Give every AI agent a decision-rights contract

An AI agent needs an explicit decision-rights contract before it receives authority to act: define its permitted work, the human commitment point and the conditions that stop execution.

Why is “human in the loop” an incomplete instruction?

A manager can approve an elegant summary without knowing that the agent has already updated the customer record. Another can review a recommendation without noticing that it was produced from material the receiving team cannot access. “A human will check it” does not specify which action is being checked, when it happens or whether the reviewer can still reverse it.

Microsoft’s 2026 Work Trend Index frames the shift toward agents around human agency and organizational redesign. Its survey evidence describes reported experiences, not proof that one authorization pattern works everywhere. The practical management question is narrower: which actions create a commitment, and who has the right to make each one? MS26

Split the work into prepare, recommend and commit

Preparation gathers authorized inputs, compares evidence and drafts an output. Recommendation proposes a choice with assumptions and alternatives. Commitment changes the world: sending a message, granting access, booking spend, changing a promise or publishing an instruction. A system can do the first two well and still be the wrong owner of the third.

The contract should name the business owner, permitted data, allowed actions, forbidden actions, approval checkpoint, recipient and stop condition. It is an operating agreement between the process owner and the people building or configuring the workflow. It is not a substitute for platform permissions, security controls or an organization’s legal obligations.

Put review before the action that matters

Morgan Stanley’s June 2024 Debrief announcement describes a concrete separation: client consent for meeting notes, action items after the meeting, and an email that the advisor edits and sends at their discretion. This is a company-described workflow, not independent evidence of its effectiveness. It illustrates why a draft and a send action should not share an ambiguous status. MS24

For a commercial team, apply the same distinction to an account-review brief. AI can organize the account facts and propose a follow-up. The account owner confirms any promise before sending. The exact boundary depends on the stakes and reversibility; it should not be inferred from a button labeled “continue.”

Make approval an informed choice, not a rubber stamp

The reviewer needs the proposed change, its audience, the relevant sources and the exceptions that could matter. Approval should refer to an identifiable version. Require fresh approval when the recipient, commitment or material underlying evidence changes; the previous approval covers a different action.

Automation remains useful when the action is narrowly scoped, reversible and objectively checked. The strongest alternative to repeated approval is not unlimited autonomy; it is a permission envelope with monitored exceptions. Use that alternative only after the owner can show what is permitted and what happens at its boundary.

Start with a shadow workflow and test its stop conditions

Next week, the workflow owner and technical control owner can select one routine management task and write its decision-rights contract. Run it in a shadow mode that proposes actions without sending, spending or changing access. Test a missing source, an unauthorized recipient, a changed draft and a failed downstream step. Record whether each produces the expected stop.

The contract also needs an action-instance identifier and an outcome receipt. Approval of a version does not prevent the same action from being executed twice. If a downstream result is unknown, stop and reconcile the receipt before retrying; a timeout is not evidence that nothing happened. The technical owner must test duplicate requests and ambiguous outcomes as separate cases.

Revisit the contract after the first week of representative tasks. A single unauthorized commitment is a stop signal for that action path. If most low-risk proposals require the same predictable approval, investigate a narrower preauthorized envelope rather than remove review across the entire process. The contract is useful only when the real workflow and its permissions enforce it.

Questions managers ask

What is a decision-rights contract for an AI agent?

A concise operating specification naming the owner, permitted inputs and actions, commitment point, approval rule and stop conditions. The technical workflow must enforce it.

Must every agent action need human approval?

No. A process owner can preauthorize a narrow class of low-risk actions when permissions, checks, exception handling and reversal are reliable. Consequential commitments require the authority specified for that workflow.

Does a review button establish safety?

No. Review must happen before the consequential action and cover an identifiable output, recipient and scope. A button alone does not prove authorization enforcement.

Sources and scope

MS26: Agents, human agency, and the opportunity for every organization — Microsoft WorkLab; 2026-05-05. Primary vendor-sponsored research; public report.

MS24: Launch of AI at Morgan Stanley Debrief — Morgan Stanley; 2024-06-26. Primary company launch announcement.

MS-EVAL: Morgan Stanley uses AI evals to shape the future of financial services — OpenAI; Undated; accessed 2026-10-27. Primary vendor/customer account; commercial interest disclosed.

Research checked 2026-10-27. Next review 2026-11-27. Suggested pilots are editorial proposals, not observed Upskill Pro customer results.