Model Governance
Which models do we trust — and how do we know?
Choosing, validating, and tracking the models themselves: a curated catalog with an approval path, provenance checks on third-party and open weights, evaluation as a repeatable gate rather than a launch ritual, version pinning against silent provider churn, and bias and safety testing proportionate to what the model decides. Classic model risk management supplies the skeleton; GenAI forces it to evaluate behavior distributions, not fixed test vectors.
An unevaluated model in production is an unread contract you signed on behalf of the business.
Why leadership should care
- Frontier models are third-party: weights, training data, and alignment process are not inspectable. Trust must come from evaluation, contract, and provenance — not inspection.
- Providers update hosted models continuously; behavior shifts with no change ticket on your side. Version pinning and regression evals are the only counterweight.
- Where models touch decisions about people, bias liability is live now: courts allowed a nationwide collective action over AI hiring screens and held that vendors can be liable as employers' agents.
- Do we run one vetted model catalog with an approval path — or do teams pick models ad hoc?
- What evaluation bar must any model clear before production, and who owns the bar?
- What is our position on open-weight models, and who governs them once downloaded?
Model Garden
GAGemini Enterprise Agent Platform (formerly Vertex AI)
Curated catalog of 200+ Google, open, and partner models.
One vetted front door for model supply instead of ungoverned model sprawl.
Model allowlisting (org policy)
GAOrganization Policy
vertexai.allowedGenAIModels restricts which models any project may call.
Central model approval enforced preventively across the resource hierarchy.
Model Registry
GAGemini Enterprise Agent Platform
Central inventory of customer models: versions, aliases, lineage.
Version control and audit trail for first-party and tuned models.
Version pinning & retirement policy
GAGemini models
Pinned stable versions with published retirement dates.
Managed change control against silent provider model churn.
Gen AI evaluation service
GAGemini Enterprise Agent Platform
Computation and LLM-as-judge metrics for models, apps, and agents.
Documented pre-deployment quality gates and regression evidence.
Generative AI indemnification
GAContractual
Two-pronged IP indemnity: training data and generated output.
Shifts copyright-infringement risk for covered services to Google.
Cards link to official documentation. Status is a snapshot (August 2026) — verify per component before contractual commitments. Full mapping and honest gaps: 08 · Google Cloud.
You no longer validate an artifact you built — you continuously evaluate behavior you rent.
- Morgan Stanley's famous move wasn't the chatbot — it was writing the evals before the rollout. Evaluation discipline is the control regulators will ask to see.
- One vetted catalog beats per-team model choice: Goldman and Walmart both built exactly this before scaling.
- Silent model churn is real: providers update behavior under your feet. Pin versions and make updates rerun your regression suite.
- Open models move governance onto you: once weights are downloaded, platform guardrails and logging no longer apply.
- How many distinct models are in production, and who approved each one?
- What evaluation evidence exists for your most business-critical AI use case?
- What happens on your side when a provider updates a hosted model?
- Do any models influence decisions about people — and when were they last bias-tested?
- Who governs open-weight models your teams have downloaded?
Deutsche Bank
Google CloudFinancial Services
'Audit-ready gen AI' concretely means: metric dictionary, versioned test sets, pinned baselines, logged prompts.
Commerzbank
Google CloudFinancial Services
Use a second service to judge and explain the model's work — the explanation becomes the audit trail.
TELUS
Google CloudTelecom
Model choice and governance aren't in tension when the gateway owns the controls.