An AI risk assessment records what an AI system does, what could go wrong, how likely and severe that is, and what you're doing about it. Below is a practical template aligned to the four NIST AI RMF functions — Govern, Map, Measure, Manage — that a SaaS company can complete in about an hour per system and reuse for SOC 2, ISO 42001 and customer questionnaires.
Section 1: System profile (Govern + Map)
| Field | What to record |
|---|---|
| System name | e.g. "Support reply assistant" |
| Owner | Accountable person |
| Purpose | What business problem it solves |
| Your role | Developer, provider, deployer — or several |
| Users | Internal staff, customers, end users |
| Affected people | Who is impacted by outputs, even if they never use it |
| Model(s) and provider | Model, version, hosting |
| Data inputs | Types and sensitivity (customer data, personal data, confidential) |
| Outputs | What it produces and where outputs go |
| Autonomy | Suggests only, acts with approval, or acts automatically |
| Regulatory flags | EU AI Act category, sector rules, contractual commitments |
Section 2: Risk identification (Map)
Assess each category and note specific scenarios:
| Category | Example risks |
|---|---|
| Security | Prompt injection, data exfiltration, over-privileged agents, API key abuse |
| Privacy | Personal data in prompts, logs or training; retention by providers |
| Accuracy / reliability | Hallucinations, outdated knowledge, drift after model updates |
| Fairness | Biased outcomes in decisions about people |
| Transparency | Users unaware they interact with AI; unmarked generated content |
| Safety / misuse | Harmful content, use outside intended purpose |
| Third-party | Provider outage, terms changes, sub-processors |
| Legal / contractual | EU AI Act obligations, customer contract terms, IP |
Section 3: Scoring (Measure)
Use a simple 1–5 scale for likelihood and impact, then multiply:
| Score | Rating | Action |
|---|---|---|
| 15–25 | High | Treat before launch or immediately; leadership sign-off |
| 8–14 | Medium | Treatment plan with owner and date |
| 1–7 | Low | Accept or monitor |
Base likelihood on evidence where you can: test results, red-team findings, incident history, monitoring data.
Section 4: Treatment (Manage)
For each medium or high risk record: the control (e.g. tenant-filtered retrieval, human approval, output filtering), the owner, the due date, the residual risk after treatment, and the decision — mitigate, transfer, avoid or accept (with who accepted it).
Section 5: Monitoring and review (Manage)
- Metrics to watch (error rates, user complaints, flagged outputs, abuse attempts)
- Review triggers: model or provider change, new data types, new user groups, incidents
- Review date — at least annually
Example: support reply assistant
| Risk | L | I | Score | Treatment |
|---|---|---|---|---|
| Reply includes another customer's data via retrieval | 2 | 5 | 10 | Permission-filtered retrieval; cross-tenant tests in CI |
| Inaccurate answer sent to customer | 3 | 3 | 9 | Agent reviews before sending; feedback loop |
| Prompt injection through ticket content | 3 | 3 | 9 | No tool access beyond drafting; output filtering |
| Provider retains ticket data | 2 | 3 | 6 | Zero-retention terms; vendor review |
How it connects to SOC 2 and ISO 42001
Add AI systems to your existing SOC 2 risk register rather than keeping a separate one. For ISO 42001, extend this template with an AI system impact assessment covering effects on individuals and society. See NIST AI RMF explained and ISO 42001 readiness.
FAQ
How many AI systems need an assessment?
Every system in your AI inventory, scaled to risk. An internal summarisation tool needs a short assessment; a customer-facing agent needs the full template.
Who should complete it?
The system owner, with input from engineering, security and — for decisions about people — legal or HR.
How often should it be updated?
At least annually, and whenever the model, provider, data or use changes materially.
Want help assessing your AI systems? Book a free scoping call or see our AI governance service.