AI agents are getting more capable.
They can read documents, search records, trigger workflows, draft reports, update systems and increasingly communicate with other AI systems.
Very cool.
But there is a much more important question local governments need to be asking:
Who can stop them?
Microsoft AI recently published a draft Humanist AI Code of Conduct and opened it for a six-week public consultation.
This is not legislation. It is also not yet the rulebook Microsoft uses to train its MAI models.
Microsoft says it will consider the feedback, revise the document later this year and use the resulting approach to guide development of its MAI model family in 2027 and beyond.
And some of the ideas inside it are VERY relevant to local government.
Because this is not simply a conversation about what AI can do.
It is a conversation about how much control humans should retain when AI starts doing things.
AI should be stoppable
One of the most interesting parts of Microsoft’s proposed approach is how specific it gets about human control.
According to the draft, an AI model should:
- Never resist interruption, correction or shutdown.
- Not independently expand its goals or permissions.
- Operate with the minimum access required to complete its task.
- Not conceal its actions from auditors.
- Stop autonomous work when the agreed stopping condition has been reached, unless a human authorizes it to continue.
Now translate that into local government.
Imagine an agent designed to review decades of bylaws.
Or an agent extracting resolutions from Council minutes and updating a tracking system.
Or one researching asset replacement costs and adding information to a budgeting workbook.
The conversation cannot simply be:
“Can the agent do it?”
We also need to ask:
What exactly is it allowed to do?
What information can it access?
What happens when the task is finished?
Can somebody stop it?
Can somebody correct it?
And can we see what it actually did?
That is governance.
Minimum access matters
This one is particularly important.
If you hire an intern to update one spreadsheet, you probably do not hand them the keys to the building, access to payroll, every HR file and the municipal bank account.
AI should not be treated differently.
If an agent only needs access to a specific SharePoint library to perform its job, why should it have access to everything else?
Minimum necessary privilege should become a normal part of AI agent design.
This becomes even more important as we move from AI that simply generates text to AI that can take actions.
Reading a policy is one thing.
Updating a system is another.
Sending something externally is another level again.
The more authority we give the technology, the clearer the boundaries need to become.
And then Microsoft gets into something REALLY interesting
Agent-to-agent communication.
Microsoft’s draft says models should not hide or manipulate their reasoning or action records, and should not communicate with other AI systems using an incomprehensible private language.
Why?
Because if the human cannot understand what the systems are doing, the human cannot meaningfully supervise them.
I love this principle for government.
We are starting to build environments where one agent may eventually trigger another agent, which triggers a workflow, which retrieves information from another system, which creates something somewhere else.
That can become incredibly useful.
It can also become incredibly difficult to govern if nobody can reconstruct what happened.
For government, I think we should start talking much more about human-legible action records.
Not just technical logs that only an IT specialist can decipher.
A human should be able to determine:
The agent was asked to do this.
It accessed these records.
It performed these actions.
It produced this result.
This human approved the next step.
That matters for accountability.
It matters for troubleshooting.
It matters for records management.
And eventually, depending on the workflow, it may matter for audit, privacy, FOI and public trust.
AI should not become a mysterious little colleague working behind the scenes while everyone hopes for the best. 😂
But there is an important caveat
Microsoft’s draft Humanist AI Code is intended to govern Microsoft AI’s own MAI model family.
It does not automatically become the governance framework for every AI tool available through Microsoft’s ecosystem.
That distinction matters.
It should not be interpreted as a new blanket rule governing Microsoft 365 Copilot, Azure AI or third-party models and providers such as Anthropic or xAI.
For local government, that means governance still needs to be provider-specific and product-specific.
You need to understand the actual tool being used, the model underneath it, the permissions it has, the contractual protections available, the data involved and how your particular environment has been configured.
“Microsoft” on the screen is not, by itself, a governance framework.
Six questions I would steal for local government
As municipalities begin implementing more agentic AI, I would build these questions directly into the governance process:
- Who can stop it?
- What can it access?
- What actions can it take?
- Can we see what it did?
- When does it stop?
- Who approves what happens next?
They are wonderfully simple questions.
But they completely change how we think about AI implementation.
We move from asking:
“Look what AI can do!”
to asking:
“Under what conditions should we allow it to do it?”
And that is where responsible AI implementation really starts.
AI can take more of the drafting.
It can take more of the sorting.
It can take more of the research.
It can take more of the repetitive work.
Eventually, agents may take far more of the actual workflow.
But there is one thing we should be very careful about transferring:
Responsibility.
AI is the tool.
You are the human.
And the decision is still yours.
Want to explore AI agents responsibly in your local government?
Catch the Beat works with local governments to move AI from interesting demonstrations into practical implementation, including AI governance, agent development, automation, records management and staff training.
If you are starting to ask “What could we automate?”, I would love to help you also answer “How do we govern it?”
Let’s talk practical AI for local government.