What are the five stages of earned autonomy in AI media buying?

Earned autonomy in AI media buying has five stages: observe, recommend, draft, human-approved action, and bounded automatic action. An agent must demonstrate reliable performance at each stage before advancing to the next, and advancement is a decision made by the human operator, not the agent itself.

Earned autonomy: ** The model in which an AI agent earns expanded operating permissions through demonstrated performance, rather than starting with full autonomy and having it restricted after incidents.

Stage progression: ** Movement from one autonomy stage to the next requires a deliberate decision by the human operator, based on evidence that the agent's behaviour at the current stage is consistent with the operator's intent.

Mandate: ** The set of constraints within which a buying agent is permitted to operate autonomously; the mandate defines the boundaries of stage five operation, beyond which the agent must escalate to human approval.

Escalation: ** The mechanism by which an agent at any stage pauses and requests human input when it encounters a situation outside its authorised operating parameters.

Stage one: observe

At stage one, the agent observes buying activity without taking any action. It reads briefs, monitors deal negotiations carried out by humans, and logs what it sees. The agent is building a model of how the buyer's team operates: what deal types they pursue, what publishers they work with, what pricing they accept, and what they reject.

Stage one typically runs for two to four weeks. The output is not recommendations but familiarity: the agent needs enough context to make useful recommendations at stage two without constantly asking for clarification.

Stage two: recommend

At stage two, the agent surfaces recommendations for human review. A human buyer initiates a deal search; the agent recommends specific publishers, deal parameters, and pricing based on the brief and its accumulated context. The human evaluates each recommendation and decides whether to proceed.

The agent tracks which recommendations are accepted and which are modified or rejected. This feedback loop trains the agent's understanding of what the buyer's team actually values in a deal, beyond what is written in the brief. Stage two typically runs for four to six weeks before advancing.

Stage three: draft

At stage three, the agent prepares deal proposals and sends them to human reviewers for approval before they go to the publisher. The agent drafts the full deal enquiry, including specific inventory parameters, proposed pricing, volume commitment, and brand safety constraints. A human reviews the draft, approves it with or without modifications, and the approved version is sent to the publisher.

Stage three tests whether the agent can produce deal-ready proposals, not just high-level recommendations. The gap between a recommendation and a compliant deal proposal is where many agents need additional calibration. Stage three typically runs for four to eight weeks.

Stage four: human-approved action

At stage four, the agent takes actions directly but every significant action requires a human sign-off before execution. The agent negotiates with the publisher's agent in real time, proposes and evaluates counter-offers, and reaches a tentative agreement. Before committing to the DealSheet, the agent pauses and presents the final terms to a human approver.

The key difference from stage three is that the agent is now negotiating interactively, not just drafting for review. The human is approving outcomes rather than drafts. This stage tests whether the agent can navigate multi-round negotiations within acceptable parameters.

Stage five: bounded automatic action

At stage five, the agent operates autonomously within the boundaries of its mandate. It negotiates deals, commits to DealSheets, and manages delivery without per-deal human approval. Outside the mandate : deals in unauthorised publisher categories, pricing above the ceiling, volume above the cap : the agent escalates to human approval before acting.

Stage five is the operational state most buyers are working toward. It does not mean unsupervised operation: the human operator reviews the agent's deal record regularly, adjusts the mandate in response to performance, and retains the ability to pause or override the agent at any time. Autonomy is bounded, not absolute.

Frequently asked questions

How long does it take to reach stage five?

The full progression typically takes three to six months, depending on the volume of deals, the buyer's risk tolerance, and how quickly the agent demonstrates reliable performance at each stage. Skipping stages to accelerate the timeline is the most common cause of stage five failures.

Who decides when the agent advances to the next stage?

The human operator makes advancement decisions, not the agent. The agent can flag that it believes it is ready for advancement, but the decision belongs to the operator based on their review of the agent's performance record.

Can an agent be moved back to an earlier stage?

Yes. If an agent at stage five makes a decision that falls outside acceptable parameters, the operator can move it back to stage four or earlier until the cause of the error is understood and corrected. Earned autonomy can be unearned.