TL;DR: AI agents in advertising are software systems that can make decisions and take actions autonomously within a defined scope, without a human authorising each individual step. In advertising, they fall into three functional roles: buy-side agents that act on behalf of advertisers, sell-side agents that act on behalf of publishers, and infrastructure agents that manage the transaction layer between them. Understanding which role an agent plays changes how buyers evaluate the claims made about it.
The term "AI agent" is used in advertising to describe a range of things that differ significantly in what they actually do. At one end, it describes a sophisticated algorithmic bidding system that automates bid submission. At the other, it describes an autonomous system that negotiates deal terms, applies mandate constraints, escalates decisions to a human principal, and produces a structured record of everything it agreed. Both are called agents. They are not the same thing.
A useful working definition: an AI agent in advertising is a software system that operates from a structured mandate, makes decisions within the scope of that mandate autonomously, escalates decisions that fall outside the mandate to a human principal, and produces an auditable record of what it decided and why. By that definition, the mandate is not an optional feature; it is what makes the agent an agent rather than a script.
What does a buy-side agent do?
A buy-side agent acts on behalf of an advertiser or agency. It operates from a mandate that the buyer has approved: a set of parameters that defines the inventory it can target, the price it can pay, the audience it can reach, the creative it can serve, and the approval thresholds that require human sign-off before the agent proceeds.
Within those parameters, the buy-side agent makes decisions autonomously: it evaluates inventory opportunities, submits bids or negotiates deal terms, adjusts pacing based on performance, and applies brand safety constraints at the impression level. Where a decision falls outside the mandate parameters, the agent escalates to the appropriate human principal and waits for authorisation before proceeding.
The defining characteristic of a buy-side agent, as distinct from algorithmic buying automation, is the mandate. Algorithmic systems apply rules; agents apply mandates. The difference is that a mandate is a governed document, approved by the budget owner, that can be audited after the fact to verify that the agent acted within authorised scope.
What does a sell-side agent do?
A sell-side agent acts on behalf of a publisher or inventory owner. It operates from a sell-side mandate: a set of parameters that defines the inventory it can offer, the price floors it must enforce, the deal types it can accept, and the audience segments it is authorised to make available.
When a buy-side agent proposes deal terms, the sell-side agent evaluates whether those terms fall within its mandate. If they do, it confirms the deal. If they do not (the offered price is below the floor, the targeting segment is restricted, the format is not available), it declines or proposes alternative terms. The resulting negotiation produces a deal record that reflects the terms both agents agreed on.
Sell-side agents are less developed in 2026 than buy-side agents, because publisher-side adoption of mandate frameworks has lagged behind advertiser-side adoption. As SSP-level AdCP integration expands through the Prebid adapter and direct integrations, sell-side agent deployments are accelerating.
What do infrastructure agents do?
Infrastructure agents manage the transaction layer: the protocol layer where deal messages are exchanged, the record layer where agreed terms are stored, and the verification layer where delivery is compared against the deal record. These agents are not directly instructed by advertisers or publishers; they operate as part of the marketplace or exchange infrastructure.
The mandate enforcement agent is the most important infrastructure role: it is the system that checks, at the point of a buy-side agent attempting to execute a deal, whether that deal falls within the buy-side mandate. Without a mandate enforcement agent, the mandate is a reference document rather than an operational constraint.
How should buyers use these definitions?
The taxonomy is useful for evaluating vendor claims. A vendor claiming to offer a "buy-side AI agent" should be able to describe the mandate interface, the escalation workflow, and the audit trail. A vendor claiming to support "agent-to-agent negotiation" should be able to describe the sell-side mandate framework, the protocol used for message exchange, and the deal record format. A vendor that uses "agent" to describe algorithmic optimisation without a mandate layer is selling a less capable product under a term that has come to imply more.
The definitions also clarify the scope of human oversight. A buy-side agent with a well-defined mandate does not remove the buyer from the process; it relocates the buyer's involvement to the mandate governance level rather than the individual transaction level. The buyer is still deciding what the agent is authorised to do; they are just not making each decision in real time.