TL;DR: AI agents help with media buying in two distinct ways. As optimisation tools within conventional programmatic, they reduce cost per outcome by making faster and more precise bid-level decisions than a human can at scale. As mandate-governed negotiation agents in bilateral deal markets, they remove the execution burden from traders entirely and handle deal negotiation with publisher sell-side agents, producing a shared deal record that supports cleaner post-campaign reconciliation.

The case for AI agents in media buying is not primarily about replacing human traders. It is about changing what human traders spend their time on, and what the media buying supply chain looks like for inventory categories where bilateral negotiation is more appropriate than open auction.

AI agents are good at specific tasks: evaluating a structured proposal against a set of defined parameters, adjusting a bid based on a performance signal, and executing a repetitive decision at a speed and scale that a human cannot match. They are not good at the tasks that require contextual judgement: setting a campaign strategy, writing a mandate that captures campaign intent as testable parameters, or deciding whether a deal that falls outside mandate parameters is worth an exception.

The division of labour that works is: agents handle execution, humans handle governance.

Where agents add value in conventional programmatic

In conventional auction-based programmatic, AI agents help in three specific ways.

Bid management at scale. A programmatic campaign may involve millions of individual bid decisions per day. Each decision involves evaluating an impression opportunity against the campaign's goals and constraints and determining an appropriate bid price. A human cannot make these decisions individually; a rule-based system can make them, but within fixed rules that do not adapt to changing conditions. An AI agent evaluates each opportunity against a performance model and adjusts bids dynamically, finding the optimal price for each impression given the campaign's current performance against its goal.

Budget pacing. Spending a campaign budget evenly across a flight period, while concentrating spend in moments of higher opportunity, is a pacing problem that AI handles better than manual management. An agent that monitors delivery rate, remaining budget, and remaining flight time can adjust spend rates continuously to avoid both underpacing (budget not spent) and overpacing (budget exhausted early).

Audience targeting precision. AI optimisation tools use performance signals to identify which audience segments, contexts, and times of day are producing the best outcomes against the campaign goal, and shift spend toward those conditions automatically. This is faster and more adaptive than a human reviewing performance data weekly and making manual adjustments.

Where agents add value in bilateral deal negotiation

In agent-to-agent bilateral deal markets, the agent's role is qualitatively different. The agent is not optimising within a fixed market structure; it is conducting a negotiation with a counterparty agent.

Executing negotiations at scale. A buy-side agent can run parallel negotiations with multiple sell-side agents simultaneously, across multiple publishers, at a speed no human team could match. A trading desk with access to bilateral agent negotiation can evaluate more publisher opportunities and complete more deals in a campaign period than a team running manual negotiations would be able to.

Applying mandate parameters consistently. A human trader applying mandate parameters is subject to variation: they might accept a deal slightly above the mandate ceiling because the relationship is important, or decline a deal that is technically within parameters because of an intuition about the inventory. An agent applies mandate parameters consistently, which makes the mandate a genuine governance document rather than a guideline.

Producing deal records automatically. Every bilateral deal an agent negotiates results in a DealSheet: a structured record of the agreed terms that both parties hold. A human-managed bilateral deal produces an email chain and an IO. The DealSheet is more complete, more structured, and more useful for post-campaign audit than any documentation a manual process is likely to produce at scale.

What agents do not replace

Agents do not replace the human functions that require judgement, context, and strategy.

They do not write their own mandate. The mandate is the governance document that defines what the agent is authorised to do; it requires a human who understands both the campaign's strategic intent and the operational parameters that translate that intent into testable rules.

They do not determine campaign strategy. What inventory to target, what audience to reach, what creative context is appropriate for the brand: these are strategic decisions that require human input, not agent execution.

They do not resolve governance gaps. When an agent encounters a deal that falls outside its mandate parameters, it escalates. A human must decide whether to approve the exception, update the mandate, or decline the deal. The agent flags the gap; the human closes it.

The most effective use of AI agents in media buying is treating them as execution infrastructure with clear governance boundaries: capable within their mandate, dependent on humans for anything outside it.