Image
Concepts

Who approves a model going to production

When a model goes live, somebody is accountable for it, whether or not anyone wrote that down. This page sets out who that approver should be, what they need in front of them, and how the decision gets recorded.

The short answer

Short answer

One named person should approve each model release: the person accountable for the product or process the model affects, not the engineer who built it.

Before deciding, they see the evaluation results against the last approved version, the known risks and failure cases, and the rollback plan. Their decision, and its reasons, is recorded.

Today

Definition

Model release approver: the single named person who decides whether a model, prompt or AI configuration change goes into production, after reviewing its evaluation results, known risks and rollback plan, and whose decision is recorded.

What goes wrong when nobody owns the decision?

Without a named approver, the model goes live because the engineer who changed it was satisfied. Nobody outside the build saw the results, the product owner learned about the change from users, and when something goes wrong there is no record of who decided or what they knew. Alternatively, approval becomes a committee that meets too rarely, so changes pile up and people find ways around it.

Who has to settle this question?

A CTO, AI product lead or business owner at a company putting models or AI features in front of customers or into internal decisions. It matters most when a model affects money, people or customers directly, and when changes are frequent enough that an informal habit has already formed.

What should the approver see, and who should it be?

QuestionInformal habitA named release approver
Who decidesThe engineer who made the changeThe person accountable for what the model affects
What they seeA few examplesEvaluation results against the last approved version
Known risksIn the builder's headWritten down, with failure cases
If it goes wrongWorked out at the timeA rollback plan agreed before release
The recordNoneThe decision, the reasons and what was seen

How it works, step by step

The approval is the same short sequence for every release.

01

Name the approver for each model

For every model in production, write down the one person who approves its releases. NIST's AI RMF Playbook suggests planning for risks related to human-AI configurations, and documenting requirements, roles, and responsibilities for human oversight of deployed systems.
02

Prepare the release packet

For each change, the builder or an agent assembles the packet: what changed, evaluation results against the last approved version, regressions, known risks and failure cases, and the rollback plan.
03

Separate the builder from the approver

The person who made the change prepares the packet but does not approve it.
04

Human gate: the named approver decides go or no-go

The approver reviews the packet and decides: release, release with a recorded accepted risk, or send back. This is the determination NIST's Playbook describes under its Manage function: a determination is made as to whether the AI system achieves its intended purpose and stated objectives and whether its development or deployment should proceed.
05

Record the decision and its reasons

The decision, the approver's name, the date and the reasons are recorded with the packet.
06

Watch after release

The first period after release is monitored against the risks the approver accepted, and the approver hears about anything that crosses the line agreed before release.

Keeping approval meaningful

What keeps the approval meaningful?

One name

A committee can advise, but one person decides, as approval gates describes.

A full packet. No packet, no decision.

Accepted risks in writing

Shipping with a known issue is allowed, and recorded.

A rollback plan agreed first

The approver knows how to undo it before saying yes.

No self-approval. Builders do not approve their own changes.

Where this fits

Concept

This principle sits under every AI workflow in the library, and the Platform page describes how such gates are built. The evaluation results the approver reads come from running the suite on every change, and trust in a new configuration is earned in shadow mode. Use the readiness checklist to see whether a process has the owner such approvals need, and the broader case for keeping some decisions human is when not to automate. The use case library shows the workflows these approvals govern.

How does release approval differ from code review?

QuestionCode reviewModel release approval
Who does itAnother engineerThe person accountable for the outcome
What it checksWhether the change is built correctlyWhether it should go live at all
What it relies onThe diffEvaluation results, risks and a rollback plan
What is recordedApproval of the changeThe decision, the reasons and accepted risks
AAOS Platform
How It Works

Questions, answered

Only if they are accountable for what the model affects. Usually the approver is the product or process owner, with the data science lead preparing the packet and advising.
AI Opportunity Audit · Model release

Put one name on every model release

Bring the last model change your team shipped and see who, if anyone, approved it.