AI Risk Assessment Checklist

Use this checklist when introducing or expanding an AI capability — whether you build with a foundation model API, embed a vendor copilot, or automate a workflow with agents.

It is intentionally practical for technology, product, and risk stakeholders working together. It is not a certification, regulatory filing template, or substitute for legal counsel.

How to use it

1. Pick one AI use case (not “all AI in the company”).
2. Walk the groups in order: use case → data → model → privacy → security → oversight → fairness → third parties → monitoring → incidents → business impact → resilience.
3. Record evidence (links, docs, ticket IDs) next to each item — “assumed fine” does not count.
4. Turn gaps into owners and dates. Re-assess after material model, data, or vendor changes.

Pair with the AI Vendor Risk Assessment Guide when a third party supplies the model or hosts prompts.

Use case

Clarify what the system does for whom, and what must never happen.

• Intended users, channels, and geographies
• Assistive vs automated decisioning
• Explicit out-of-scope and misuse scenarios
• Named owner and internal risk tier

Data

Map what enters the model context and what is retained afterward.

• Categories of personal and confidential data
• Minimisation and redaction before send
• Retention for prompts, outputs, embeddings, and logs
• Special handling for sensitive categories

Model

Know which model you are actually running and how it fails.

• Provider, model ID/version, region
• Evaluation on representative tasks
• Known limitations and hallucination risk for this domain
• Change control when prompts or models update

Privacy

Treat AI paths as processing activities, not magic side effects.

• Notices and lawful basis considerations (as applicable to your organisation)
• Provider training / data-use settings
• Deletion and export coverage for AI stores
• Cross-border transfer awareness

Security

Protect keys, tenancy, and untrusted content.

• Server-side secrets only
• AuthN/AuthZ on every AI and mutation path
• Prompt-injection and document-ingestion hygiene
• Controls against sensitive data exfiltration

Human oversight

Decide where a human must remain in the loop.

• Review gates for high-impact actions
• User challenge / escalation path
• Kill switch or feature disablement

Bias and fairness

Ask who could be harmed by skewed or stereotyped outputs for this use case.

• Affected groups and scenarios
• Targeted testing
• Documented mitigations (including “do not automate this”)

Third parties

Inventory the chain: model host, plugins, vector DBs, logging tools, support access.

• Subprocessors and data flows
• Contract / DPA review for AI-specific terms
• Exit and portability assumptions

Monitoring

Production AI needs ongoing observation, not a one-time demo review.

• Quality and failure signals
• Cost and abuse spikes
• Drift and silent degradation

Incident handling

Prepare for harmful output, data exposure, and provider outages.

• Playbook and severity criteria
• Escalation and communications
• Evidence retention for investigation

Business impact

Connect risk work to why the capability exists.

• Success metrics and value hypothesis
• Dependency of critical processes
• Manual fallback if AI is off

Resilience

Assume models, regions, and rate limits will fail.

• Timeouts, retries, degraded UX
• Surge / cost behaviour
• Backup path for critical workflows

Closing note

Complete this checklist per use case, keep evidence with the answers, and revisit after material changes. For identity-related controls that often sit underneath AI apps, see the Identity & Access Management Risk Checklist.