How AI Product Managers Use ChatGPT and Other AI Tools to Do the Work of an Entire Product Team — and Where They Still Need Humans?

It’s Monday morning. Your learning app has a problem: lots of new users start a course, but few finish one. By Friday, you need to read 400 support tickets, check three competitors and write a product requirements document (PRD) for a fix. The QA (testing) team also wants a list of edge cases, the odd situations nobody planned for.

That used to keep a whole team busy. Today, a product manager (PM) with ChatGPT can get a first draft of every one of those tasks.

So can one PM with AI really do the work of an entire product team? Not quite. AI is fast at writing drafts. It is much worse at knowing which draft is right.

How do AI product managers use ChatGPT and other AI tools?

They use them to summarize customer feedback, compare competitors, draft PRDs and user stories, test ideas and write test cases. People still decide which problem matters, check the facts, make the trade-offs and own the result.

The research agrees. In a small McKinsey study of 40 PMs, AI cut time on writing-heavy tasks by about 40% (McKinsey, 2024). But only the senior PMs kept their quality. The junior PMs got faster, and their work got worse (McKinsey, May 2024).

Below are five practical workflows with prompts you can copy, the AI tools for product managers worth knowing, and the decisions you should never hand over to AI.

Key takeaways

• AI speeds up content-heavy product work
• Output quality follows input quality. 
• Generated output is not verified insight. 
• Judgment and accountability stay human. 
• Protect customer data. 
• Start small. 

What does an AI product manager actually do?

A product manager decides what to build and why. The job connects four things: a customer problem worth solving, a business goal, what design and engineering can realistically build, and the outcome the team will measure after launch.

Much of that work happens during product discovery: the research and testing that show whether a problem is real and a solution is worth building, before engineers commit time.

The term “AI product manager” is used in two ways:

• A PM who uses AI tools to do product work faster, such as research synthesis, documentation, analysis and prototypes. That is the focus of this article.
• A PM who builds AI-powered products, such as a support chatbot or a recommendation engine. That role adds model evaluation, data quality and responsible-AI decisions.

Many roles blend both. Either way, the core skill is framing a problem clearly enough for a person or a model to work on, then judging the result.

Five AI product management workflows (with ChatGPT prompts you can copy)

To keep things simple, all five examples use one made-up company.

The example: Coursewren is a fictional app that sells online courses for a monthly fee. Lots of new users start a course, but most never finish it. The PM wants to find out why, and fix it.

In each prompt, replace the words in [BRACKETS] with your own details.

Workflow 1: How can AI help with user research and customer feedback?

AI can read hundreds of customer comments in minutes and sort them into groups. Those groups are a starting point, not proof.

Try this prompt:

Here is customer feedback about [YOUR PRODUCT]:

[PASTE TICKETS, REVIEWS OR INTERVIEW NOTES, WITH NAMES REMOVED]

Group the feedback into the top 3 to 5 problems.

For each problem, tell me how many comments mention it

and quote 2 comments word for word.

Only use what is in the feedback. Do not make up quotes or numbers.

If there is not enough feedback to be sure, say so.

What you might get back (example):

• Too busy in week one: 41 tickets and 7 of 12 interviews
• Didn’t know how long courses were: 23 app reviews
• Lost progress after an app update: 9 tickets

Before you use it: Check every quote against the original. When product coach Teresa Torres tested an AI summary of customer interviews, about 30% of the “quotes” were wrong or not in the transcript at all (Product Talk). Nielsen Norman Group also warns that AI can “miss, misinterpret, or even manufacture insights” (NN/g).

Protect your customers: Remove names and emails before you paste anything, and use only AI tools your company allows. Business versions of ChatGPT and Claude don’t train on your data by default (OpenAI, Anthropic), but your company’s rules come first. More on this: Is It Safe to Give AI Access to Your Tools, Data and Systems?

Saves you: hours of reading and tagging. Can’t tell you: which problem actually makes people quit. You still need to test that.

Workflow 2: How can AI help with competitor research?

AI research tools can search the web and build a comparison with links. Every claim still needs a quick check.

Use a tool that shows its sources, such as ChatGPT deep research, Perplexity or Gemini Deep Research.

Try this prompt:

Compare how [COMPETITOR 1], [COMPETITOR 2] and [COMPETITOR 3]

help new users finish their first [COURSE / PROJECT / TASK].

Use only their official websites, help pages and app store pages.

Give me a table with: competitor, what they do, a link to the source,

and the date of that page.

If you can’t find something, write “not found”. Don’t guess.

What you might get back (example, made-up competitors):

• Competitor A asks new users how many hours a week they can study, then builds a plan. Source: their help page.
• Competitor B sends a reminder after two missed days. Source: a blog review, so check it in the app.
• Competitor C: not found.

Before you use it: Click every link and make sure the page really says that. AI can sound sure even when it’s wrong (OpenAI). Better still, sign up and try each competitor’s app yourself.

Saves you: hours of searching. Can’t see: private data, like how many users a competitor keeps.

Workflow 3: Can ChatGPT write a PRD, user stories and acceptance criteria?

Yes, it can write a good first draft. Three terms first:

• PRD (product requirements document): a document that says what problem you’re solving, for whom, what you’ll build and how you’ll measure success.
• User story: a short need written from the user’s side: “As a [user], I want [action], so that [benefit].”
• Acceptance criteria: simple checks that tell the team when a story is done.

This is the most popular use of AI in product teams. In Atlassian’s State of Product 2026 survey, 60% of respondents used it for product documents (Atlassian, PDF).

Try this prompt:

Write a first-draft PRD for [FEATURE] in [PRODUCT].

The problem: [WHAT IS GOING WRONG, AND HOW YOU KNOW]

Who it’s for: [USERS]

Goal: [WHICH NUMBER SHOULD CHANGE]

Limits: [DEADLINE, TEAM SIZE, ANYTHING OUT OF SCOPE]

Include: the problem, the solution, what’s not included,

how we’ll measure success, open questions, and 3 user stories

with acceptance criteria.

Mark anything you assumed with [ASSUMPTION]. Don’t invent numbers; write [TBD].

If something important is missing, ask me before you write.

What you might get back (example, one user story):

• User story: As a new user with a busy week, I want to set a weekly study goal when I sign up, so the app plans lessons that fit my time.
• Acceptance criteria: (1) When I pick 1 to 5 hours a week, I see a weekly plan that matches. (2) If I miss two sessions, I get no more than one reminder a day.

Before you use it: Ask engineers if it can be built in time, designers if the steps make sense, and testers if every check can be tested.

Saves you: the blank page. Can’t do: decide if this is the right problem. A neat PRD for the wrong problem is still wrong.

Workflow 4: How can AI help with brainstorming and prioritization?

AI is good at coming up with options and spotting weak points. It can’t tell you which idea to build. (Prioritization means deciding what the team works on first.)

Try this prompt:

Problem: [PROBLEM]. Goal: [GOAL]. Limits: [TIME, TEAM, BUDGET].

Give me 6 different ideas to solve this. For each one, tell me:

– what has to be true for it to work

– the biggest risk

– the cheapest test we could run in two weeks

Don’t rank the ideas unless you can explain why using the facts I gave you.

What you might get back (example):

• Idea: let users set a weekly study goal. It works only if people quit because they’re busy. Quick test: show it to 10% of new users.
• Idea: make the first lesson just 10 minutes. It works only if an early win keeps people going. Quick test: give half of new users the short lesson.

Before you use it: Time and effort estimates must come from your engineers, not AI. Only you know the budget, the strategy and what your boss already promised. In Atlassian’s survey, planning and prioritization were among the least common uses of AI (Atlassian, PDF).

Saves you: a stuck brainstorm. Can’t do: pick the winner. Real tests and real data do that.

Workflow 5: Can AI generate test cases and edge cases?

Yes, as a first draft. An edge case is an unusual situation, like a user with no internet or a phone with notifications turned off. AI is good at thinking of these.

Try this prompt:

Here is a user story and its acceptance criteria:

[PASTE]

The app runs on [iPhone / Android / web].

List test cases in a table: what we’re testing, the steps,

and what should happen.

Include unusual situations, like no internet, notifications off

or wrong inputs.

Then list any questions the story doesn’t answer.

What you might get back (example):

• Test: a user picks the maximum goal of 5 hours. Should happen: the plan shows 5 hours of lessons.
• Test: notifications are off and the user misses two sessions. Should happen: the story doesn’t say. Question for the team.

That second test is the real win: AI found a gap the PRD missed.

Before you use it: Testers and engineers must check every result. In one study, developers rated about 87% of ChatGPT’s test cases as valid, and 15% of those covered situations they hadn’t thought of. The rest didn’t apply or were repeats (Bhatia et al., arXiv). For more testing tools, see Top AI Tools Every Software Tester Should Try in 2026.

Saves you: hours of writing tests. Can’t do: know how your system really behaves.

Which AI tools for product managers are worth using in 2026?

There is no single best AI tool for product managers. General-purpose assistants (ChatGPT, Claude, Gemini) handle most drafting and analysis. Specialized tools win where your team’s data already lives: specs in Notion, tickets in Jira, research in Dovetail, designs in Figma.

Features below come from official help pages, checked October 2026. Plans, limits and regional availability change often.

ToolBest PM tasksRelevant strengthsWatch out forChatGPTPRDs, user stories, analyzing CSV exports, competitor researchProjects keep files and instructions together; deep research produces cited reports; data analysis runs Python on spreadsheetsAvailability depends on plan and country; review generated code and assumptionsClaudeLong documents, PRD reviews, research synthesisProjects with uploaded knowledge; Research across the web and connected apps; creates Word, Excel and PowerPoint filesResearch is on paid plans and uses limits faster; project sharing is Team and Enterprise onlyGeminiTeams working in Google Docs and SheetsDeep Research with an editable plan; Gemini Notebook (formerly NotebookLM) grounds answers in your uploaded sources, with citationsDocs and Sheets features need an eligible Workspace or Google AI planPerplexityQuick market and competitor questionsReal-time web answers with numbered citations; Research mode reportsA citation still needs a click to confirm it supports the claimNotion AISpecs, meeting notes, workspace Q&AAnswers across your workspace and connected apps such as Slack and Google Drive; meeting notes with action itemsFull AI is on Business and Enterprise plansFigma Make and Figma AITurning a PRD into a clickable prototypePrompt-to-prototype that can be published to a URLMake is for Full seats on paid plans; a prototype is not a tested designJira with RovoBreaking large features (epics) into smaller work items, summarizing ticketsSuggests child work items from a parent item; summarizes status, blockers and commentsStandard, Premium and Enterprise plans; review suggested items before acceptingDovetailResearch repositories, feedback at scaleClusters interview highlights into themes; classifies tickets and reviewsHighlight clustering is Enterprise only

What should product managers never delegate entirely to AI?

Let AI write the drafts. Keep the decisions for people. People choose the problem, check the facts and make the final call, because they are the ones who answer for it. The biggest danger is when AI is wrong but sounds right.

A large study shows why. Researchers gave 758 Boston Consulting Group consultants real work tasks. On tasks AI handles well, the people using GPT-4 finished 12.2% more work and were about 25% faster. But on a task that needed careful reading of interview notes and data, they got the right answer less often: 70.6% of the time, compared with 84.5% for people working without AI (Harvard Business School AI Institute; published in Organization Science, 2026). Worse, their wrong answers were well written, so they looked convincing (Ethan Mollick, One Useful Thing). The study was about consultants, but PMs do this kind of work every day.

Here are eight things to keep in human hands, with Coursewren examples:

1. Understanding real customers. AI can read 400 tickets. It can’t see a night-shift nurse trying to finish a lesson in a 15-minute break, or notice her pause before she taps “skip.”

2. Picking the right problem. AI counts how often a complaint comes up. The “lost progress” bug had only 9 tickets. But if it hits paying users right when they are most motivated, it may matter more than the most common complaint.

3. Balancing users and business. Making every course free might help more people finish, but it would also cut revenue. The people who own the results should make that call.

4. Sorting out mixed signals. Interviews say courses feel too long. The data says people quit at lesson 2, no matter how long the course is. A person has to decide what to believe and what to test next.

5. Questioning AI’s guesses. The brainstorm assumed people quit because they’re busy. If people with free evenings quit just as often, that guess is wrong. Always ask: what would prove this AI answer wrong?

6. Working with the team. Only engineers know if reminders can follow each user’s time zone. Only the support team knows which tickets were really refund requests. Agreeing on what to build, and earning trust, is human work.

7. Privacy, fairness and safety. Should the app use learners’ behavior to personalize lessons? Are some users under 18? Will a reminder feature leave out people with unusual work hours? These need human judgment, and often a policy owner.

8. Owning the result. AI can suggest a plan. It can’t be responsible for what happens after launch. If completion drops or a customer complains about privacy, the PM answers for it.

How can beginners use AI in product management? A 7-step workflow

If you are new to AI in product work, use the same seven steps for any task. Here they are applied to Coursewren’s completion problem.

1. Define the problem or question. “Why do new subscribers stop before finishing their first course?” One question per session keeps the output focused.
2. Gather relevant, trustworthy context. Ticket exports, interview transcripts, the completion funnel and last quarter’s goals. Remove personal data first.
3. Choose the right tool. An approved general assistant for synthesis, a cited research mode for competitors, Jira or your test tool for QA work.
4. Give clear instructions and inputs. Paste the cleaned tickets and transcripts into the Workflow 1 prompt, which states the task, format, constraints and what to do when information is missing.
5. Review the output. Spot-check the quotes behind “too busy in week one,” and look for invented numbers, bias, missing segments such as web-only users, and assumptions stated as facts.
6. Validate with evidence and teammates. Check quotes against transcripts, competitor claims against live pages, and estimates with engineering. Run the cheapest experiment that could prove the idea wrong.
7. Decide and document. Write down what you decided, the evidence behind it and what would change your mind. The PRD or decision log is where that lives.

If step 5 or 6 fails, go back to step 2.

Real questions from product communities

These are real questions people asked on Blind and Hacker News, copied word for word (typos included).

“AI basically doing my PM job now and idk if I should be happy or terrified”

Blind, January 2026

Be happy, but don’t relax. AI now writes docs fast and cheaply. It can’t decide what to build, handle office politics, help stuck engineers or test ideas with real users. Spend the time AI saves on that work.

“What are we even doing here?”

Hacker News, June 2026, asked after a coworker spent 3 days writing a prompt to get Claude to write a PRD

AI can move work from the writer to the reader. A spec that takes minutes to create can take hours to review. Think first, agree on the direction, then ask AI for a short draft. A short, correct PRD beats a long, polished one.

“do people feel comfortable talking when they know the other end is an AI?”

Hacker News, October 2024

Not always. The person who replied said they wouldn’t take part. Research also shows that when AI pretends to be users, its answers are far less varied than real people’s (NN/g). Let people run the interviews. Use AI to prepare questions and sum up the notes afterward.

“Tried using chatgpt but the results are not good…”

Blind, April 2025, from a post about writing test cases from a PRD

Bad test cases usually come from vague input. Give AI one user story at a time, with its acceptance criteria, and use the Workflow 5 prompt. Jira users can try Rovo, but treat it as a helper, not a full testing tool (Atlassian Community). Testers still check every case.

“Can my company track if I accidently share confidential information with ChatGPT or Claude through Zscaler?”

Blind, October 2026

Assume yes. Replies pointed out that company security tools can log or flag what you type into AI. The better question is what you’re allowed to share. Use only approved tools, remove private details first, and if you already shared something sensitive, tell your security team.

Conclusion

AI is unlikely to replace a good product manager, but it can make a good one much faster. Let it write drafts, sort feedback and list test cases. Keep the big calls for yourself: which problem to solve, what the evidence really says and what to build.

Start small. Pick one task this week, try it with AI, and compare the result with your own work.

If you want structured, guided practice with these skills, explore Syntax Technologies’ AI Product Management Fellowship.

Frequently asked questions

How do product managers use ChatGPT?

To summarize customer feedback, draft PRDs and user stories, compare competitors, brainstorm ideas and write test cases. Treat ChatGPT for product management as a fast first-draft partner, not a source of truth.

What are the best AI tools for product managers?

There’s no single best tool. Start with ChatGPT, Claude or Gemini. Then add tools where your work already lives, like Notion AI, Jira with Rovo or Figma Make.

Can ChatGPT write a PRD?

Yes, a good first draft, if you give it a clear problem, users, goals and limits. Engineers, designers and testers should review it before anyone builds.

Can AI replace product managers?

Not today. AI handles more of the writing, but deciding what to build and owning the result still need people.

Can ChatGPT be used for user research?

Yes, for preparing interview questions and summarizing notes. It can’t replace talking to real users, and you must check its quotes. Remove personal data first.

What skills does an AI product manager need?

Core product skills first: understanding users, setting priorities and communicating clearly. Then writing good prompts, reading data and spotting when AI is wrong.