Make decision and incident ownership explicit

An original ownership sheet can identify who approves the use case, controls data access, reviews evaluations, and responds to incidents. Make handoffs explicit so a failed model output does not become everybody’s responsibility and nobody’s action.

Policies need triggers that affect deployment

A written policy that is never connected to deployment decisions offers little operational value. Include review triggers for model changes, new data sources, and expanded autonomy.

What the owner has to be able to do

An owner who cannot disable a tool or roll back a prompt does not own the system.

Name the person who can take the endpoint out of production, the person who approves a new data source, and the person who talks to the user who was harmed. They can be the same person in a small team.

They cannot be “the AI working group” if that group has no on-call.

Ownership includes the supplier. If a vendor changes the model, the owner decides whether that change is accepted.

If nobody is assigned, the change is accepted by default. Write that default down so it is a choice.

What to record when something goes wrong

An incident note should say what the system did, what data it used, who was affected, and what was turned off. It should not be a paragraph about how impressive the model usually is.

Keep the tool log. Without it, the owner is guessing, and the next owner will inherit the guess.

What to record before you rely on it

Publish the owner’s name next to the system, the rollback they can perform, and the vendor notice they are expected to read. Review it when the person changes teams.

A name that has left the company is not an owner.

Keep the model id or the paper citation, the date you read it, and the inputs you used beside that note, so a later change can be compared with the same test.

When the result is a failure, store the artifact and change the design. Do not relax the check until the model happens to pass.

Name a person who owns that failure.

Common questions

Is a model card the governance?

A model card describes a checkpoint. Governance is the operating assignment.

You can have a thorough card and still have nobody who can turn the feature off on a Friday.

Do we need a committee?

You need a person who can act. A committee can review.

If the committee is the only name on the page, ask who gets paged.

Sources & further reading

  1. NIST: AI RMF Core ↗
How this story was made

Written by Kristian Kostov with AI assistance and checked against the linked sources. Company performance claims are attributed to the company. Analysis reflects AiLookout’s interpretation; we have not independently tested the products discussed. Cover photography is illustrative and does not depict the specific announcement or product.

Our editorial standards
Back to all stories