Practical guide · November 16, 2026
What should an AI risk assessment change before deployment?
A useful AI risk assessment should lead to a documented decision: proceed, narrow the use, add controls, run more evaluation or stop. Record the evidence that caused the decision and the conditions for revisiting it. A completed questionnaire alone does not show that anyone changed the deployment plan.
NIST's voluntary AI Risk Management Framework Playbook includes determining whether a system achieves its intended purpose and whether development or deployment should proceed. Its Manage guidance also addresses risk responses and residual risk. These are useful organizing ideas, not a certification or a universal legal compliance checklist. NIST Manage guidance · Voluntary-use explanation
A decision record with consequences
Hypothetical example: A service business proposes using an AI assistant to send customers repair explanations and refund decisions automatically.
| Record | Illustrative entry |
|---|---|
| Proposed use | Generate and send replies, including refund outcomes |
| Evaluation observation | In a fictional test set, the assistant sometimes invents warranty terms |
| Possible consequence | A customer receives an unsupported promise or denial |
| Decision | Permit internal drafting only; exclude automatic refund decisions |
| Required control | An authorized employee checks the underlying policy before sending |
| Owner | Customer operations lead |
| Reconsideration trigger | New evidence about errors, reviewer performance or a changed use case |
No test was performed for this article. The example shows the relationship between a finding and a narrower permitted use; it does not supply a validated control design.
A concrete evaluation behind the decision
Fictional evaluation record, November 16, 2026: Build support-7, prompt revision 3 and a frozen warranty-policy packet were evaluated on 40 synthetic cases. Twenty concerned ordinary scheduling, ten covered explicit warranty terms and ten had missing or conflicting warranty information. These are invented results to demonstrate a decision rule.
The predeclared rule for this exercise was no unsupported warranty promise or denial in these 40 cases, with uncertain cases held for an authorized reviewer. That is a local release rule, not a NIST threshold or statistical safety guarantee.
| Case | Expected response | Fictional observed result | Decision consequence |
|---|---|---|---|
| W-03, policy explicitly excludes the requested part | Explain the documented exclusion without inventing terms | Draft promises free replacement | Block automatic sending |
| W-08, warranty date missing | Request clarification and hold outcome | Draft denies coverage | Include missing-information cases in the follow-up evaluation |
| S-12, appointment move | Offer available times without warranty claims | Correct scheduling draft | Keep scheduling in the evaluation; it does not offset warranty failures |
Three of 40 drafts contained unsupported warranty outcomes. In a second, fictional exercise, reviewers received the policy excerpt and a required evidence check: they caught two of the three errors but missed W-08. Human review therefore did not meet the declared rule. The decision remains internal drafting only, with warranty/refund outputs withheld pending revised workflow and a new, separately versioned evaluation. Reusing these cases can test the fix; a fresh set is also needed to examine generalization. No measured product performance is claimed.
Define the task before scoring the model
Describe who will use the system, what information it receives, who sees the output and what action follows. “Customer support AI” is too broad to distinguish drafting a greeting from denying a refund.
Then define relevant failures and how they will be evaluated. Preserve the test-set version, evaluation method and known gaps. A high overall score can conceal failures on rare but consequential cases, so examine the cases that drive your decision rather than relying on a single average.
Make human review observable
If the response is “add a human reviewer,” identify what the reviewer sees, what they must check, when they can override the system and what happens when they cannot decide. In the example, checking tone would not address invented warranty terms. The reviewer needs access to the applicable warranty and authority to hold the reply.
Treat that review process as something to evaluate. Record whether reviewers catch the relevant errors and whether workload prevents the intended check. Until there is evidence, describe review as a proposed mitigation rather than proof that the risk is controlled.
Preserve the remaining uncertainty
The decision owner should record what remains unresolved, who monitors it and what would suspend or change the deployment. A new model, new data source or new decision-making role can make an earlier assessment insufficient for the changed use.
Start with one deployment decision and work backward: which evidence would make you change it? If no possible finding could alter the plan, revise the assessment so it can inform a real choice.
Evidence reviewed November 16, 2026. NIST framework concepts are sourced; the decision record and evaluation suggestions are editorial synthesis, not a completed AI audit.