TL;DR: Agencies are using AI in programmatic buying in two structurally different ways: AI-assisted optimisation within conventional auction-based programmatic, and agentic bilateral deal negotiation where the agency deploys an agent operating under a mandate. The first is incremental improvement to existing processes. The second is a new operating model that changes what the trading desk does and what governance it needs to have in place.

Agency use of AI in programmatic buying is real and broad in 2026, but the term covers significantly different implementations. A trading desk that has switched on a DSP's automated bidding feature is using AI in programmatic. A trading desk that has deployed a buy-side agent to negotiate bilateral deals with publisher sell-side agents under a mandate framework is also using AI in programmatic. These are not the same thing, and the governance requirements, the skills needed, and the supply chain implications are different.

AI-assisted optimisation within conventional programmatic

The most common form of AI use in agency programmatic is AI-assisted optimisation: automated systems that adjust bids, budgets, targeting parameters, and creative selection based on performance signals, operating within a conventional DSP and auction framework.

These tools have been available in major DSPs for several years. They work by predicting which impression opportunities are most likely to achieve the campaign's goal and adjusting the bid accordingly. The human campaign manager sets the goal (conversions, reach, viewability) and the constraints (budget, audience, placement restrictions); the AI tool handles the moment-to-moment bid management within those constraints.

This approach is well-established and the evidence for its effectiveness at reducing cost per outcome is strong for most campaign types. It does not change the supply chain: the buyer is still going through a DSP and one or more SSPs to reach publisher inventory, and the fee structure is the same as manual programmatic. What changes is who manages the bid-level decisions: the AI system rather than a human trader.

The trading desk's role in this model shifts from manual bid management to goal setting, constraint definition, and performance monitoring. This is a meaningful change in how traders spend their time, but it does not require the governance infrastructure that agentic bilateral negotiation requires.

Agentic bilateral deal negotiation

The more structurally significant use of AI in agency programmatic is the deployment of buy-side agents that negotiate bilateral deals with publisher sell-side agents. This is emerging in 2026, with a small but growing number of agencies building or piloting mandate frameworks and agent deployments.

In this model, the agency writes a mandate for each campaign or client: a governance document that defines the CPM range the agent is authorised to agree, the inventory scope, the audience data conditions, and the financial thresholds at which the agent must escalate to a human. The agent connects to a marketplace and negotiates deals with publisher sell-side agents within those mandate parameters, without requiring a human trader to approve each negotiation step.

The trading desk's role in this model is mandate governance rather than execution management. Traders write and review mandates, monitor the agent's decision log, review the deal records produced by negotiations, and manage the approval workflow for deals that fall outside the agent's autonomous authority. The execution is delegated to the agent; the governance is retained by humans.

Where agencies are in the transition

Most agencies in 2026 are at the early stages of exploring agentic bilateral negotiation rather than running it at scale. The common starting point is a pilot with one or two high-value publisher relationships, using a limited mandate scope, with close human oversight of every deal the agent negotiates. This gives the trading desk the experience needed to calibrate mandate parameters and build confidence in the agent's decision-making before expanding the scope.

The barrier to broader adoption is not technology. The barrier is the mandate governance capability that most trading desks do not yet have. Writing a mandate that covers all the parameters an agent needs to negotiate autonomously, with appropriate approval thresholds and clear escalation paths, is a skill that is genuinely different from DSP configuration. Agencies that are building this capability are doing so through a combination of hiring people who understand governance frameworks and retraining existing traders in mandate writing.

What clients should understand

Agency clients should understand the distinction between AI-assisted optimisation and agentic bilateral negotiation when their agency describes their AI capabilities. The questions to ask are: is the AI operating within a conventional auction framework, or is it negotiating bilateral deals? If bilateral, who wrote the mandate, and what are the approval thresholds? And what does the deal record look like?

The answers to these questions determine whether the agency's AI use is an incremental improvement to existing execution, or a genuinely different operating model that requires different governance from the client as well as the agency.