Skip to content
AI in Operations | Part 9 of 10

Governance for AI in Operations: Owner, Risk Class, Model Card

A lightweight governance for factory AI: every application has an owner, a risk class, a model card and human oversight.

When the first AI application goes live, governance sounds like overhead. When the fifth one is running and nobody knows which model version checks which parts, it becomes urgent.

Good governance for AI in operations is small, practical and in place before the second use case goes live.

Every application has an owner

Each model in use belongs to a named person in operations, usually the process expert who defined the problem and the labels. The owner knows its purpose, watches its performance and decides when it needs care. An AI application without an owner is like a machine without a maintenance plan.

Classify by risk

Three classes are enough for most plants:

  • Low: shift report summaries, document search, energy forecasts. Owner and model card, users informed.
  • Medium: quality decisions with a human check, maintenance planning. Additionally a parallel run before go-live, human oversight, monthly monitoring and approval by the sponsor.
  • High: safety functions, decisions about people, products with safety relevance. Specialist and legal review before any pilot. Often a learning model is simply not the right tool.

The class decides the checks, not the enthusiasm of the team.

The model card

One page per model, kept with the application:

  • purpose and limits, for example “surface defects on housings H2 and H3, not trained for new colors”,
  • training data and its date,
  • performance measured on the real process,
  • human oversight: who checks what and how often,
  • owner and version.

The model card answers the questions auditors, customers and new colleagues will ask, and it forces the team to write down the limits while they still remember them.

Monitor like equipment

Define one or two measures per model with a threshold. For a vision station: detection rate from the sample check below 97.5 percent in a month, or false rejects above 2.5 percent, triggers a review. Any change of material, supplier or variant triggers a test before it goes live.

Data and rules

Avoid personal data wherever possible. Keep the machine network separated from the office network and document every connection. Agree with suppliers who owns the labeled data and the trained models, and where data is processed.

Rules on AI, data protection and product safety differ by region and use, and they keep developing. Governance in the plant does not replace legal advice. It makes sure the right questions reach the legal function early, especially for medium and high risk uses.

Review it monthly

A thirty-minute model review per month, with the owners of all models in use, is enough for a plant of medium size: monitoring results, changes coming, retraining needs, new risks.

Governance done this way is not bureaucracy. It is the reason the fifth application is as reliable as the first.

Newsletter

One Lean idea.
Every week.