Write an AI product strategy that tells teams what to stop
An AI strategy that lists every opportunity guides nobody. Name where AI changes the customer's outcome most, choose two or three places to win, say what teams stop doing, set how progress will be judged, and state what would make you change course.
LESSON 1 OF 22 · 16 MIN ESTIMATED STUDY
On this page 7 sections
TL;DR: A usable AI strategy fits on two pages and makes choices: the customer outcomes AI should change, the two or three areas where the company will invest to win, what teams stop or decline, the shared foundations everyone uses, how progress is measured each quarter, and the signals that would change the plan. If every team can read it and conclude their current roadmap is fine, it is not a strategy yet.
Where you are. In Set direction. Leadership loops ask how you would set direction for many teams. This lesson builds the document the rest of the course depends on.
Case note. Corvane is fictional; its figures and dialogue are teaching assumptions.
Why most AI strategies fail teams
Corvane's executive team asks its new head of AI product for "our AI strategy". The tempting output is a long list: an assistant in every screen, AI-written job descriptions, candidate matching, payroll anomaly detection, performance review summaries, a chatbot for employees. Each idea is reasonable. Together they say "do everything", which six teams will interpret as "keep doing what you planned, with AI added".
A strategy is useful when it changes what teams do on Monday. That requires choices, and choices mean some teams hear no.
The parts of a strategy that works
Customer outcomes. Start from what Corvane's customers pay for: HR teams want to hire well, pay people correctly and keep good employees. Where can AI change those outcomes measurably? Payroll errors are costly and rare; hiring is slow and expensive; HR teams spend hours answering repeated employee questions.
Where to win. Pick the areas where AI changes outcomes most and where Corvane has an advantage competitors cannot copy quickly. Teaching fixture:
| Area | Customer value | Corvane's advantage | Choice |
|---|---|---|---|
| Employee questions about policy and pay | HR teams spend about 30% of time on repeated questions | Holds each customer's policies and payroll data | Win |
| Payroll anomaly detection | Rare but costly errors | Years of payroll history across customers | Win |
| AI job description writing | Moderate time saving | None; general tools do it well | Stop investing; integrate a basic version |
| Candidate matching | High value, high regulatory risk | Some data, legal exposure in several markets | Decline for now |
What to stop. Say it plainly: "We will not build our own job description writer. We will not ship automated candidate ranking this year." Each stop frees people and names a risk avoided. Candidate matching is a hard call: it may be valuable, but employment decisions are treated as high risk under the EU AI Act, and a later lesson covers how risk tiers decide review effort.
Shared foundations. Evaluation tooling, the policy on customer data, the model access layer and the quality bars. Built once, used by every team. Without them each team reinvents them, slower and inconsistently.
How progress is judged. Quarterly measures tied to outcomes: HR time spent on employee questions, payroll errors caught before payment, share of customers using the assistant weekly. Feature counts stay off the list.
Signals that would change the plan. "If model costs fall enough that the employee assistant costs under 0.5 USD per active employee per month, we include it in every plan instead of selling it as an add-on." "If a competitor ships compliant candidate matching with audited results, we revisit the decline." Writing these down makes the strategy a living document and shows leadership you are watching the right things.
Test the draft with three questions
- Could a team lead change their roadmap from this? Ask two of them to read it and tell you what they would do differently. If neither changes anything, sharpen the choices.
- Does it say no to something someone wants? If not, it has avoided the hard part.
- Will we know in one quarter whether it is working? If the measures only move annually, add leading indicators.
In the interview
Leadership prompts such as "write the AI strategy for this company" or "you have just joined as head of AI product; what do you do in 90 days?" reward this shape. Lead with the customer outcomes, name two or three places to win with the reason, state what you would stop, and end with how you would judge progress and what would change your mind. A list of features, however good, scores as a list.
Do this before moving on
Choose a company you know. Write its AI strategy on one page using the six parts above. Include at least two things the company should stop or decline, each with the reason, and two signals that would change the plan.
Expected reasoning: the stops that read as most credible are areas where the company has no advantage over general tools, or where the risk is high relative to current capability. The signals are usually cost thresholds, regulatory changes or a competitor proving something works.
Go deeper
- The moat ladder helps judge where the company's advantage is real when choosing where to win.
- Write the AI strategy for an online travel company is a full graded answer to a strategy prompt.
- AI roadmaps turns the strategy into a roadmap with capability bets and kill criteria.
- Set direction for a product org new to AI practises the 90-day version of the same work.
Key takeaways
- A strategy that lists everything guides nobody; it must choose.
- Start from customer outcomes, pick two or three places to win, and name what stops.
- Shared foundations prevent six teams reinventing the same things.
- Judge progress quarterly by outcomes, and write down what would change the plan.
Check yourself
Answer out loud first, then reveal. Retrieving it yourself is what makes it stick; reading ours and agreeing does very little.
1Which test best shows a strategy draft is too vague?
2Why name signals that would change the plan?
Sign in to track which lessons you have finished.