The phrase "AI agent" has been stretched so far it has lost meaning. Some vendors use it to describe a dashboard with a chatbot bolted on. Others use it for a script that runs once a week. Neither is what we mean, and neither is what your media budget deserves.

This article describes an illustrative operating model for a media agent. It is not a demonstration of a released Maaten feature or a claim that all agents operate this way.

The loop, not the moment

An AI agent is defined by a loop, not a single action. It observes, reasons, proposes, waits for a signal, acts, then observes again. The loop runs continuously. There is no shift change, no Monday morning catch-up, no quarterly optimisation sprint. It is always mid-cycle.

That loop has four stages in paid media: monitor, propose, approve, execute. A fifth stage, learn, is what separates agents that compound value from those that plateau.

Stage 1: monitor

A monitoring workflow receives available campaign signals at the refresh cadence supported by its sources. Define latency, missing-data behaviour and platform limits; do not assume continuous real-time access.

A monitoring agent can compare available signals with agreed thresholds and surface a change for investigation. The operator should see the source, time window and uncertainty.

Stage 2: propose

Reasoning is where agentic systems do work that humans genuinely struggle to match at scale. An agent can hold dozens of variables simultaneously: budget pacing against target, creative fatigue scores, audience saturation, day-parting patterns, channel-level ROAS trends. It surfaces a proposed action with a rationale.

A useful proposal identifies the campaign, suggested change, supporting evidence, uncertainty and approval required. For example: review a creative whose response has declined, check whether audience or placement changes explain it, then decide whether to test a replacement.

Specificity matters because vague recommendations cannot be meaningfully approved or rejected. They just create noise.

Stage 3: approve

This is the stage most vendors skip over in their marketing, and it is the stage that matters most for trust.

Define the permission boundary before operation. Routine supported actions may run within pre-approved limits; changes outside them require the named human approver. Test that these controls behave as intended.

You define the thresholds. The agent respects them. This is not a limitation of the technology; it is the correct design. Media buyers who trust a system are media buyers who will give it meaningful autonomy over time. Systems that grab autonomy before it is earned get switched off.

The approval interface should be fast. You are reviewing a specific proposal with full context, not digging through dashboards to reconstruct what the agent saw. One screen, one decision, thirty seconds. That is the standard.

Stage 4: execute

For supported actions, the system sends the approved request to the platform and records the response. Execution can fail or be delayed. Retries, reconciliation and human escalation remain necessary.

Automation can reduce some manual entry errors but introduce other errors through incorrect logic, stale data or failed integrations. Evaluate the complete workflow rather than claiming that automated execution eliminates mistakes.

Stage 5: learn

This is what separates an agent from a rule-based automation. After execution, the agent tracks the outcome. Did CPM stabilise? Did conversions hold? Did the creative perform better or worse than the model predicted?

Review outcomes against the proposed change. Feeding results into later analysis can improve context, but improved prediction or performance must be demonstrated rather than assumed.

What the human does in this picture

The straight answer: strategy, creative judgement, and accountability.

The agent handles surveillance, pattern recognition, and routine execution. The human sets the goals, approves non-trivial changes, makes calls the agent cannot make (a brand safety issue, a competitive response that requires market knowledge, a client relationship dynamic), and is accountable for the overall result.

This is not a threat to media expertise. It is a reallocation of it. The hours that used to go into pulling reports, adjusting bids manually, and chasing down creative trafficking can go into thinking about strategy, testing hypotheses, and building better briefs.

What to ask any vendor claiming to offer this

Four questions that cut through the noise:

  • What does the agent monitor, and how frequently?
  • Can you show me an example proposal with its rationale, not a summary?
  • Where exactly does human approval sit in the workflow, and how do I configure the thresholds?
  • How does the agent update its model based on outcomes?

If the answers are vague, you are looking at a dashboard, not an agent.

Ask for evidence of the complete workflow, including failures and approval controls. An agent label alone does not establish capability.