---
title: "How to Build an AI Product Strategy"
description: "How to build an AI product strategy framework in five decisions: the user problem, an AI fit test, your right to win, unit economics and the bets you sequence."
canonical_url: "https://builderscamp.com/guides/other/ai-product-strategy"
date_published: "2026-10-26"
date_modified: "2026-10-26"
author: "Joy Chatterjee"
publisher: "Builders Camp"
guide_class: "other"
---

# How to build an AI product strategy framework

**TL;DR:** Build an AI product strategy framework from five decisions: the user problem, whether AI is the right tool for it, where your right to win comes from once models are commodities, the cost per successful outcome, and a sequence of bets each with a quality bar and a kill criterion. Write all five on one page, and if AI fails the fit test for a problem, the right strategy for that problem is to ship without it.

An AI product strategy is a normal product strategy with three extra decisions bolted on: whether AI fits the problem, where you win when everyone has the same models, and what each successful output costs. The first decision is the one teams skip. Martin Zinkevich's [Rules of Machine Learning](https://developers.google.com/machine-learning/guides/rules-of-ml) at Google opens with Rule #1: "Don't be afraid to launch a product without machine learning." His reason comes right after: "Machine learning is cool, but it requires data."

The other two extra decisions exist because the model layer moves under you. The [Stanford HAI 2025 AI Index](https://hai.stanford.edu/ai-index/2025-ai-index-report) reports that the inference cost for a system performing at the level of GPT-3.5 dropped over 280-fold between November 2022 and October 2024. A strategy whose only advantage is access to a good model loses that advantage as fast as prices fall.

A strategy still has to answer the plain questions first. Marty Cagan's [Product Strategy Overview](https://www.svpg.com/product-strategy-overview/) frames the job as answering "how do we make the product vision a reality, while meeting the needs of the company as we go?" If you have no vision above the strategy yet, start with [product vision vs strategy vs roadmap](https://builderscamp.com/guides/other/product-vision-vs-strategy-vs-roadmap), then come back.

## What are the five decisions in an AI product strategy framework?

The five decisions are the problem, the AI fit, the right to win, the economics and the bets, made in that order because each one constrains the next.

| Decision | The question it answers | Evidence that settles it |
|---|---|---|
| 1. Problem and job | Which user, doing which job, wants which outcome? | Interviews, support tickets, usage data |
| 2. AI fit | Does AI solve this better than rules, search or a manual flow? | A baseline built without AI, plus a prototype with it |
| 3. Right to win | Why will users choose you when competitors use the same models? | Workflow position, distribution, unique context or data |
| 4. Economics | Does each successful outcome cost less than it is worth? | Cost per successful outcome against price or value |
| 5. Bets and quality bar | What ships first, how good must it be, and when do you stop? | Eval pass rates, kill criteria, staged rollout plan |

Each row ends in a written choice, not a theme. Exploring AI in onboarding is a theme. Replacing the setup wizard with an assistant for new admins, shipped only if it completes setup at least as often as the wizard, is a decision you can be wrong about, which is the point.

## How do you write the problem statement before any model choice?

Write one sentence that names the user, the job and the outcome, with no mention of AI in it. Google's [People + AI Guidebook chapter on user needs](https://pair.withgoogle.com/chapter/user-needs/) puts the reframe directly: instead of asking whether you can use AI for something, ask "How might we solve" the problem, and only then ask "Can AI solve this problem in a unique way?"

The sentence looks like this: for support leads at mid-size software companies, triaging the morning queue, we want urgent tickets in front of the right agent before the customer follows up. Nothing in that sentence commits you to a model. If you cannot write it without the word AI, you have a technology looking for a job, and the [jobs to be done](https://builderscamp.com/guides/glossary/jobs-to-be-done) glossary entry is the better place to start.

## How do you test whether AI is the right tool for the problem?

Test AI fit by checking the problem against known good and bad matches, then by building the cheapest non-AI baseline and asking how far it gets. The same [PAIR chapter](https://pair.withgoogle.com/chapter/user-needs/) lists cases where AI is likely the better choice, such as "Recommending different content to different users" and "Natural language understanding", and cases where it is likely not, such as "Maintaining predictability", "Minimizing costly errors" and "Optimizing for high speed and low cost".

The baseline matters more than the list. Zinkevich's [Rules of Machine Learning](https://developers.google.com/machine-learning/guides/rules-of-ml) warns that borrowing data from another problem to start a model "will likely underperform basic heuristics". Build the rule, the search box or the template first, measure it, and make AI beat it on the outcome from decision 1. If the baseline is good enough, ship the baseline; that is a strategic result, not a failure of ambition. The [AI-native vs AI-enabled](https://builderscamp.com/guides/other/ai-native-vs-ai-infused-products) guide covers the related call of whether AI is the product or a feature inside it.

## Where does an AI product's right to win come from when everyone has the same models?

The right to win comes from what sits around the model: your position in the user's workflow, your distribution, the context only you can feed the model, and how completely you solve the job end to end. The model itself is a weak answer, because the gap between models keeps closing.

| Model-layer trend | Earlier figure | Later figure | Source |
|---|---|---|---|
| Inference cost at GPT-3.5 level | Nov 2022 baseline | over 280-fold cheaper by Oct 2024 | [Stanford HAI AI Index 2025](https://hai.stanford.edu/ai-index/2025-ai-index-report) |
| Score gap, top vs 10th-ranked model | 11.9% | 5.4% a year later | [Stanford HAI AI Index 2025](https://hai.stanford.edu/ai-index/2025-ai-index-report) |
| Open-weight vs closed model gap, some benchmarks | 8% | 1.7% in a single year | [Stanford HAI AI Index 2025](https://hai.stanford.edu/ai-index/2025-ai-index-report) |

Those figures come from public benchmarks, not from your users' tasks, so they say nothing about whether a cheaper model is good enough for your product. What they do show is direction: an advantage built on having access to the best model has a short shelf life.

Proprietary data is the usual fallback answer, and it is weaker than it sounds. Martin Casado and Peter Lauten argue in [The Empty Promise of Data Moats](https://a16z.com/the-empty-promise-of-data-moats/) that there is generally not "an inherent network effect that comes from merely having more data", and conclude that "defensibility is not inherent to data itself". Write your right-to-win line as a mechanism a competitor would have to copy: we sit inside the ticketing tool agents already use, we see the full account history, and we close the loop by routing the ticket, not just labelling it.

## How do the economics of an AI product change the strategy?

The economics change because every successful output has a running cost, so margin is a strategy decision rather than an accounting detail. Martin Casado and Matt Bornstein wrote in [The New Business of AI](https://a16z.com/the-new-business-of-ai-and-how-its-different-from-traditional-software/) that, anecdotally, the financial data of AI companies they had seen showed "gross margins often in the 50-60% range", below the 60-80%+ benchmark they gave for comparable SaaS businesses.

That essay dates from February 2020 and from the companies a16z saw at the time, and inference prices have fallen sharply since, so treat the figures as a warning about shape rather than a forecast of your margin. The shape still holds: cost scales with usage, retries and context length. Put one number in the strategy, the cost per successful outcome, and the price or value it has to stay under. The [LLM cost estimation guide](https://builderscamp.com/guides/other/how-to-estimate-llm-costs-for-a-feature) shows how to build that number from tokens and call volume.

## How do you turn the strategy into bets you can sequence?

Turn the strategy into a short list of bets, each with an outcome metric, a quality bar the AI path must clear, and a kill criterion written before work starts. Order them by how much each one teaches you about the riskiest of the first four decisions, not by how impressive the demo will be.

The quality bar is where AI strategy and delivery meet. A bet with no eval set has no definition of good, so the first deliverable of most AI bets is the set of test cases, built as described in [how to write evals for AI products](https://builderscamp.com/guides/tools/how-to-write-evals-for-ai-products). The kill criterion covers the other failure: if the AI path cannot beat the baseline from decision 2 on the outcome from decision 1 within the time you set, the bet ends and the baseline ships. Launch gates come later and live in the [AI feature launch checklist](https://builderscamp.com/guides/other/ai-feature-launch-checklist).

## What does a one-page AI product strategy look like?

A one-page AI product strategy is five filled-in lines plus the evidence behind each, and the hypothetical support-triage team from the problem section would fill it in like this.

| Line | Hypothetical support-triage example |
|---|---|
| Problem and job | Support leads triaging the morning queue want urgent tickets with the right agent before the customer follows up |
| AI fit | Rules on keywords and plan tier are the baseline; AI is tested on reading free-text urgency the rules miss |
| Right to win | Runs inside the existing ticketing workflow, uses full account history, routes instead of only labelling |
| Economics | Cost per correctly routed ticket stays below the agent time it saves, checked monthly against model prices |
| Bets and quality bar | Bet 1: urgency classifier on one queue, must beat the rules baseline on missed urgent tickets, stopped if it does not by the agreed date |

The example is illustrative, not a case study, and it has no numbers on purpose: your thresholds come from your own baseline and cost model. The useful test is whether a new engineer could read the page and say what the team will not build this quarter.

## What is the strongest argument against having an AI product strategy at all?

The strongest argument is that the technology moves too fast for a strategy to survive, so teams should run experiments and let strategy emerge. The objection is right about the model layer and wrong about the rest. Prices and capabilities will change within a year, which is exactly why the right-to-win and economics lines have to be revisited on a schedule.

The user problem, the job and your position in the workflow change slowly. A team without written decisions on those still runs experiments, but it cannot tell a failed experiment from a wrong bet, and it ships demos that nobody asked for. Keep the strategy to one page, date it, and put the next review on the calendar when you publish it.

## Where can you practise building an AI product strategy?

The five decisions overlap with the [guide to building an AI product from scratch](https://builderscamp.com/guides/tools/how-to-build-an-ai-product), which picks up where strategy ends and covers the build itself. To measure whether a bet worked once it ships, use [how to measure AI feature success](https://builderscamp.com/guides/other/how-to-measure-ai-feature-success).

Builders Camp's AI Product Management bootcamp runs over 2 weeks with 4 live sessions and 7 self-paced microlessons, and covers how to design AI-native experiences, evaluate model behavior, manage risk, and deliver AI features users trust. The bootcamp is part of the AI Product Expert Track.

[See the AI Product Management bootcamp](https://builderscamp.com/bootcamps/ai-product-management?utm_source=guide&utm_medium=organic&utm_campaign=ai-product-strategy)

## Frequently asked questions

### What is an AI product strategy framework?

An AI product strategy framework is a short, repeatable set of decisions that turns a user problem into AI bets you can defend. The version on this page has five: the user problem and job, whether AI is the right tool for it, where your right to win comes from once models are commodities, the unit economics per successful outcome, and the sequence of bets with a quality bar and kill criteria for each.

### How is an AI product strategy different from a normal product strategy?

The core questions are the same: who it is for, what problem it solves and why you will win. Three things change. You have to test whether AI fits the problem at all, the model layer you build on gets cheaper and more interchangeable every year, and every successful output has a running cost that a normal software feature does not have.

### Should an AI product strategy start with the technology?

No. Google's People + AI Guidebook suggests that instead of asking whether you can use AI for something, teams start by asking how they might solve the user's problem, and then whether AI can solve it in a unique way. Start from the job the user is trying to get done, then test whether AI solves it better than rules, search or a manual workflow.

### Is proprietary data a moat for an AI product?

Usually less than teams assume. In The Empty Promise of Data Moats, Martin Casado and Peter Lauten of a16z argue that there generally is not an inherent network effect from merely having more data. They note that accumulating proprietary data is strongest when the sources are scarce, and that in some domains more data makes a dramatically better product, so treat data as one input to your right to win rather than the whole answer.

### How long should an AI product strategy document be?

One page for the decisions and a few pages of evidence behind them. If the five decisions do not fit on one page, the strategy has not made its choices yet. The supporting pages hold the evals, cost model and research that back each decision up.

### How often should you revisit an AI product strategy?

Revisit the right-to-win and economics decisions whenever model prices or capabilities shift enough to change them, and review the bets at each planning cycle. The Stanford HAI 2025 AI Index reports that inference cost for a system performing at GPT-3.5 level fell over 280-fold between November 2022 and October 2024, which is the kind of move that can flip a cost decision inside a year.

### Who should own the AI product strategy?

The product leader for the area owns it and makes the calls, with engineering owning the feasibility and cost evidence and data or ML specialists owning the quality evidence. A strategy written by one function and handed to the others tends to fail at the economics or the quality bar.

## Sources

- [Martin Zinkevich, Google: Rules of Machine Learning](https://developers.google.com/machine-learning/guides/rules-of-ml)
- [Google People + AI Guidebook: User Needs + Defining Success](https://pair.withgoogle.com/chapter/user-needs/)
- [Stanford HAI: The 2025 AI Index Report](https://hai.stanford.edu/ai-index/2025-ai-index-report)
- [Martin Casado and Matt Bornstein, a16z: The New Business of AI](https://a16z.com/the-new-business-of-ai-and-how-its-different-from-traditional-software/)
- [Martin Casado and Peter Lauten, a16z: The Empty Promise of Data Moats](https://a16z.com/the-empty-promise-of-data-moats/)
- [Marty Cagan, SVPG: Product Strategy Overview](https://www.svpg.com/product-strategy-overview/)

## How this guide was made

Researched from Builders Camp's bootcamp, track and masterclass material and the sources listed on this page, drafted with AI, and fact-checked against every source cited.