When AI runs in production, it needs the same control as any other system
What does AI governance actually involve?
Making AI features in live operations as accountable as any other production system: what ran, on which inputs, with which output, who can override it, and what the business does when the model is wrong or down. At Entrivis this is reliability engineering applied to AI: the same “we stay” discipline, not a separate compliance product we overclaim.
The failure mode is familiar. A feature ships, the demo impresses, and nobody owns the question of what “wrong” looks like. There is no evaluation set, no drift signal, and no fallback path. Six weeks later the team has stopped trusting the outputs, which is how AI projects die without a dramatic incident.
We design the control plane with the feature: audit log, eval harness, human override, and a documented unavailable path. That is the difference between a demo and something operations will keep.
What we put in place
Audit and logging
Who triggered the run, which inputs were used, what the model returned, and what was written back: reconstructable after the fact.
Evaluation harness
Test sets and acceptance criteria before go-live, so “good enough” is a number you can argue about, not a feeling in a meeting.
Drift and health monitoring
Signals that the model or the upstream data has shifted before customers feel it. Owned dashboards, not hope.
Fallback architecture
What the system does when the model is slow, wrong, or unavailable: degrade, queue for humans, or block the write. Written down, not improvised.
How we treat AI in live systems
Built after the system could be trusted
For VitaOne, the AI-assisted Functional Lab Report loop came after stabilization and migration: not as the pitch. Reporting was useful because the data underneath finally was.
Read the VitaOne case study →The section most proposals omit
We price evaluation, logging, and fallback with the feature. Removing them later is not a saving; it is how the second partner gets hired. About a third of our clients came to us second. AI produces a clean version of that pattern.
Governance is not a slide. It is the runtime
We are not selling an AI audit certificate. We are building the same monitoring and control we expect on any operations-critical path, applied to features that call a model. If the feature cannot say what it did yesterday, it is not ready for production.
Custom models and product surfaces live under AI custom solutions. Queues and human review gates live under AI workflow automation. Hosting and data posture for the wider firm is on Trust.
How governance work runs
Usually scoped with a custom build or a rescue of an AI feature already live. Rarely a standalone “governance package.”
-
01
Inventory what is live
Which features call a model, where outputs land, and who believes they own failures today.
-
02
Define wrong and unavailable
Acceptance criteria, override roles, and the degrade path when the model is down. Written before tooling.
-
03
Instrument and fall back
Logging, eval sets, health signals, and the queue or block behaviour when checks fail.
-
04
Handover and sustained ownership
Prompts, eval sets, and runbooks in your repository. Managed Continuity when you want us watching the signals with you.
Common questions about AI governance
Is this a compliance or certification service?
No. We do not sell ISO-style AI certifications. This page is engineering: logs, evals, drift signals, and fallback for AI features inside your operations. Broader firm posture is on Trust.
Can governance be added after a feature already shipped?
Yes. That is often how the work arrives. We inventory what is live, turn off what cannot be trusted, and instrument what stays. If the system is already failing users, start at Rescue & Stabilization.
Do you need to have built the model to govern it?
No. We can instrument and wrap features another team shipped, provided we can see inputs, outputs, and write-back paths.
Who owns the eval sets and runbooks?
You do. They hand over with the feature. We do not hold your prompts hostage to a retainer.
How does this relate to the other AI pages?
Custom solutions own the model surface. Workflow automation owns the operational path. This page owns accountability when either is live. Durable work usually needs all three; buyers should enter at the gap they feel first.
Ask what happens when the model is wrong
If the proposal has no answer, that is the engagement. Tell us what is live or what you are about to ship.