How do you govern AI agents in a company? Permissions, evals and owners
Governing an agent means deciding what authority it has and what evidence it needs to earn more. Every agent needs a named owner, a written permission level (read, draft, recommend, act with approval, write within limits, or run on its own), a test set built from real cases, and a way to take authority away when it fails.
Where it breaks
It rarely breaks in the demo that fails. It breaks in the demo that works. Someone sees a useful draft and says "great, ship it". That's when the agent goes from suggesting a reply to sending it, from recommending a CRM change to making it, and nobody decided that jump.
The model can do the task. What's missing is the company knowing what it allowed the model to do.
Levels of authority
An agent moves up a level when it leaves enough evidence, however excited the team is.
- Read. It sees an authorized source. It doesn't reply or change anything.
- Draft. It writes or summarizes. A person reviews before anything is sent, published or closed.
- Recommend. It proposes a decision. It doesn't execute it.
- Act with approval. It does something after someone approves, and leaves a record, an owner and a way to undo it.
- Bounded write. It touches a real system, but only within defined limits.
- Recurring operation. It runs on its own, with sampling, alerts and periodic review.
Comparing three quotes, recommending one and committing to the purchase are three different levels. Write it down that way for each agent.
The four things every agent needs
- Owner. Someone from the area that lives the workflow, who can say fix it, keep going or switch it off. If the "human in the loop" has no name, what you have is someone who'll take the blame.
- Permissions. Which sources it reads, what it can write, how much it can spend and when it has to escalate. If it only reads, you govern sources, freshness and traceability. If it acts, you govern consequences, approvals, rollback and monitoring.
- Evals. Good examples and rejected examples, an acceptance rule and a threshold. Run them before every change and sample in production. Without a threshold, all you have is hope.
- Traces. What it did, from which source, who approved. Without that you can't fix it or promote it.
What to do, in order
- List the agents and automations that already exist, including the ones people built on their own. If the company offers no approved routes, AI comes in through the back door.
- Give each one an owner and an authority level. Start low.
- Build the test set from real cases, exceptions included.
- Define what evidence is needed to move up a level and who decides.
- Define how it moves down or gets switched off, and test that once before you need it.
- Review on a fixed date. What you learn goes into the agent's instructions, not into someone's head.
The skills and instruction packs you install on an agent also give it authority. Review them the same way: when they trigger, what they touch and how they're removed.
An example
For a leadership team we built an agent that looks up board decisions, risks and follow-ups. It has permission and verification limits for sensitive material. Open access to confidential documents and any commitment on behalf of leadership were left out on purpose. More in the Board Agent case.
At a Colombian food company that replaced its marketing agency with an agent-based operation, the design had three clear roles and left out publishing without approval. See the case.
When Business Cyborgs helps and when it doesn't
If you're about to give an agent write access, spending authority or control over something customers see, we first map what it can do and who answers for it. For one workflow we do it inside an Agent Sprint: one problem, one workflow, 2 to 4 weeks, projects from USD 5,000, and we deliver recommendations on permissions and owners. For several workflows with scattered data, in a Transformation.
It isn't for you if you want autonomy without limits or if nobody inside is willing to answer for the result.
Jurisdiction Map · Agent Control Plane · Agent Sprint · Talk to a Business Cyborg
FAQ
Who should own an AI agent?
Someone from the area that uses the result and can decide to fix it, keep it or switch it off. Not the team that built it, and not the vendor.
What permissions should an agent start with?
Read or draft. Move it up one level only when tests and sampling show it has earned it.
How do you evaluate an AI agent?
With real cases, good and rejected examples, an acceptance rule and a threshold. Run it before every change and review a sample of what it does in production.
Is a human in the loop enough?
Only if that human has a name, knows what they're approving and has criteria to say no. Otherwise the review is in name only.
What about agents people already built on their own?
Inventory them and give them an approved route. Banning them with no alternative pushes them further out of sight.