The phrase "human in the loop" has become a standard assurance in any conversation about AI-driven media buying. It is also, in most contexts, almost entirely undefined. A system that requires a human to click one button before a campaign launches technically has a human in the loop. So does a system that stops for human review at every individual negotiation round. These are not equivalent. Understanding what human approval actually means in a specific system requires looking at exactly which decisions require sign-off, what information is presented at the approval gate, and what the deal record shows about who approved what.

What is earned autonomy in AI media buying?

Earned autonomy is a model for deploying AI systems in which the scope of autonomous action expands only as the system demonstrates reliable performance within narrower boundaries. The stages run in sequence: observe, recommend, draft, human-approved action, bounded automatic action. At each stage, the system earns the right to operate with less oversight by producing good outcomes at the stage before.

Most production deployments of AI media buying sit in the middle of this progression. The agent drafts deal terms and flags them for human review. A human approves or amends before the commitment executes. The agent does not bypass that gate. The value is in the speed and consistency of the drafting, not in removing the human from the decision.

What does the approval gate actually cover?

An approval gate covers the decision point immediately before a consequential action executes. In a media buying context, the most significant consequential actions are: committing to a deal that creates a financial obligation, adjusting deal terms after a counter, and cancelling or renegotiating an existing commitment. Each of these should require explicit human approval in a governed system.

What the approval gate does not cover is equally important to understand. It does not cover background processes that the agent runs before presenting a recommendation, such as pricing analysis, inventory assessment, or brief compliance checking. Those processes are internal to the agent. The human sees the output of that processing, not the processing itself. The gate intercepts the recommended action, not every step that produced it.

What information should be present at the approval gate?

Approval without adequate information is not meaningful governance. For a human reviewer to make an informed decision at the gate, the approval interface should show: the proposed deal terms in full, the brief against which the terms were evaluated, the specific reason the agent is recommending approval at this point in the negotiation, the current budget position including existing commitments, and any flags the agent has raised about risk or uncertainty.

A system that presents only the headline CPM and a yes/no button is technically requesting human approval but is not enabling meaningful human oversight. The human cannot evaluate whether the deal is right without seeing the context in which it was generated. Approval quality is a function of information quality.

How does the DealSheet make approval concrete?

Alkimi's DealSheet is the deal record that captures what was agreed, who approved it, and when. It is not a reporting layer applied after the fact. It is the record created at the moment of approval. When a human reviewer approves a deal at the gate, the approval is recorded in the DealSheet alongside the terms at the point of approval. This creates an immutable connection between the decision and the decision-maker.

The DealSheet is bilateral: both parties to the deal have access to the same record. The buyer sees what the seller committed to. The seller sees what the buyer approved. There is no separate version of the deal terms for each side. This shared record is what makes the approval meaningful as a governance mechanism rather than an internal administrative step.

What does the deal record need to show about approval?

A deal record that is useful for governance needs to show, at minimum: which human approved the deal, at what time, on what terms, and with what information presented at the approval gate. If the deal was amended after agent recommendation but before approval, the record should show both the agent's original recommendation and the human's amendment. If the deal was rejected at the gate and a revised recommendation was generated, the record should show the full sequence.

This level of detail matters because approval disputes are rarely about what was eventually agreed. They are about whether the person who approved had adequate information and authority at the time of approval. A deal record that captures only the final agreed terms, without the approval history, cannot resolve those disputes.

How does approval scope change as autonomy is earned?

Under the earned autonomy model, the approval gate does not disappear. Its scope narrows. An agent that has consistently produced deals within brief across many campaigns might be granted the right to approve deals below a certain size or within a defined rate range without human sign-off. This bounded automatic action is not the removal of oversight. It is the extension of trust to a defined scope, with the boundary enforced by the harness and the history of performance recorded in the deal record.

The boundary matters. Bounded automatic action means the agent acts autonomously within parameters that have been explicitly defined and approved by a human. Any action outside those parameters still requires human sign-off. The agent earns the right to move faster within the defined scope. It does not earn the right to define the scope itself.

This distinction is what separates a governed AI buying system from one that is simply running without oversight. The governance is not in the model's capability. It is in the design of the approval gates, the quality of information at each gate, and the completeness of the deal record that captures every approval decision made across the campaign's life.