An AI assistant that can take action needs more than a friendly personality. It needs boundaries—and someone who can reach the stop button.
What Microsoft announced
On September 14, Microsoft AI published a draft Code of Conduct for its MAI models and opened a six-week public consultation. The proposal describes intended model behaviour, including human oversight, limits on autonomous action, and acceptance of interruption or shutdown. Microsoft plans to review feedback and publish a revised version later this year. Read Microsoft’s announcement.
The draft code is a statement of intended training and behaviour standards. Its publication does not establish that every deployed model already meets those standards, or that every Microsoft product uses the same model.
Guava analysis: turn principles into purchasing questions
For a business evaluating an AI agent, the useful conversation starts with what the system can actually do in your environment. A supplier’s safety policy belongs alongside permission settings, testing evidence, and a clear operational owner.
- Which systems can the agent read or change, and who grants that access?
- Which actions require a person’s approval before execution?
- Can administrators interrupt a task and revoke access promptly?
- What records show the instructions, approvals, and actions involved?
- How does the supplier test failures, and how will you learn about significant changes?
Ask for a demonstration using an ordinary business workflow. For example, let an assistant prepare a supplier email, then check whether it waits for approval before sending. Repeat the exercise with an ambiguous instruction and a document containing conflicting directions.
Give your pilot a clear finish line
Choose one bounded task, identify the data it needs, and record what success looks like. Measure both useful results and the work required to review them. A fast draft that takes longer to correct has simply moved the queue.
Assign an owner who can suspend the pilot, document how staff report mistakes, and review access before expanding the deployment. These are Guava’s practical evaluation suggestions, not claims that Microsoft’s consultation mandates particular business controls.
Planning your next step? Explore Guava’s AI and Copilot collection for more practical guidance.
