AI Agent Governance Starts With Five Decision Rights
AI Agent Governance Starts With Five Decision Rights
Say your support team wants an AI agent to handle refund requests. The agent can read orders, look up a customer's history and update the ticket. It has every permission it needs to do the job.
Then a customer writes in about an order that's two days past the return window. The package arrived damaged, they say. Can the agent refund it? Can it offer store credit instead? Can it decide the return window doesn't apply to damaged goods?
The agent's access doesn't answer any of that. Someone has to decide it ahead of time. If nobody has, the agent ends up making a policy call that no one handed it.
Teams usually begin AI agent governance with access controls, approval gates, audit logs and risk categories, and all of those matter. But each of them assumes someone has already settled who may decide what. An approval gate on the refund agent still needs to know whose approval counts.
Until that's settled, "human in the loop" is a reassuring phrase with no named person, no named decision and no point in the work where anyone steps in.
Settle five decision rights for the refund work before the agent starts, and you won't need a person behind every action it takes. We call the result the Authority Map.
Permission is not authority
Four words get used as if they mean the same thing, and they don't.
Permission is what the agent's account can technically do: read the order, update the ticket, send a refund through the payment system. Approval is a recorded yes or no on one specific thing, like this refund for this customer. Authority is the standing right to make a kind of decision, such as approving refunds that fit the written policy, up to an amount finance has set. Accountability belongs to the person who answers for how that authority was set up and used.
The refund agent can have permission to send money back without anyone giving it the authority to. That's easy to end up with, because the key that connects it to the payment system knows nothing about your refund policy.
Anthropic's April 2026 piece on trustworthy agents puts the risk plainly: an agent "directs its own processes and tool use," so there is "more room for them to misread users' intent and take actions with unintended consequences." Part of its answer is letting users choose which tools to enable and set permissions such as always allow, needs approval or block. That limits what the refund agent can touch. It doesn't say who wrote the policy the agent applies, or who may bend it.
Asking a person every time isn't the fix either. In a May 2026 engineering post on containing Claude, Anthropic reported that Claude Code users approved roughly 93% of permission prompts, and that "the more approvals a user sees, the less attention they pay to each." A click records that someone said yes. It doesn't show they decided anything. Anthropic's response in Claude Code included a sandbox the agent can work inside without asking, which it reports cut permission prompts by 84%.
The refund agent's sandbox is the written policy and the limit finance has set. If finance and the head of support drew those lines, they're the people to go to when one turns out to be in the wrong place.
Five questions for every agent workflow
An Authority Map answers five questions:
- Who sets direction? Someone decides what the workflow is for, which policy it follows and how much risk is acceptable.
- What may the agent decide on its own, within its scope, its tools and its limits?
- What do people keep? Some decisions need a person's approval, judgment or sign-off, and the map says whose.
- Who owns exceptions and outcomes, and where do problems go when the agent can't settle them?
- Who can pause it, narrow it, take away its access, hand it to someone else or retire it?
You don't answer these for every keystroke. Answer them for the kinds of action that carry real consequence: sending money, telling a customer something the company has to honor, changing a record someone else relies on.
NIST's voluntary AI Risk Management Framework asks for several of these answers. Roles and lines of communication should be documented and clear to the people involved, so support, finance and IT each know their part of the refund agent. The people overseeing it, here the support lead and the finance approver, need defined roles. And there should be a safe way to phase the system out, which is question five.
Who sets direction
Direction belongs to the person allowed to say why the agent exists, and it's often not the person who set it up.
In the refund workflow, the head of support might own how customers get treated. Finance might own the refund limits, even though support runs the agent day to day. Whoever in IT connected the agent to the order system may own neither. If you ask "who's in charge of the refund agent?" and get three different names, that's normal. The map just has to say which of them owns which part.
Direction should answer questions like these:
- What is the workflow for? Resolving requests quickly within policy, or keeping refunds down? Those can pull against each other.
- Which policy governs it, and which version?
- What's already decided, such as the return window and what counts as damaged?
- Which tradeoffs may the agent make, like refunding now versus asking for a photo first?
- Who may change any of this later?
That last question is the easiest one to skip. If any support rep can type "approve these, they're good customers" into the agent's instructions and have it stick, you don't have one governed workflow. You have a shared tool taking orders from whoever spoke last.
The OECD AI Principles, updated in May 2024, say that those who develop, deploy or operate AI should be accountable for it working properly, "based on their roles, the context, and consistent with the state of the art." Nobody can hold "the AI" accountable for a refund. The person who owns the refund policy can be.
What the agent may decide
Teams often sum this up in one word, "autonomous" or "supervised." A more useful line separates how the work gets done from what it's for. The refund agent can choose its method: which records to pull, when to ask for a photo, how to word the reply. People keep the choices that change the purpose, the cost, whether it can be undone or what the company has promised the customer.
Your refund agent might be allowed to:
- pull the order, shipping and payment records;
- check the request against the published refund policy;
- ask the customer for a photo when they report damage;
- draft the reply;
- approve refunds that fit the policy, up to the limit finance has set.
And not allowed to:
- change the refund policy;
- make up an exception because the customer sounds upset;
- refund above the limit;
- promise credit on a future order;
- share another customer's details;
- change its own access.
OpenAI's white paper on governing agentic AI systems says some decisions "may be too important for users to delegate to agents," such as "independently initiating an irreversible large financial transaction." One refund is far smaller than that. A batch of wrong ones adds up, and money already back in customers' accounts is hard to recover.
So the closer a decision gets to money leaving, a promise to a customer, private data or a change in policy, the more likely it sits outside the agent's discretion. The refund agent can still approve refunds under finance's limit, because someone set that limit on purpose.
What people keep
"Human in the loop" doesn't say which human, or what they're deciding. Go back to the late refund for the damaged package. The support lead might decide the damage claim is credible. If the amount is over the limit, finance might still have to approve the money. That's two decisions by two people, and a single "approve" button can blur them into one.
Those decisions also land at different points in the refund. Before the agent starts, the head of support signs off on which requests it handles and what it may approve. Partway through, the support lead answers its question about the damage photo, then reads the reply before it goes out. Finance authorizes the money if the amount is over the limit.
Researchers have found oversight spread out like this in software work too. In a small, exploratory study, Shipi Dhanorkar, Samir Passi and Mihaela Vorvoreanu interviewed 17 experienced developers about working with software agents and describe at least four forms of oversight: setting controls up front, planning the work with the agent, watching it run and reviewing what it produced.
For the refund workflow, the reserved decisions come down to questions like these. Is this still the right goal? May the policy bend for this customer? Is this reply good enough to send? May this money go out? Has the agent earned a higher limit? For each one, write down who holds it, what they need to see, how long it can wait and what happens if nobody answers. Skip that last part, and the late refund sits in finance's queue until the customer writes in again.
Who owns exceptions and outcomes
The refund agent will run into things nobody planned for. An order record is missing. Two policies disagree. A customer asks for something the store doesn't offer. Governance breaks down when the agent can spot the problem but has nowhere to send it, or when everyone can step in and nobody owns closing it.
An exception path should say:
- what forces the agent to escalate;
- what evidence goes with the escalation;
- who can resolve it;
- what happens to the customer while it waits;
- whether the answer is a one-time exception or a new rule;
- who tells the customer what was decided.
Owning the outcome doesn't mean doing the work by hand. The operations lead accountable for refunds doesn't need to read every routine ticket. They do need to own the policy, the limits, how the agent's work gets reviewed, the exceptions nobody has resolved and the decision to keep the workflow running or change it.
Logs help here if they record the right things. Say the damage photo turns out to come from another order. The operations lead needs to piece back together which instruction the agent followed, which records it read, what the support lead decided and who approved the money. The OECD principles ask for this kind of traceability, covering "datasets, processes and decisions made during the AI system lifecycle." A log that holds the people's decisions as well as the agent's actions tells the operations lead whose call to revisit.
Who can stop it
The refund agent will outlast the way it was first set up. Whoever launched it may move on. The workflow can grow to cover exchanges, the payment system can change, and sometimes a run of bad refunds shows that finance's limit was wrong from the start.
Someone has to hold each of those rights, and everyone should know who before they need it. The OECD principles ask for mechanisms so that AI systems that "risk causing undue harm or exhibit undesired behaviour" can be "overridden, repaired, and/or decommissioned safely as needed." In the refund workflow, the operations lead might hold that right, with the head of support able to pause the agent the moment refunds start going out wrong.
Build one map, then test it on a bad day
Start with one workflow that matters, not a company-wide policy. Refunds are a good first choice because the consequences are concrete. The same questions work for publishing a campaign, reviewing a vendor or updating a system.
Fill this in:
Workflow and what it's for:
Who owns it:
- Who sets or changes direction?
- What may the agent decide without approval?
- Which decisions do people keep, and who holds each one?
- Who owns exceptions, incidents and the final outcome?
- Who can pause, narrow, revoke, hand over or retire it?
What each person needs to see before they decide:
Permissions and limits that put this map into practice:
Review date, and what triggers an earlier review:
Then test it on the days things go wrong. The order record is missing. The finance approver is on vacation. Two policies conflict. The support lead changes jobs. A batch of refunds goes out wrong. If the map doesn't say what happens next, that's the gap to close before you give the agent more room.
Where Supanova fits
Say your refund work runs on Supanova, with atoms (our AI workers) doing the agent's job. Several of the map's answers become settings. You choose how often your workforce checks in: every step, every task, or once per project. Until the support lead trusts the replies, the refund work can check in at each task. Later, green-light a project once and your team runs it end to end. Finished work comes back for your review, and nothing ships without you.
The atoms also need your payment system. Connect it as a Bond: give us its website and your key, and we read its docs, prove what your key can actually do, and only then let an atom touch it. Like the refund agent's key at the start, your key knows nothing about your refund policy, so whether an atom may send refunds at all, and which ones, is still up to the people who own the policy and the limit.
And if a batch of refunds goes out wrong, you can stop everything in one action. Your map says who should be the one to do it.
Pick the agent workflow that worries you most and walk one hard case through it, the way we walked the late, damaged package through the refund agent. Write down who answers each of the five questions, including which decisions stay with a person however well the work has gone. If every line has a name, you have governance you can put into practice. If one is blank, start there.
When you're ready to put a workforce to work inside those decisions, start at Supanova.
Sources
- Trustworthy agents in practice, Anthropic, April 9, 2026.
- How we contain Claude across products, Anthropic, May 25, 2026.
- AI Risk Management Framework Core, National Institute of Standards and Technology.
- OECD AI Principles, OECD, adopted May 2019 and updated May 2024. See Accountability (Principle 1.5) and Robustness, security and safety (Principle 1.4).
- Practices for Governing Agentic AI Systems, Shavit, Agarwal, Brundage and colleagues, OpenAI.
- Human oversight of agentic systems in practice, Shipi Dhanorkar, Samir Passi and Mihaela Vorvoreanu, arXiv, June 3, 2026.