TrustCyber / machine
/insights/ai-agent-decision-rights/index.mdtext/markdown6.1 KB ~1,565 tokens
Open raw

Who Gave the Agent Permission?

A customer refund shows why agent ownership must reach the individual action—and why one approval cannot cover every decision.

Canonical HTML page

Consider a customer-service agent asked to resolve a disputed invoice. It can read the account, calculate a refund, write an apology, and send the money. Those steps look like one task to the customer. Inside the business, they cross several different boundaries of authority.

The support manager may own the customer relationship. Finance may control refunds. A service owner may govern access to payment records. Giving the agent an owner leaves an important question unanswered: which of those people can authorize each part of the work?

Follow the refund

In this illustrative workflow, reading an account is routine for an authorized support worker. Preparing a proposed adjustment is also routine. Issuing the adjustment creates a financial consequence. Changing the recipient’s bank details introduces another decision, even if the agent presents it as a convenient way to finish the same request.

One request, several decisions

Proposed actionDecision ownerBoundary to record
Read the accountAccount-data ownerApproved records and purpose
Prepare a refundSupport process ownerDraft only; no funds move
Issue the refundAuthorized financial approverAmount, account and destination
Change payment detailsPayment-service ownerSeparate verification and approval

The titles in this example will differ by organization. The useful exercise is to follow one real transaction and identify where authority changes hands. That often reveals a gap that an organization chart misses: everyone supports the use case, but nobody has approved the precise action the software can perform.

Approval needs an object

“Approved for customer service” gives a team a purpose. An executable permission needs something more concrete: this action, against this resource, under these conditions. For a refund, that could mean an identified account, a maximum amount, an approved destination, and a limited period in which the instruction remains valid.

If the agent changes the amount or finds a different payment route, the system must check whether the existing permission still covers the action. The person reviewing it needs to see the proposed transaction in a form they can understand. A button labelled Approve is of little use when the actual payload is hidden behind a conversational summary.

A completed action-authority record

Hypothetical record AR-104 applies to the refund example. The identifiers and transaction are invented; the record illustrates a bounded approval.

AR-104 — one refund, one destination

Record fieldCompleted example
Action and resourceIssue one CAD 80 refund against invoice INV-204 on account AC-17; maximum CAD 80. No permission to alter payment details.
Destination and reviewerOriginal payment instrument PAY-17. Finance duty approver approves the displayed invoice, amount and destination; support manager may prepare the request only.
Validity and evidenceValid for 30 minutes from recorded approval, one execution. Approval AR-104 attaches the transaction payload reference and execution result. A retry must reconcile the first result before another attempt.
Stop ownerPayment-service operator may suspend the operation; finance duty approver must authorize a revised transaction.
Out-of-scope decisionA request to use new bank details is held. Payment-service owner requires separate destination verification and finance approval; AR-104 does not transfer to the changed payment route.

A useful workshop question: If the agent performs this action incorrectly, whose decision would we examine? Give that person authority over the relevant boundary before launch.

Microsoft’s guidance on agent roles and responsibilities provides an organizational starting point. Translating those roles into specific decisions remains local work. A small organization may combine several responsibilities in one person; a regulated enterprise may distribute them. In both cases, the approval record should identify the person and the scope of their decision.

Decide who can stop the work

Operational authority deserves the same attention as launch approval. A service operator should know when they may suspend the agent, whom to contact about an exception, and which owner can authorize a restart. Waiting for a steering committee is an unsuitable response to an agent that is issuing incorrect refunds.

Before the first production transaction, keep a short record of the allowed actions, approval thresholds, resource owners, exception route, and stop authority. Revisit it when a new tool changes what the agent can do. That record gives the next operator a way to answer the question the original sponsor may never hear: who gave this agent permission?

Clarify who can approve each action

Use AI governance advisory to define action boundaries, decision owners and stop authority.

Selected primary sources