Why AI Products Need Their Own PRD
Most AI product requirements documents fail because they treat AI features like traditional software features. This template is designed specifically for AI products.
A standard requirements document says what a feature does. An AI feature also needs a statement of how well it must do it, what it does when it is unsure, and how you will know. The six steps below cover those questions in order.
The Six Sections
Step 1: Define the user problem and AI hypothesis
What problem does the user have? How does AI solve it better than non-AI approaches? What is your hypothesis about model capabilities? Document assumptions explicitly.
Write: the user, the problem, and one sentence that begins "We believe the model can".
Step 2: Specify AI behavior requirements
Instead of just listing features, define expected AI behavior: What should the AI do when it is confident? What should it do when uncertain? What is the acceptable error rate? When should it escalate to a human?
Write: the behavior in three cases (confident, uncertain, wrong) and the rule for passing a case to a person.
Step 3: Define data requirements
What training data or knowledge base does the AI need? Where does this data come from? How frequently does it need updating? What are the data privacy requirements?
Write: each data source, its owner, how often it changes, and who may see it.
Step 4: Specify model requirements
Which LLM or ML model should be used? What are the latency requirements? What is the acceptable cost per inference? Are there any compliance constraints on model selection?
Write: the model choice and the reason, the latency limit, the cost limit per request, and any rules on where data may be processed.
Step 5: Define evaluation criteria
How will you measure if the AI is working correctly? Define quantitative metrics (accuracy, latency, cost) and qualitative criteria (user satisfaction, output quality). Include example inputs and expected outputs.
Write: the metrics, the target for each, and a set of real example inputs with the output you would accept for each.
Step 6: Plan for iteration
AI products require more iteration than traditional software. Plan for prompt tuning, model switching, and feedback loop integration in your timeline.
Write: how often you will review failures, who decides on prompt and model changes, and how a change is tested before release.
The Template
Copy the headings below and answer each line. The example under each heading is illustrative, written for a made-up support-ticket product. It isn't a client project.
1. Problem and AI hypothesis
- User and problem: who has the problem, and what do they do about it today?
- AI hypothesis: we believe the model can ...
- Example: support agents spend time sorting incoming tickets. We believe a model can label the topic and urgency of a ticket accurately enough to save that sorting time.
2. AI behavior
- When confident: what does the AI do?
- When uncertain: what does it do?
- When wrong: how is the error found and fixed?
- Escalation: when does a person take over?
- Example: when confident, the label is applied; when uncertain, the ticket goes to a person with the AI's suggestion shown; every label can be corrected in one click and the correction is stored.
3. Data
- Sources, owner, update frequency, privacy limits
- Example: past tickets with their final labels, owned by the support team, refreshed monthly, with personal details removed before any text is sent to a model provider.
4. Model
- Model choice and why, latency limit, cost limit per request, compliance limits
- Example: a hosted language model, a response within a few seconds, a cost per ticket the support budget can carry.
5. Evaluation
- Metrics and targets, plus the example inputs and accepted outputs
- Example: label accuracy on a set of real past tickets that the team labelled by hand, measured before launch and again after every prompt or model change.
6. Iteration
- Review rhythm, decision owner, release test
- Example: a weekly review of corrected labels, with prompt changes released only after the evaluation set is re-run.
Review Checklist
- Is the AI's behavior defined for the confident, uncertain and wrong cases?
- Is there a rule for handing a case to a person?
- Is every data source named, with an owner and privacy limits?
- Does every model choice have a latency limit and a cost limit?
- Are there example inputs with accepted outputs?
- Is the evaluation set built from real inputs?
- Is the plan for iteration and failure review written down?
The complete AI PRD template, with fill-in sections, example completions, and a review checklist, is on this page, so there is nothing to download. For what comes next, see how to validate your AI MVP idea, the AI MVP to production timeline and AI MVP development, or book a strategy call to turn the document into a scoped build.
