Turn AI interest into testable requirements
A useful requirement states what the system must do, the evidence that will prove it and the consequence if it fails. Adapt this checklist to the actual use rather than sending every question to every supplier.
Separate non-negotiable gates from preferences. A product that fails a critical data, security, accuracy, accessibility, legal or workflow requirement should not pass because it scores highly on attractive features.
Use case and people
- The supported task and excluded uses are written clearly.
- Users, reviewers, administrators and affected people are identified.
- The expected benefit and baseline process are measurable.
- The consequence of a wrong, missing or delayed output is understood.
- Human approval and escalation responsibilities are assigned.
- Accessibility and language needs are included.
Output and evaluation
- Representative test cases cover routine and difficult examples.
- Acceptance measures include usefulness and material failure rates.
- Factual claims can be checked against reliable evidence.
- Uncertainty, refusal and missing-information behaviour are tested.
- Bias or uneven performance across relevant groups is assessed.
- Model or product changes trigger proportionate re-testing.
Data, security and administration
- Allowed and prohibited input data are defined.
- Retention, provider training use and deletion are understood.
- Identity, multi-factor authentication, roles and account removal are supported.
- Logs, usage controls, incident routes and supplier evidence are adequate.
- Connectors and actions use the minimum necessary access.
- Prompt injection, sensitive disclosure and unsafe output handling are considered.
Commercial, operational and exit
- Usage units, limits, add-ons and likely volume are modelled.
- Review, correction, training and administration time are costed.
- Support, availability and change-notice expectations are recorded.
- Intellectual-property and output-use terms are reviewed.
- Data, prompts, configurations and records can be exported where needed.
- A safe suspension, fallback and exit process is documented.
AI requirements FAQs
What is a must-have AI requirement?
It is a condition necessary for the intended workflow, risk control, accessibility need, contractual position or applicable obligation. Failure means the product is unsuitable for that use.
Should every AI requirement be scored?
No. Use pass or fail gates for genuine non-negotiables. Score preferences only after every surviving option has met the essential conditions.
How often should AI requirements be reviewed?
Review them when the use, provider, model, data, integration, law or consequence changes, and at a proportionate scheduled interval for continuing use.
Continue your AI decision
Use the next guide that matches the question or risk you still need to resolve.