Skip to main content

AI risk assessment checklist for SaaS: review a feature before launch

Assess an AI feature's intended use, affected people, data, failure modes and controls before release, then assign owners and revisit the risks as the product changes.

In this guide

How do you assess risk in a SaaS AI feature?

Review the complete product workflow, not just the model. Record the intended task, users and affected people, data sources, provider, interface, automation authority, likely failures and controls. NIST's AI Risk Management Framework groups activities under Govern, Map, Measure and Manage; its use is voluntary, not a certification, and NIST states that AI RMF 1.0 is being revised. Use the framework as a structured conversation and tailor it to the actual feature and its impact.

Define the use case and where the AI is allowed to act

Describe what problem the feature solves, who uses it, who may be affected, what inputs it sees and what output it produces. State what it must not do. Separate drafting or search assistance from decisions about eligibility, safety, employment, finances or access to services. Document when a person must verify an answer or approve an external action.

Map the full data and dependency path

List the application, model and provider, prompt and tools, retrieval sources, tenant boundary, human review, logs and downstream systems. Identify personal or confidential information, external actions, shared indexes and third-party dependencies. Ask what can change independently, who owns each component, what users have been told and how data is retained or deleted.

Identify who could be harmed and how

Consider inaccurate or missing answers, biased treatment, inaccessible interactions, language gaps, privacy exposure, prompt injection, excessive tool authority, fraud and service outages. Think about people who are not the direct account holder but may be affected by an output. Estimate severity and exposure using evidence; do not reduce a serious consequence to a single unexamined average score.

SaaS AI feature risk assessment
Use case and limitsAffected people and dataFailure and impactControl and verification evidenceOwner and release decision

What should the AI pre-launch review measure?

Choose tests that match the real user task

Build representative evaluations for correct answers, missing evidence, refusals, language and accessibility variants, malicious inputs, out-of-scope requests and boundary cases. Measure the quality dimension that matters: source support, task completion, harmful errors, false refusals or unauthorized actions. Keep subgroup results visible where sample sizes allow, and document what the tests do not measure.

Set controls and stop conditions before testing

For each material risk, name a control, accountable owner and test. Controls might include server-side authorization, restricted tools, human approval, rate limits, source filters, a user correction route or a disable switch. Define what result blocks release, who can accept residual risk and how an exception expires. A disclaimer is not a substitute for a working safeguard.

Make the release decision explicit

Record the feature version, evaluated model and prompt, evidence reviewed, known limitations, mitigations, unresolved risks, approver and rollout scope. Start with a reversible deployment and monitor the agreed signals. NIST's Playbook is voluntary guidance with suggested actions, not a mandatory ordered checklist; select practices that fit the use case and retain a clear rationale for the decision.

How should an AI risk assessment stay current?

Reassess when a meaningful dependency or use changes

Review risk after changing the model, provider, system prompt, tools, retrieved corpus, target users or level of automation. A change that looks like routine configuration can alter privacy, accuracy or authority. Use change-impact review to decide which tests must rerun, who needs to approve and whether the release needs a staged rollout.

Connect incidents and user reports to the review

Feed confirmed failures, appeals, safety reports, accessibility barriers and privacy events into the risk register and evaluation suite after review. Track the risk, mitigation, owner, due date and evidence that the control works. Do not treat report counts as ground truth without considering exposure, reporting friction and changes in usage.

Keep the review proportionate and understandable

A small internal assistant and a feature that can affect a customer's rights or money need different scrutiny. Explain the rationale, evidence and remaining uncertainty in language product, engineering, security, support and affected-user representatives can understand. Reuse a common template, but avoid giving every AI feature the same risk score or approval path.

SaaS AI risk assessment: FAQs

Is the NIST AI RMF a legal requirement or certification?

The NIST AI RMF is voluntary guidance. It does not certify a product or settle which laws apply. Check the current rules, contracts and sector requirements for the product and jurisdiction with qualified advisers.