Using Claude as a PO: where it helps, what you still do

The user story looks great

You have just been asked to write user stories for a returns feature. Instead of starting from scratch, you ask Claude to draft them. In under a minute you have a neat draft: the right “As a customer, I want…” format, numbered acceptance criteria, smooth wording. You send it to the team.

Two days later an engineer asks: “Our return policy is 7 days. Why does the story say 30?” QA asks how many seconds “the system processes requests quickly” means. Customer support asks how refunds work for orders that used a discount code. Nobody on the team wrote 30 days. The AI filled the gap on its own.

This is an illustrative situation, but it is an easy mistake to make. This article does not teach prompts. It covers the harder part: where AI helps in PO work, where you still have to do the work yourself, and how to check a document before it reaches the team.

Why a tidy draft hides errors

AI tools such as Claude write fluently. That fluency lowers the reader's guard. A wrong paragraph that is well presented slips past more easily than a correct paragraph that is messy. NIST also names over-reliance on machine output as a risk.

The behaviour has a name. Anthropic, the company that makes Claude, calls it “hallucination” and states in its documentation that mitigation techniques reduce it but do not eliminate it. The US National Institute of Standards and Technology (NIST) lists it among generative AI risks as “confabulation”: confidently producing false or made-up content.

For a PO, the damage is not in the wording. It is that the team builds exactly what is written. An invented business rule that slips through becomes code, then a screen, then a customer complaint.

What AI can and cannot do for you

The simplest way to think about it: AI is good at drafting, organising and suggesting. The PO owns the facts, the trade-offs and the decision.

In every task, AI has a part it can help with. The difference is who holds the right to say “this is correct, send it”. That right always belongs to the PO.

The rule: AI drafts, the PO checks, the PO decides

Whether you use Claude or another AI tool, one rule works for every product document.

Step 1 matters more than it looks. AI has less to guess when you already hold confirmed facts: how many days the return policy allows, who can approve refunds, how the system handles orders today. You still have to check. And if you do not know yet, the AI fills the gap with whatever sounds most plausible, not with what is true.

Four mistakes AI makes in product documents

These are the first things to look for when reading a draft written with AI.

There is a fifth, harder to see: using the wrong terms for your team. AI uses the words common in the market, while your team may use another name for the same thing, or one word for two different things. For example, this team treats “refund” and “return” as separate flows, but the draft mixes them up.

Example: fixing a returns user story

Back to the opening situation. The table shows an illustrative draft and what the PO finds when comparing it with the policy of a fictional shop.

Nothing in the “After the fix” part came from the AI's own knowledge. The 7 days, the 5-minute target and the discount rule all come from the policy and from talking to operations. AI gets you a fast draft. Making the draft correct is still the PO's job.

A checklist before sending to the team

Use this checklist to reread any document AI helped draft, before you send it. Copy it into your notes or your team's ticket template.

Before sending this document to the team:
[ ] Every number, deadline and permission points to a source (policy, data, a person who confirmed it).
[ ] Each main flow has at least one error case.
[ ] Every acceptance criterion is testable: a number, a condition or a visible result.
[ ] Terms match what the team actually calls things.
[ ] There is no point I could not explain again in my own words.
[ ] It contains no customer data or internal information the company has not allowed in AI tools.
[ ] I am ready to own this document as if I had written every word.

The “explain it in my own words” item is the one most often skipped. If an engineer asks “why is it like this?” and the only answer is “the AI wrote it”, the document is not ready.

When not to use AI

• When the document contains customer personal data, contracts or internal figures, and the company has no policy allowing it. Ask about the policy first; do not decide alone. NIST also lists data privacy as one of 12 generative AI risks.
• When you do not yet understand the problem. AI will give you an answer that looks complete, and you lose the chance to ask the right questions yourself.
• When a priority decision is needed. AI can list criteria, but the trade-offs between people and the responsibility for the result belong to the PO.
• When a hard conversation is needed. A draft helps you organise your points, but a refusal or an apology should sound like you.

Mistakes new POs make with AI

• Sending the draft as it is. Not rereading each line because it “looks fine”.
• Asking the AI instead of the person who knows. Questions about real policies, systems or customers go to whoever owns them.
• Letting the AI decide. Using an AI ranking as the backlog order without checking each reason.
• Hiding that you use AI. If the company has a policy, follow it. If not, ask your manager how the team handles it.

Summary

• AI writes fluently, so errors in a draft are harder to see. Read the parts that look best most carefully.
• AI drafts, the PO checks, the PO decides. The right to say “send it” always belongs to the PO.
• Check four mistakes first: invented rules, missing error cases, untestable criteria, confidence without data.
• Do not put sensitive data into AI tools unless a policy allows it.

This article is not affiliated with Anthropic. Claude is a trademark of Anthropic, PBC.

I am writing more PO situations in this same format, with templates you can use straight away. If you would like the next one by email, subscribe to the newsletter below.

• Get new articles by email

Frequently asked questions

Should a PO use Claude or other AI to write user stories?

AI can draft, but the PO must check every rule, number and criterion before sending it to the team. AI drafts, the PO checks, the PO decides.

What mistakes does AI make in product documents?

The common four are invented business rules, missing error cases, untestable criteria, and confident conclusions without enough data.

Should I put customer data into AI tools?

Only if your company has a policy allowing it. If unsure, ask about the policy first and keep personal data, contracts and internal figures out.

Sources to check

• Anthropic: Reduce hallucinations
• NIST AI 600-1: Generative AI Profile
• Atlassian: Acceptance criteria

Related articles

• Can AI replace a Product Owner? What still needs a PO: A balanced answer for new POs: AI helps with drafting, summarising and suggesting, while accountability for product value and the decisions behind it stays with a person.
• What data should you keep out of AI tools at work?: Six kinds of data a PO should keep out of AI tools, how to filter each piece before pasting, an anonymised complaint example and a ready-to-use checklist.
• Using AI to summarise customer feedback: what to check: AI tools group feedback and name themes fast, but they can invent quotes, lose frequency and drop small groups of customers. This article shows each failure mode and how a PO checks a summary before using it for a decision.
• Numbers from AI analysis: what to check before reporting: How a PO checks numbers an AI tool calculated: reconcile totals, check the denominator and the period, recompute a few key figures, and label what is still uncertain.