AI Risk Assessment for GxP Systems: Template

What should an AI risk assessment for GxP systems include?

Context of use, model type, decision influence, consequence of error, existing compensating controls, required assurance activity, human oversight requirement, baseline performance reference, and a reassessment trigger. The risk rating comes from combining decision influence with error consequence — not from how sophisticated or accurate the model is in isolation.

Two AI models can use the exact same underlying algorithm and land at opposite ends of the risk scale. The model isn't what you're rating — the decision it feeds, and what happens if that decision is wrong, is.

The AI Risk Assessment Template

Field	What It Captures
Context of use	What question the model answers, who acts on the output, autonomous or advisory
Model type	Static/deterministic, dynamic/adaptive, or generative
Decision influence	Sole basis for action, or one input among several a human weighs
Consequence of error	Impact on patient safety, product quality, or data integrity if the output is wrong
Compensating controls	Independent human review, secondary checks, or other safeguards already in place
Risk rating	Combination of decision influence and consequence — the output of this assessment
Required assurance activity	Scripted testing, scenario-based testing, or vendor evidence, scaled to the rating
Human oversight requirement	Mandatory review before use, or autonomous operation permitted
Baseline performance reference	What the model is being measured against — the process or method it supports or replaces
Reassessment trigger	Retraining, vendor model update, or drift monitoring threshold that requires a fresh review

How to Score Risk: Influence × Consequence

Risk isn't a property of the model — it's a property of the decision the model feeds. Score two factors independently: how much the output influences the decision (sole basis for action versus one input a human weighs), and how severe the consequence would be if that output were wrong. A model scoring high on both is high risk. A model scoring high on influence but genuinely low on consequence, or vice versa, sits somewhere in the middle — and the same model can score differently depending entirely on how it's deployed.

Common Mistakes in AI Risk Assessments

Rating the model instead of the decision. A sophisticated, highly accurate model deployed autonomously for a critical decision is higher risk than a simple model whose output a human always double-checks — accuracy and sophistication don't override decision context.
Skipping the baseline performance reference. Without knowing what the model is being measured against, "acceptable performance" has no defensible meaning.
Treating vendor risk documentation as sufficient. A vendor's assessment reflects general model behavior, not your specific context of use and compensating controls.
No reassessment trigger defined. A model's risk rating from initial deployment doesn't automatically update itself when the context of use changes or the model is retrained.

How GoVal Supports AI Risk Assessment

GoVal captures the AI risk assessment as a structured record tied to each AI-embedded system, scaling required validation evidence, oversight requirements, and reassessment triggers to the documented risk score rather than applying a fixed checklist to every model. Context of use, decision influence, and consequence are recorded alongside the resulting risk tier, kept in the same audit-trailed structure as the rest of the system's validation history.

Frequently Asked Questions

What should an AI risk assessment for GxP systems include? +

A complete template captures the context of use, the model type (static, dynamic, or generative), how much the output influences a decision, the consequence of an error, existing compensating controls like human review, the assurance activity required, whether human oversight is mandatory before use, a baseline performance reference, and the trigger for reassessment. The risk rating comes from combining decision influence with error consequence, not from evaluating the model's sophistication alone.

How do you determine the risk level of an AI model? +

Combine two factors: how much the model's output influences a decision, and how severe the consequence would be if that output were wrong. Risk increases as the model has greater influence over a GxP decision and as the consequence of an incorrect output becomes more severe. Independent human review and other compensating controls can reduce the effective risk when they are consistently enforced and documented.

Is an LLM used for drafting always low risk? +

Only when a qualified person reviews and takes documented responsibility for the output before it is used, and the output cannot bypass that review to directly affect a GxP decision. The risk level depends on the context of use, decision influence, consequence of error, and effectiveness of the required controls.

Does a vendor's own AI risk assessment satisfy the requirement? +

No. A vendor's risk assessment reflects their model's general behavior and intended design, not the specific context of use, decision consequence, and compensating controls that exist in your environment. The same model can carry different risk in different deployments depending on what decision it feeds, so the regulated company remains responsible for its own documented risk assessment.

How often should an AI risk assessment be revisited? +

Whenever the context of use changes, the model is retrained or updated, or drift monitoring shows performance has shifted from its original baseline — not on a fixed calendar alone. A model deployed for one purpose that later gets applied to a higher-stakes decision needs a fresh risk assessment reflecting that new context.

How does GoVal support AI risk assessment for GxP systems? +

Score AI risk by decision consequence, not model hype

Context of use, risk scoring, and reassessment triggers — captured as one structured record, in GoVal.

Book a Free Demo →