TL;DR: Earned autonomy is the principle that an AI agent's independent operating scope expands only as it demonstrates reliable, mandate-compliant behaviour over time. An agent does not begin with full autonomy and have it constrained; it begins with minimal autonomy and earns additional scope through evidence. This is the governance model that distinguishes responsible agentic deployment from automation without accountability, and it is the model that Alkimi uses to describe the appropriate progression for buy-side agents in advertising.
The phrase "earned autonomy" sits in contrast to two failure modes in AI deployment that are common in early-stage agentic advertising.
The first failure mode is full autonomy from day one: a buyer deploys an agent with broad operating parameters, no approval thresholds, and no human review loop. The agent runs without governance; its decisions cannot be audited; and when something goes wrong, the buyer has no record of why.
The second failure mode is permanent minimal autonomy: a buyer deploys an agent with such narrow operating parameters and such frequent approval requirements that the agent delivers no efficiency advantage over a human trader. The overhead of constant escalation erases the benefit.
Earned autonomy is the framework for navigating between these two failure modes: begin with minimal scope, demonstrate compliance, and expand scope incrementally based on evidence.
The five stages of earned autonomy
The earned autonomy model describes five stages that a buy-side agent progresses through as it demonstrates reliable behaviour.
Observe. In the initial stage, the agent has no authority to act. It observes deal proposals and buying opportunities, generates recommendations for human review, and produces a record of what it would have done if it had authority. This stage exists to establish a baseline: the buyer can evaluate the agent's judgement before giving it authority.
Recommend. The agent has authority to generate recommendations that humans review before any commitment is made. Deal proposals are presented with the agent's evaluation and recommendation attached; the human trader decides whether to proceed. The agent does not execute; it advises.
Draft. The agent has authority to draft deal proposals and prepare commitments, which a human reviews and approves before execution. The agent does the work of negotiation preparation; the human controls the final execution step.
Human-approved action. The agent has authority to execute deals within its mandate parameters, with human approval required for any decision that exceeds defined thresholds. The agent handles routine deal execution independently; approval thresholds route significant or unusual decisions to humans.
Bounded automatic action. The agent has authority to execute deals within its full mandate parameters, including decisions that previously required approval, based on demonstrated compliance in the human-approved action stage. Approval thresholds may be higher or fewer, reflecting the evidence that the agent operates reliably within its mandate.
How progression between stages works
Progression from one stage to the next requires evidence of reliable behaviour at the current stage. The evidence takes the form of a review of the agent's decision record: how many decisions did the agent make, how many were within mandate, how many triggered approval thresholds, and how many were overridden by human reviewers.
A progression review is not automatic: it requires the buyer to actively evaluate the evidence and decide that the agent has earned additional scope. In well-governed deployments, the progression review is documented and approved by the person with budget authority for the campaign, not only by the trader managing the account.
Regression is also possible and appropriate. If an agent at the human-approved action stage produces decisions that are systematically approaching or exceeding approval thresholds, the appropriate response may be to return to the recommend stage and review the mandate parameters before reinstating execution authority.
Why earned autonomy matters for buyers
Earned autonomy provides a governance framework that buyers can explain to their own organisations. When a board, a finance team, or a legal department asks "how do you ensure the agent acts within your intended scope?", the answer is the mandate and the earned autonomy progression model. The combination of these two instruments describes both the boundaries of the agent's authority and the evidence-based process by which those boundaries were established.
It also provides a practical path for buyers who want the efficiency benefits of agentic buying but cannot yet justify full autonomous operation to their governance structures. Beginning at the observe or recommend stage allows a buyer to build the evidence base for expanded autonomy without accepting the governance risk of deploying an unsupervised agent from the outset.