AI for product managers: research, specs and judgement

How can AI be used in product management?

AI is most useful to a product manager as a fast, tireless reader and a first-draft writer. It can process a pile of notes, group feedback and produce a spec skeleton in minutes. It is weakest at the things that make a product manager valuable: knowing which evidence is representative, which customer matters to the strategy and which trade-off the business can live with.

These are tasks a general assistant can help with, using material your company allows you to share:

What goes wrong when AI summarises customer interviews?

The usual problem is overstatement. A model asked for 'key themes' presents a point two people made as what 'users want', smooths away disagreement and can produce a quotation that nobody said. A product decision built on that summary rests on evidence that does not exist.

A fictional example. Ledgerly is a small invoicing app. The product manager has notes from six customer interviews. Two customers asked for automatic payment reminders. One said reminders would annoy their clients and they would switch them off. Three did not mention reminders and talked mostly about slow bank reconciliation.

A loose prompt, 'Summarise the key themes from these interviews', produced: 'Users strongly want automated payment reminders. As one customer put it, “I would pay extra for reminders.” Reconciliation is a secondary concern.' Nobody said the quoted sentence. The theme with most support, reconciliation, was demoted. The objection disappeared.

A better prompt: 'For each theme, give the number of interviews that mention it, the interview IDs, and one exact quote copied from the notes. List any interview that contradicts the theme. Do not rank themes by importance.' The checked summary then read: reconciliation, raised in 3 of 6 interviews (I2, I4, I6); payment reminders, requested in 2 of 6 (I1, I3), with one objection (I5). The product manager verified each quote against the notes before sharing the summary.

How do you write a better spec with AI?

Use AI to find what the spec is missing rather than to write the spec for you. The product manager who writes the problem statement and goals understands them; one who accepts a generated version often cannot defend it in review.

Continuing the Ledgerly example, suppose the team does build reminders. The product manager writes a one-paragraph description and asks: 'List edge cases this feature must handle. Group them by billing, timing and the client's experience.' A useful list includes partially paid invoices, invoices disputed by the client, clients in a different time zone, a reminder scheduled for a public holiday, a client who has asked not to be contacted, and an invoice edited after a reminder was sent.

The product manager then decides which edge cases are in scope for the first release, which are explicit non-goals and which need an engineering opinion. That decision, and the reasoning behind it, is the part of the spec that matters most, and it is not something to delegate.

Can AI do product design, or a product manager's job?

AI can help with parts of both. It can draft interface copy, suggest layout options, produce rough prototypes in some tools and generate many variations quickly. That is useful for exploring, and a designer or product manager still has to judge which option fits the users, the brand and the constraints.

The product manager's core work is choosing: which problem, which customers, which trade-off, which thing not to build. Those choices depend on context a model does not have, such as the company's strategy, its commitments to particular customers and what the engineering team can sustain. We do not make predictions about which roles will change and how fast.

Is an AI product manager the same as a product manager using AI?

No. An AI product manager usually means someone who builds AI features: defining how a model should behave, designing evaluations, handling failure cases and managing the cost of inference. That is a specialist area with its own courses and hiring market.

This guide is about the much larger group of product managers who use AI tools to do their existing job better. If you are building AI features, you need both: the everyday skills on this page and a separate grounding in model evaluation and behaviour, which this page does not cover.

What product evidence should stay out of AI tools?

Anything your company has not approved for the tool you are using. Interview notes often contain names, employers and details that identify people; remove them before pasting, and replace them with interview IDs. Unreleased roadmap items, pricing plans and contract terms may be restricted by policy.

Use the company's approved workspace rather than a personal account, and check your organisation's AI policy before uploading research recordings or transcripts. Where personal data is involved, the organisation has data protection duties regardless of which tool is used.

How should a product team practise using AI?

Practise on a past decision whose evidence and outcome you already know.

When is this approach not the right answer?

If your research is a handful of interviews, read them yourself. The time saved by AI synthesis is small, and reading the notes gives you context a summary cannot. AI earns its place when there is more material than one person can read carefully, and even then it should point you to what to read rather than replace reading.

Tosheo.ai is not a course in AI product management or model evaluation. It is being built to teach people to use general AI tools in their day-to-day work, with a planned Product role path, short lessons and practice tasks graded against a visible rubric. If you need a career programme in building AI products, a specialist course is the better choice.