Pre-Production AI Evaluation Gates

Objective

Require explicit, documented evaluation against defined thresholds — covering performance, fairness, robustness, safety, and (where applicable) drift baselines — as a gate to production deployment for every AI system.

Applicability

AI types

LLM, Agentic AI, Traditional ML, Computer Vision, Generative AI, Multi-modal, Recommender

Deployment models

Self-hosted, Hybrid, Edge

Lifecycle stages

Evaluation & Testing, Deployment

Risk domains

Model, Data

Regulatory regimes

EU AI Act, ISO 42001, NIST AI RMF, Banking, Healthcare

Company size

MidMarket, Enterprise

Rationale

Deploy-to-production without an evaluation gate is the single most common source of post-deployment AI incidents. EU AI Act Annex IV requires evaluation records for high-risk systems; NIST AI RMF MEASURE-2 requires repeatable measurement; ISO/IEC 42001 clause 9.1 requires evaluation. A formal gate ensures evaluation happens, that thresholds are met, and that exceptions are documented at an appropriate authority.

Control narrative

Every AI system must pass a documented evaluation gate before promotion to production. The gate verifies: performance metrics meet documented thresholds against representative evaluation data; fairness metrics meet thresholds across protected subgroups (where applicable); robustness/adversarial-testing results are within tolerance (AI-CTRL-003 / AI-CTRL-014); safety testing for systems with material impact (e.g., red-team evaluation for LLM agents); drift baseline established for monitoring; documentation per AI-CTRL-012 is complete. Failure to meet a threshold blocks promotion; exception requires sign-off from the AI Governance lead and the business sponsor, with compensating controls and remediation plan recorded.

Test of Design

Procedures

1. Obtain the Evaluation Gate standard and confirm scope, threshold definitions, and gate authority.
2. Confirm evaluation categories required (performance, fairness, robustness, safety, drift baseline, documentation).
3. Confirm exception process with sign-off authority and compensating-control requirement.
4. Confirm linkage to AI-CTRL-010 (Change Control), AI-CTRL-011 (Fairness), AI-CTRL-003/014 (Adversarial), AI-CTRL-012 (Documentation).

Inquiries

• Who is the gatekeeper? Is this separate from the model author?
• How are thresholds set for each system?
• How are evaluation datasets kept representative of production?
• What is the override path and how often is it invoked?

Inspections

• Evaluation Gate standard.
• Threshold definitions per system type / risk tier.
• Sample of evaluation gate decisions (approved, rejected, exception).
• Exception records with sign-off and compensating controls.

Test of Operating Effectiveness

Procedures

1. From the inventory, sample model promotions to production in the audit period.
2. For each, confirm a documented evaluation gate decision exists, pre-dating promotion.
3. Inspect the evaluation record for completeness across required categories.
4. Confirm thresholds were met or exception sign-off is documented.
5. Confirm gatekeeper identity is separate from model author.
6. For exceptions, confirm compensating controls were implemented (trace to operational evidence).
7. Confirm drift baselines were established and are referenced in monitoring (AI-CTRL-008).

Sample-size guidance

Reperformance

• For 1 sampled promotion, independently re-run a subset of the documented evaluation suite and confirm results match the gate decision.

Evidence requirements

Required

• Evaluation Gate standard PDF/Word · At fieldwork
• Gate decision records for sampled promotions System / PDF · Per sample
• Evaluation reports underlying gate decisions PDF/Notebook · Per sample
• Exception sign-off records (where applicable) PDF/email · Per sample

Supporting

• Threshold definitions per system / risk tier Reference doc · At fieldwork

Retention: Lifetime of model + 7 years post-retirement (regulated); + 3 years (otherwise)

Framework mappings

Related controls

• AI-CTRL-003

  Adversarial Robustness Testing for LLM Systems

• AI-CTRL-010

  Model Versioning and Change Control

• AI-CTRL-011

  Bias Testing and Fairness Validation

• AI-CTRL-014

  Adversarial Robustness Testing for Predictive ML and Computer Vision

References

• NIST AI RMF — MEASURE function
• EU AI Act Annex IV — Technical documentation
• ISO/IEC 42001:2023, clause 9.1