# Governed by Human: which decisions stay with people?

Fabio Munhoz · 2026-11-20

Canonical: https://fabiomunhoz.com/blog/governed-by-human/

Language: en

Useful autonomy starts with a clear agreement about what an agent may decide, when it must stop, and who owns the outcome.

## In brief

Governed by Human is a design principle: people define objectives, boundaries, and intervention criteria. Agent autonomy should reflect impact and reversibility. Meaningful approval requires enough evidence and a real ability to stop execution.

A recommendation arrives ready for approval: change the audience, increase spending, and offer a different commercial arrangement. The agent provides a persuasive explanation. A manager approves it between meetings.

Who made the decision?

An approval button does not settle that question. If the person never saw the sources, alternatives, or scope of the change, they may have confirmed a choice they were not equipped to assess.

I use *Governed by Human* as a design principle. People set objectives, boundaries, and intervention criteria. Agents work within those boundaries. Human approval should represent an informed decision, with enough information and authority to reject the recommendation.

## Start with the consequences

I would classify actions by their impact and reversibility. Preparing an internal summary differs from sending a commercial proposal. Fixing a label differs from deleting an account's history. Sharing the same CRM does not give those actions the same risk.

A practical policy could look like this:

| Action | Proposed autonomy | Evidence required |
| --- | --- | --- |
| Organize information for review | Execution allowed within a defined scope | Sources and a record of changes |
| Prepare commercial outreach | Drafting allowed; sending follows the channel policy | Audience, message, and purpose |
| Change pricing or a customer commitment | Decision by an accountable person | Current terms, proposed change, and impact |
| Change the priority market segment | Leadership decision | Market evidence, delivery capacity, and alternatives |

This is a working proposal. Each business needs to adapt it to the access granted and the actual consequences of the action. A low-value field in one workflow may trigger a consequential action in another.

## Approval needs enough context

Consider a fictional company evaluating a recommendation to expand a campaign. The report shows more booked meetings. It omits that some meetings involve unsuitable accounts and that the sales team is already at capacity.

The approver needs the whole decision. I would ask for a short brief covering the proposed change, supporting evidence, remaining uncertainty, execution boundaries, and stopping conditions. It should also explain what happens if the person declines.

That final point matters. Keeping the current setup can be a reasonable choice while evidence is incomplete. A recommendation should not turn uncertainty into manufactured urgency.

The [NIST AI Risk Management Framework](https://www.nist.gov/itl/ai-risk-management-framework) is a voluntary reference for incorporating trustworthiness considerations into the design, use, and evaluation of AI systems. The commercial decision matrix here is my editorial application of that problem. It is not a certification or a NIST-published operating model.

## Define stopping conditions before execution

An agent may encounter an out-of-scope request, conflicting sources, or insufficient permissions. Those situations need a defined destination.

If a message depends on a promise absent from the approved offer, the agent should raise the question with the commercial owner. It should not fill the gap with a plausible statement. If it cannot confirm whether a record update succeeded, it should inspect the current state before repeating the operation.

Leadership also needs to assign someone to handle exceptions and define a response expectation. A policy that escalates everything without an available owner turns governance into a waiting room.

The handoff should carry the work already completed, the reason for stopping, and the decision needed. Otherwise, the person has to reconstruct the task before they can help.

## Review the boundary of autonomy

I would not treat the initial permission as permanent. Autonomy can increase as the team observes performance and understands failures. It can decrease when the offer, process, or data source changes.

To make that decision, I would track consequential errors, required interventions, review time, and recovery effort. Fast execution that repeatedly requires reconstruction may be more expensive than it appears.

I would also test whether the responsible person can still stop the workflow. A control documented in a policy but absent from execution does not provide the expected protection.

## Put judgment where it matters

This does not require a meeting for every field update. The aim is to release predictable work under clear conditions and focus human attention on choices that change commitments, priorities, or business exposure.

In the article on [coordinating AI agents](https://fabiomunhoz.com/blog/who-coordinates-ai-agents/), I discussed shared objectives and responsibilities. Governance makes those agreements inspectable: someone can explain what was authorized, what happened, and why an intervention became necessary.

Choose a workflow that already uses AI and identify its next decision that could affect a customer. **Who has authority to make it, what evidence will they receive, and how can execution be stopped?**

Clear answers turn autonomy into an operating decision the team can evaluate.

## Questions and answers

### What does Governed by Human mean?

Here it means that people define agent objectives, autonomy boundaries, and intervention criteria, with explicit accountability for decisions.

### Does every agent action need approval?

No. The proposed approach allows predictable tasks within defined boundaries and reserves approval for actions that change commitments, priorities, or business exposure.

### What does an approver need to see?

The proposed change, available evidence, uncertainty, execution boundaries, stopping conditions, and the option of keeping the current setup.
