AI in product management
AI for product managers: where it helps, where it hurts, and how to keep your judgement
Updated
Nearly every product manager now uses AI, and nearly all of them say it makes them faster. Far fewer say it makes their products better. This article looks at why that gap exists, walks through the product workflow step by step, and sets out the habits that keep AI working for your judgement instead of replacing it.
Faster is not the same as better
Product Focus’s 2026 survey of 677 product people in 40 countries found that 69% use AI frequently, up from 49% a year earlier. 97% said it improved their personal productivity, but only 64% said it improved their product outcomes: better decisions, better products or better business results. Product Focus calls the difference the 33-point gap.
The symptoms it lists will sound familiar: AI output that nobody verified, less experienced product managers leaning on AI instead of building judgement, and speed without a clear direction. A faster first draft of the wrong thing is not progress.
Step by step through the product workflow
AI is not equally useful everywhere. A useful test for each task: if the AI is wrong, will anyone notice before it matters?
- Capturing evidence: strong. Transcription and note-taking are accurate and cheap now. Turning raw notes into tidy, searchable insights is a good use, as long as quotes are checked against the source. See using AI to analyse customer interviews.
- Synthesis: useful as a second opinion, risky as the first. Asking AI to “find the themes” across twenty interviews produces confident, generic summaries that lose who said what. Do your own synthesis first, then ask AI what you missed.
- Hypotheses: best as a critic, not an author. AI is good at poking holes: “What would have to be true? Who might this hurt? What is the riskiest assumption?” It is poor at knowing which bet matters to your customers. Write the hypothesis yourself and let AI challenge it.
- Experiment design: good at checklists. AI can spot a missing baseline, a metric that cannot fail, or an obvious guardrail you forgot. It should not choose your target; that is a business decision. See success metrics and guardrails.
- Requirements and specs: good first drafts. Acceptance criteria, edge cases and test ideas are where AI saves real time. Each requirement should still point to the evidence that justifies it, which AI cannot invent. See requirements traceability.
- Communication: strong. Release notes, stakeholder updates and summaries of long threads are low-risk, high-volume work. Check the facts and send.
Five habits that close the gap
- Verify before you rely. Treat AI output as a draft from a fast, confident colleague who was not in the room. Check quotes, numbers and claims against the source before they reach a decision.
- Keep the evidence, not just the conclusion. A summary that says “users want favourites” is worthless six months later. Keep the quotes and data that led there, linked to the decision.
- Use AI to argue with you. Ask it to make the case against your plan, list what would prove you wrong, or play a sceptical stakeholder. This is where AI adds judgement instead of replacing it.
- Give it context, not just a task. A prompt with your customer, your goal, your constraints and an example of good output beats a clever one-liner. Save that context in a project or a reusable document so you are not retyping it.
- Stay in contact with customers yourself. AI can summarise what customers said. It cannot give you the instinct you build by hearing it first-hand, and that instinct is what lets you catch AI when it is wrong.
A prompt pattern that works
Context: We build [product] for [customer]. This quarter's outcome is [goal].
Task: Review this hypothesis as a sceptical senior product manager.
Input: [paste the hypothesis and the insights behind it]
Ask: List the three assumptions most likely to be wrong,
what evidence would test each one cheaply,
and anything in my wording that cannot be proven false.
Rules: Do not rewrite the hypothesis. Quote my text when you criticise it.Mind the data you share
Customer transcripts, support tickets and analytics exports often contain personal data. Before pasting them into an AI tool, check that your company has approved it and what it does with your data. Strip names and contact details first: a role and a participant reference such as P07 is enough for analysis.
Sources
Keep your evidence and decisions linked
Groundwork takes your team from customer insights to evidence-based requirements. Free to start, with your whole team.
Get started