ten, start with a general assistant and a checklist. Over ten, test a dedicated tool. This is the single biggest factor.
2. Who reviews them? If it is you alone, a playbook mostly lives in your head and a written checklist does the same job. If five people negotiate for the company, a shared playbook stops each of them accepting different terms.
3. Where do your contracts live? Mostly Word files you redline back and forth? An add-in fits. Mostly signed PDFs you need to search later? A platform with a repository fits.
4. Do you have written positions? Playbook tools are only as good as the playbook. If you cannot yet write down “our liability cap is X” or “we never accept automatic renewal over 12 months,” write that list first. It makes any tool better, including a free one.
5. What happens to your data? Contracts hold names, prices and confidential terms. Read the tool’s data terms before uploading. For example, on Claude’s Free, Pro and Max plans, chats can be used for model training if you allow it in your privacy settings, while Team, Enterprise and API use fall under separate commercial terms that this setting does not cover (Anthropic). The Law Society’s guidance for solicitors makes the same point: know how a tool uses what you put in before you share client data (Law Society).
6. Which languages and countries? If you sign contracts under Dutch, German or UK law, test the tool on one of those, not on the vendor’s demo contract.

The 30-minute test before you pay for anything

Every vendor demo uses a contract that makes the tool look good. Use your own instead.

1. Pick one real contract you already negotiated and know well. You know which clauses were a problem, so you can judge the tool’s answer.
2. Write down the three to five issues you found before you upload anything.
3. Run the same contract through two options: a general assistant with the prompt below, and a free trial of one dedicated tool if your volume justifies it.
4. Score them: did each one find your issues? Did it invent problems that are not there? Did it quote the right clause numbers? Could you use its suggested wording?

If the general assistant finds your issues and the paid tool adds only formatting, you have your answer. If the paid tool catches things the assistant missed, and you review contracts every week, the subscription likely earns its place.

Here is a first-pass prompt to use with Claude or ChatGPT. Upload the contract first, then paste it:

You are helping me do a first-pass review of the attached contract. I am the [customer / supplier / freelancer] in this agreement. I am not a lawyer.

My positions:
- Liability: [e.g. capped at the fees paid in the last 12 months]
- Payment terms: [e.g. 30 days, no longer]
- Termination: [e.g. I can end it with 30 days notice]
- Renewal: [e.g. no automatic renewal longer than 12 months]
- Anything else that matters to me: [list]

1. List every clause that conflicts with my positions. For each one, quote the clause number and the exact wording, explain the risk in one plain sentence, and suggest replacement wording.
2. List any clause that is unusual for this type of agreement, even if it is not on my list.
3. List all dates and deadlines (start, renewal, notice periods).
4. If you are not sure about something, say so. Do not guess. Only quote text that is actually in the contract.

Then check every quote against the contract yourself. AI tools can produce text that sounds right and is not; the Law Society’s guidance calls this hallucination and warns against treating outputs as accurate without checking (Law Society). If a quote does not appear in the document, the answer around it is not reliable either. For contracts longer than about 100 pages, see our guide on analyzing long PDF files with Claude.

A real example of where the line sits

Take a freelance consultant who signs a client agreement every few weeks. The work is the same each time: check payment terms, liability, who owns the deliverables, and notice periods. A checklist of five positions and the prompt above catches the usual problems in about ten minutes, and a dedicated tool would cost more than the time it saves.

Now take a 40-person software company where sales, procurement and HR all send contracts to one in-house lawyer. That lawyer is the bottleneck. A shared playbook that lets sales handle standard NDAs without waiting, and only escalates the odd ones, is exactly what the dedicated tools are built for.

Most readers of this site sit closer to the first case than the second. Start there.

What AI contract review will not do for you

• It will not replace legal advice on high-stakes deals. Large sums, personal guarantees, exclusivity, intellectual property you depend on, or anything you would lose sleep over: get a lawyer to read it. The AI review makes that meeting shorter and cheaper, because you arrive with the issues already listed.
• It will not know what was agreed outside the document. Emails, calls and side letters are invisible to it unless you add them.
• It will not negotiate for you. It suggests wording. Whether to push back, and how hard, depends on the relationship.

If you are not ready to buy software yet, you can still get a plain-English read of a contract at work. Our guide to using AI for legal documents shows a safe way to do it, with copy-ready prompts and a note on when to call a lawyer.

Quick answers

Can I just use ChatGPT or Claude to review a contract?

For a first pass on everyday agreements, yes, as long as you give it your positions, ask it to quote clauses exactly, and check those quotes yourself. Check the data settings of your plan before uploading anything confidential.

What does AI contract review software cost?

General assistants have free and paid plans. Dedicated tools such as Spellbook and LegalOn price per team and ask you to book a demo for a quote, and Spellbook offers a 7-day trial. Check each vendor’s pricing page, since these change often.

Is it accurate enough to rely on?

Rely on it to find things to look at, not to decide for you. Run the 30-minute test above on a contract you know well, and you will see its accuracy on your type of agreement, not someone else’s.

Do I need a playbook to start?

You need a short list of positions, even five lines. That list is your playbook, and it improves every tool on this page.

Your next step

Pick one contract you already signed, write down its three biggest issues, and run the prompt above. In half an hour you will know whether a general assistant is enough for you or whether a dedicated tool is worth a trial. If you want more practical setups like this, from using Claude for daily work to choosing AI invoice processing software, the free Launchpad on robotshelp.me is where we start you off.

You run a small online shop. Orders come in, but so do the carts that never finish, the first-time buyers who never come back, and the newsletter you meant to send three weeks ago. You know email should be doing more of the work. You just do not have an afternoon a week to write it.

Here is the short answer. Email automation for ecommerce means setting up a few emails once, each triggered by something a shopper does (signs up, leaves a checkout, buys, goes quiet), and letting your store send them on its own. Start with five: abandoned checkout, abandoned cart, welcome, post-purchase thank you, and win-back. Use the ready-made templates in your email tool, and use AI (ChatGPT, Claude, or the assistant built into Shopify) to write the copy in your own voice. You can have the first one live today.

Below: the five flows in the order I would build them, the exact settings that stop the classic mistakes, one copy-paste prompt for the writing part, and the legal basics so your emails keep landing in the inbox.

Want one practical AI workflow like this every week?

What email automation for ecommerce actually is

Two kinds of email leave your shop. Campaigns are the ones you write and send by hand to a list: a sale, a new collection. Automations (Klaviyo calls them flows, Shopify calls them marketing automations) are written once and then sent to each shopper at the moment they trigger it.

The trigger is the whole trick. A campaign reaches people whenever you happen to send it. An automation reaches one person right after they did something, while your shop is still on their mind.

And there is a lot to catch. Baymard Institute averaged 50 studies and lands on a documented cart abandonment rate of about 70% (Baymard). Be honest with yourself about that number, though. In Baymard’s own survey, the most common reason was simply “I was just browsing / not ready to buy” (Baymard). An email will not turn every one of those into a sale. It catches the people who meant to buy and got distracted.

Which tool you need (probably the one you have)

• Shopify store: Shopify Messaging, Shopify’s built-in email app, has ready-made automations for abandoned browse, abandoned cart, abandoned checkout, welcome series, thank you, first-purchase upsell and win-back (Shopify Help Center). Every store on the Basic plan and up gets 10,000 free emails a month, then it costs $1 per 1,000 emails, and abandoned checkout emails are always free (Shopify pricing).
• Klaviyo: a dedicated email and SMS tool for shops that want more control. It has a flow library with pre-built abandoned cart, browse abandonment and post-purchase flows (Klaviyo Help Center).
• WooCommerce or another platform: Mailchimp can send abandoned cart emails once your store is connected, including through the Mailchimp for WooCommerce plugin (Mailchimp Help).

My advice: do not switch tools to start. Build the five flows in what you already pay for. Move later, when you know which flows actually earn.

The 5 flows to set up first (in this order)

1. Abandoned checkout

Trigger: someone typed in their email at checkout and left without paying. These people were almost done buying, so this flow usually comes first.

The setting that matters: make sure the flow stops the moment they buy. In Klaviyo you add the profile filter Placed Order zero times since starting this flow. Klaviyo checks it before every message, so a buyer never gets a “you forgot something” email after paying (Klaviyo: abandoned cart flow). You will then see many “skipped” profiles in the report. Klaviyo says that is normal: it means the filter works.

In Shopify the abandoned checkout automation already skips people who completed the order before the email went out (Shopify Help Center).

What to write: the product they left (your tool pulls it in with a photo), one honest line about why it is worth finishing, and a button back to the checkout. Skip the fake countdown timer.

2. Abandoned cart

Trigger: they added something to the cart but never started checkout. Lower intent than flow 1, so softer copy and fewer emails.

The setting that matters: nobody should get both the cart email and the checkout email for the same visit. Klaviyo recommends running them as two separate flows and adding Started Checkout zero times since starting this flow plus Placed Order zero times to the cart flow (Klaviyo: Added to Cart flow). On Shopify, you also need the Klaviyo setting Track behavioral events switched on, or the Added to Cart trigger never fires.

3. Welcome series

Trigger: someone joins your list through a signup form. This is the moment they are most curious about you.

Shopify’s template sends a welcome email with a discount, then two emails that tell your story, then a reminder only if they have not bought yet (Shopify Help Center). That is a sound structure to copy in any tool. You do not have to lead with a discount. A shop that sells hand-poured candles could just as well open with “how we choose our scents” and keep the discount for email three.

Watch out: run one welcome flow, not two. Shopify warns that two active welcome automations send customers twice as many emails (Shopify Help Center).

4. Post-purchase thank you (and the next step)

Trigger: an order is placed. Keep the order confirmation itself as it is, that one is a transactional email. This automation comes a day or more later and is about the relationship: how to use or care for the product, an honest request for a review, and maybe one product that genuinely goes with it.

Shopify has both a thank you automation (first and second purchase) and a first-purchase upsell. Pick one to start, not both at once.

5. Win-back

Trigger: a set amount of time has passed since their last order. You choose the time, and it should match how often people actually need your product. Coffee beans run out in weeks, a winter coat lasts years.

Keep it human: “It has been a while. Here is what is new.” If you sell something people rebuy, a simple reminder that they may be running low often fits better than a discount.

Let AI write the emails (without sounding like everyone else)

Setting up the triggers takes minutes. Writing ten good emails is the part that stalls most shop owners. That is where AI earns its place.

Inside Shopify: in Shopify Messaging, the AI assistant Sidekick can write subject lines and body text. You click the icon in the subject or text field and describe the email (Shopify: generate email text). Shopify adds a fair warning: you are responsible for the accuracy of what you publish, so read every line before it goes out.

In ChatGPT or Claude: you get more control over voice if you give the assistant real material. Paste two or three emails or product descriptions you wrote yourself, then use a prompt like this:

You are helping me write automated emails for my online shop.

About the shop: [what you sell, price range, who buys it]
My writing style: see the 3 examples below. Match the tone, sentence length and words I use. No hype, no exclamation marks in every line.

Write the [abandoned checkout / welcome series / post-purchase / win-back] emails.
For each email give me:
- 3 subject line options under 50 characters
- preview text (one line)
- the body, max 120 words, with ONE button and its label
- when it should send (time after the trigger)

Rules:
- Only use facts I gave you. If you need a fact (shipping time, return policy, discount), write [ASK ME] instead of inventing it.
- No fake urgency, no countdowns, no "only 2 left" unless I say it is true.
- [discount: none / 10% off with WELCOME10 in email 3]

My example texts:
[paste 2-3 of your own emails or product descriptions]

The [ASK ME] line matters most. Without it, an assistant happily invents a “free 30-day returns” policy you do not have. Then do one pass yourself: read each email out loud, check every price, discount and link, and send yourself a test.

Want more prompts for the campaign side? See ChatGPT prompts for marketing campaign ideas. And if replies to these emails pile up in your inbox, these customer support prompts help you answer them faster.

A simple plan for your first week

• Day 1: turn on abandoned checkout with the template text. Send yourself a test by starting a checkout with your own email.
• Day 2: rewrite that email with the prompt above, in your voice.
• Day 3: add the welcome series and check your signup form actually feeds it.
• Day 4: add post-purchase thank you.
• Day 5: add abandoned cart with the “started checkout” filter, then win-back.

After that, leave it alone for a few weeks. Check per flow: how many people entered, how many emails went out, how many orders followed. Change one thing at a time (a subject line, a delay), never everything at once, or you will not know what helped.

Stay legal and out of the spam folder

This is not legal advice, but these basics come straight from the regulators and mailbox providers.

• US (CAN-SPAM): marketing emails need an honest subject line, your valid postal address, and a clear way to opt out. Opt-outs must be honored within 10 business days (FTC compliance guide).
• UK: you may email existing customers about similar products without fresh consent only if you collected their details during a sale and gave them a simple way to opt out then and in every email (the “soft opt-in”). New prospects need consent (ICO). EU countries have similar consent rules, so check yours.
• Who gets what: Shopify automations go only to customers subscribed to marketing by default. For abandoned checkout emails you can switch the recipients to all customers (Shopify Help Center). Check your local rules before you flip that switch.
• Gmail: if you send 5,000 or more messages a day to Gmail addresses, Google requires email authentication, an unsubscribe option that works straight from the inbox on marketing mail, plus a visible unsubscribe link in the message, and a spam rate below 0.3% (Gmail sender guidelines). Most small shops are below that, but the same habits keep you out of spam at any size.

Mistakes that make ecommerce email automation backfire

• No exit filter: the customer pays and still gets “you left something behind” twice. Always stop the flow on purchase.
• Two flows fighting: cart and checkout emails, or two welcome series, landing in the same inbox on the same day.
• A discount in every email: people learn to leave carts on purpose and wait for the discount.
• Unchecked AI copy: invented policies, wrong prices, a product name spelled two ways.
• Set and forget forever: a flow that mentions last year’s product or a closed shipping option. Put a reminder in your calendar to reread every flow each quarter.

Most of these come from the same root: nobody sent themselves a test. Do that for every flow before you switch it on.

When you outgrow the templates

The built-in flows cover most small shops. When you want something custom, for example “tag VIP customers after their third order and email them personally,” Shopify points you to Shopify Flow (Shopify Help Center). For work outside your store, like a follow-up after a wholesale inquiry, a tool such as Zapier works, no programming needed. Start with automating follow-up emails with no programming, or see what the Zapier free plan can already do. For email automation ideas beyond the shop, from inbox sorting to meeting recaps, read these 8 email automation examples.

The same AI that drafts your flows can also help with the pages those emails link to. Here is how to use AI to write product descriptions that stay accurate and sound like your shop.

FAQ

How many emails should an abandoned cart flow have?

Start with the number of emails in your tool’s template. Add more only when the report shows the later emails still bring orders.

Do I need a paid tool for ecommerce email automation?

Not to start. On Shopify, the built-in automations come with 10,000 free emails a month, and abandoned checkout emails do not count toward that. Klaviyo and Mailchimp have their own plans, so check their current pricing pages before you switch.

Can AI send the emails for me?

The sending is done by your email tool, not the AI. The AI writes the copy once, you check it, and the automation sends it every time someone triggers it. That split keeps a person in control of what goes out.

Will customers notice the emails are automated?

Only if they read like a template. Write them in your own voice, sign them with a real name, and let replies reach a real inbox. Then an automated email reads like a note from the shop owner, because it is one.

Get one hands-on AI workflow every week.

Want to turn more of your shop’s busywork into automations like these? Start free on the RobotsHelp Launchpad.

Sources

• Creating and managing marketing automations in Shopify Messaging (Shopify Help Center)
• Shopify Messaging email pricing (Shopify Help Center)
• Automatically generating email subject lines and body text (Shopify Help Center)
• How to create an abandoned cart flow (Klaviyo Help Center)
• How to create an abandoned Added to Cart flow for Shopify (Klaviyo Help Center)
• Create an abandoned cart marketing automation flow (Mailchimp Help)
• Cart abandonment rate statistics (Baymard Institute)
• CAN-SPAM Act: A Compliance Guide for Business (FTC)
• Electronic mail marketing (ICO)
• Email sender guidelines (Gmail Help)

You know what the sheet needs to do. Pull the account manager in next to each order, total the North region for the quarter, flag the big deals. You just do not know the formula, and the last forum thread you opened was from 2014 and used functions your Excel does not even show.

Here is the short answer. Describe your sheet to an AI assistant (the sheet names, the column letters, what is in each column, and your Excel version), say exactly what result you want in which cell, and ask for the formula plus a one-line explanation. Then test it on three rows where you already know the answer before you fill it down. ChatGPT, Claude and Gemini all do this well in a normal chat. If you have Copilot in Excel or Gemini in Google Sheets, you can ask right inside the sheet.

Below is the prompt that gets a working formula on the first try, four real examples you can copy, and the five-minute check that stops a wrong formula from ending up in a report.

Want one practical AI workflow like this every week?

Why the AI gives you a wrong formula (and how to stop it)

Most bad formulas from AI are not the AI being dumb. They come from a vague question. “Write a VLOOKUP for my customer list” leaves the assistant guessing your sheet names, your columns, whether you want an exact match and what should happen when a name is missing. It fills those gaps with assumptions, and one wrong assumption gives you a formula that looks fine and returns the wrong number.

Give it four things and the guessing stops:

• The layout. Sheet names, column letters and headers, where the data starts (row 2 if row 1 holds headers).
• The result. What should appear, in which cell, for one example row. “In F2, the account manager for the customer in B2.”
• The edge cases. What to show when there is no match, a blank cell or a zero.
• Your version. Excel 365, Excel 2019 or Google Sheets. This matters: Microsoft notes that XLOOKUP is not available in Excel 2016 and Excel 2019, so a modern formula can break on an older install (Microsoft Support).

The copy-paste prompt for any Excel formula

Paste this into ChatGPT, Claude or Gemini and fill in the brackets. It works for one-off formulas and for the ones you will fill down a whole column.

I need an Excel formula. I use [Excel 365 / Excel 2019 / Google Sheets].

My data:
Sheet "[name]": row 1 = headers, data starts in row 2.
Column A = [header] ([what it holds, e.g. text, dates, amounts in euros])
Column B = [header] ([...])
(add every column the formula needs)

What I want:
In cell [F2], show [the result] for that row.
Example: for row 2 I expect [the answer you know is right].
If there is no match or the cell is empty, show [blank / 0 / "Not found"].

Give me:
1. The formula for [F2], ready to fill down.
2. One plain sentence per part explaining what it does.
3. Anything in my data that could make it fail (text stored as numbers, extra spaces, merged cells).

The line “for row 2 I expect” is the one people skip, and it does the most work. It gives the assistant a target to check its own answer against, and it gives you a test you can run in ten seconds.

Four formulas people ask AI for most, with working examples

Picture a simple sales sheet. On the sheet Orders: column A Order ID, B Customer, C Region, D Order date, E Amount. On a second sheet Customers: column A Customer, B Account manager. These are the formulas a good prompt gets you.

1. Look up a value from another sheet (XLOOKUP)

Prompt: “In F2 on Orders, show the account manager for the customer in B2, taken from the Customers sheet. Show Not found if the customer is missing.”

=XLOOKUP(B2,Customers!A:A,Customers!B:B,"Not found")

XLOOKUP searches column A of Customers for the name in B2 and returns the matching value from column B. The fourth part is what shows when there is no match; without it you get #N/A (Microsoft Support: XLOOKUP). On Excel 2019 or older, ask for the INDEX and MATCH version instead:

=IFERROR(INDEX(Customers!B:B,MATCH(B2,Customers!A:A,0)),"Not found")

2. Total with conditions (SUMIFS)

Prompt: “Total the Amount column for region North with an order date in the first quarter of 2026.”

=SUMIFS(E:E,C:C,"North",D:D,">="&DATE(2026,1,1),D:D,"<"&DATE(2026,4,1))

SUMIFS adds up column E only where every condition is true: the region is North, the date is on or after 1 January and before 1 April (Microsoft Support: SUMIFS). Using “before 1 April” instead of “up to 31 March” means an order logged on 31 March with a time on it still counts.

3. Flag rows with a rule (IF)

Prompt: “In G2, show Large if the amount in E2 is 5,000 or more, otherwise Standard.”

=IF(E2>=5000,"Large","Standard")

Simple, but this is where AI saves the most time once the rules pile up (“Large over 5,000, Medium over 1,000, flag anything from the West region in red”). Describe the rules in plain words, in order, and let it build the nesting.

4. Count days since a date

Prompt: “In H2, show how many days have passed since the order date in D2.”

=TODAY()-D2

If the result shows as a date instead of a number, the cell has a date format. Set it to Number. That is exactly the kind of small snag worth asking the assistant about, in the same chat.

Check the formula before you trust it

AI formulas fail quietly. They rarely throw an error; they return a plausible number that is wrong. Microsoft says the same about its own assistant: Copilot can generate formulas that are inaccurate, so you need to review and verify them (Copilot in Excel FAQ). This five-minute check catches almost all of it:

1. Test three rows by hand. Pick rows where you know the answer: a normal one, one at an edge (a missing customer, the last day of the quarter), and one with a blank. If all three match, fill it down.
2. Cross-check totals with a filter. For a SUMIFS, filter the sheet on the same conditions and look at the sum in the status bar. Same number? Good.
3. Step through it. In Excel, Formulas > Formula Auditing > Evaluate Formula shows each part being calculated, which makes a wrong range easy to spot (Microsoft Support).
4. Ask for the explanation, then read it. If the explanation says “column C” and your region is in column D, you found the bug before it found you.

When the formula breaks: let the AI fix it

Paste the formula, the exact error and one example row. Do not just say “it doesn’t work”.

This formula in F2 returns #N/A:
=XLOOKUP(B2,Customers!A:A,Customers!B:B)
B2 contains "Jansen BV". On the Customers sheet, A14 also shows "Jansen BV".
Why does it fail, and what is the fixed formula?

The usual causes are boring and very fixable. For #N/A, Microsoft lists the main ones: the value really is not in the lookup list, or the two cells hold different data types, such as a number stored as text on one side (Microsoft Support: correct a #N/A error). Invisible extra spaces at the end of a name cause the same thing; the assistant will often suggest wrapping the lookup in TRIM.

Explain a formula you inherited

The other big use: a colleague left, and their workbook is full of formulas three lines long. Paste one in and ask:

Explain this Excel formula step by step, in plain English, for someone who knows basic Excel.
Then tell me what would make it return a wrong result.
[paste the formula]
Column A holds [...], column B holds [...].

The second question is the useful one. It tells you where the old sheet is fragile before you build next month’s report on it. Copilot in Excel can do the same from inside the cell: select it and ask Copilot to explain the formula (Microsoft Support: Copilot in Excel).

Chat assistant or the AI inside your spreadsheet?

Both write formulas. The difference is what they can see and what you need to pay for.

• A chat assistant (ChatGPT, Claude, Gemini) works with any spreadsheet and any version, including the free plans. It only knows what you describe, so the prompt above matters. If you are new to these tools, start with our ChatGPT tutorial for beginners or the Claude tutorial for beginners.
• Copilot in Excel sees your actual columns, suggests formula columns and explains existing formulas. It needs a Microsoft 365 plan that includes Copilot, and Microsoft says it works on files saved to OneDrive or SharePoint with AutoSave on, with your data set up as a table or a clean range (Copilot in Excel FAQ).
• Gemini in Google Sheets lets you type = in a cell and press Ctrl + Alt + G (Cmd + Ctrl + G on a Mac) to describe the formula you want, and offers a Fix button on cells that show an error. It requires an eligible Google Workspace or Google AI plan (Google Docs Editors Help). Our Gemini tutorial for beginners covers the basics.

If you have the built-in option, use it for formulas on the sheet in front of you. Keep the chat assistant for planning a whole workbook, or when you are on a version or plan without it.

Keep company data out of the chat

A formula never needs your real customer names or revenue. Send the headers and two or three made-up rows with the same shape. The formula works the same on the real data.

If you do paste real data into ChatGPT on a personal Free, Plus or Pro plan, know that sharing data for model training is on by default there, and you can switch off “Improve the model for everyone” under Settings > Data controls. OpenAI says content from Business, Enterprise and Edu workspaces is not used for training by default (OpenAI Help Center: data controls). Your company’s own AI policy comes first either way.

Where this goes next

Once writing formulas stops being the bottleneck, the next step is asking AI what the numbers actually say. Our guide on how to use ChatGPT to analyze a spreadsheet for work picks up right there. And if your sheet is fed by copying numbers out of emails or invoices every week, that part can be automated too: see the n8n workflow for automatic invoice processing.

FAQ

Can ChatGPT write Excel formulas for free?

Yes. Writing and explaining formulas is plain text work, and the free plans of ChatGPT, Claude and Gemini handle it. You only need a paid plan for the in-sheet assistants like Copilot in Excel or Gemini in Sheets.

Does the same formula work in Google Sheets?

Often, but not always. Common functions like SUMIFS, IF and XLOOKUP exist in both, while some functions and date handling differ. Say which one you use in the prompt and you get the right version.

Why does the AI formula work on row 2 but not further down?

Usually a reference that should stay fixed moves when you fill down. Ask the assistant: “Which references in this formula need $ signs so it fills down correctly?”

Is it cheating to use AI for formulas?

No more than looking it up in a forum. The difference is speed. Read the explanation each time and you will find yourself writing the simple ones without help after a few weeks.

Get one hands-on AI workflow every week.

Want to go from one formula to a whole week of saved hours? Start free on the RobotsHelp Launchpad.

Sources

• XLOOKUP function (Microsoft Support)
• SUMIFS function (Microsoft Support)
• How to correct a #N/A error (Microsoft Support)
• Get data insights with Copilot in Excel (Microsoft Support)
• Frequently asked questions about Copilot in Excel (Microsoft Support)
• Collaborate with Gemini in Google Sheets (Google Docs Editors Help)
• Data controls in ChatGPT (OpenAI Help Center)