AI Product Requirements Document Template

Version: 1.1.0 Last updated: 2026-07-16 Status: Informative implementation template; replace every {{...}} field or state Not applicable — <reason>.

Purpose

Define user outcome, intended use, measurable requirements, risk, and lifecycle controls.

Why

AI product requirements must bind intended use and prohibited use to measurable outcomes and affected-person controls.

How

1. Identify affected users, intended use, prohibited use, and accountable product owner.
2. Set outcome, quality, safety, oversight, and unit-cost requirements with baselines.
3. Document data/model/provider flows and affected-person notices or appeal.
4. Define release, monitoring, rollback, and retirement criteria before approval.

Decision authority

Product owns outcomes; domain and risk reviewers approve high-impact intended use and control sufficiency.

Product and ownership

{{NAME_OWNER_STAKEHOLDERS}}

Intended use and prohibited use

{{USERS_CONTEXT_BOUNDARIES}}

Outcomes and requirements

Requirement	Metric	Baseline	Target	Evidence
{{REQUIREMENT}}	{{METRIC}}	{{BASELINE}}	{{TARGET}}	{{EVIDENCE}}

Data, model, provider, and affected-person impact

{{DATA_FLOWS_RISK_CLASSIFICATION_HUMAN_OVERSIGHT}}

Acceptance, rollout, rollback, and retirement

{{GATES_CANARY_ABORT_OWNER_RETIREMENT}}

Tradeoffs

Use	Benefit	Cost
Complete PRD record	Reconstructable decision and ownership	Authoring and independent review time
Proportionate abbreviated record	Lower low-risk overhead	Requires an explicit tailoring rationale

Anti-patterns

• Writing model capability as a user outcome.
• Omitting prohibited uses or affected-person impact.
• Setting quality targets without populations, denominators, or failure severity.

Enterprise considerations

• Counsel determines legal applicability; the PRD records the decision and engineering obligations.
• Procurement and data owners approve provider/data boundaries.
• High-impact products document meaningful oversight, complaint, correction, and retirement capacity.

Checklist

• Intended and prohibited uses are unambiguous.
• Requirements have baseline, target, population, and evidence source.
• Data/provider and human-decision flows are diagrammed.
• Abort and retirement rules have owners.

References

• NIST, AI RMF MAP (accessed 2026-07-16).

Changelog

Version	Date	Change
1.1.0	2026-07-16	Specialized authority, evidence, tailoring, and recovery fields for the PRD record.
1.0.0	2026-07-16	Initial enterprise template.