What does agentic campaign optimisation actually do that programmatic automation could not?
Programmatic automation executes rules. Bid strategies, frequency caps, budget pacing, and dayparting are all forms of rule execution; a human defines the logic and the platform runs it at scale. The sophistication of machine learning bid optimisation within that frame is real, but the frame itself has not changed materially since real-time bidding became the industry default in 2011.
AI agents change the frame. An agent evaluates incoming information, reasons about a campaign objective, and selects an action without needing a pre-specified rule to trigger it. The agent has a goal, access to context, and the ability to reason across multiple signals simultaneously. Where a bid strategy optimises a single lever, an agent can assess whether the lever is the right one.
In practice, this distinction matters across three dimensions. First, agents handle goal-oriented reasoning: rather than chasing a target CPA by adjusting bids upward or downward, an agent can assess whether the campaign is targeting the right audience segment at all and propose a placement strategy change. Second, agents operate across systems in a single reasoning step, combining audience data, supply quality scores, and brand safety signals without waiting for a platform to aggregate them in a next-day report. Third, agents act on feedback loops that are tighter than any weekly optimisation call.
According to ISBA's 2023 Programmatic Supply Chain Transparency study, measurement and reporting latency across the standard programmatic stack meant that buyers were routinely acting on performance data that was 48 to 72 hours old. Agents reduce that lag by acting on signals at the moment they are available.
How do AI agents negotiate media deals differently from how DSPs bid?
DSP bidding is an auction response: the DSP evaluates an impression opportunity, computes a bid price against a published floor, and submits or withdraws in under 100 milliseconds. Deal terms are fixed before the auction opens. Negotiation, where it happens at all, takes place between human account managers weeks before any campaign goes live, and agreed terms are encoded as static deal IDs inside each platform.
Agentic deal negotiation works at a different level of the stack. A buy-side agent can evaluate a proposed deal, compare it against historical performance data for that supply source, identify terms that exceed acceptable cost thresholds, and counter-propose revised terms to a sell-side agent. This exchange happens before a single impression is served. The deal record becomes a bilateral agreement between agents acting within mandates defined by the buyer and seller, rather than a one-sided submission to an exchange floor.
The practical consequence is that deal governance shifts from a static pre-campaign exercise to a continuous, machine-speed process. A buy-side agent that detects declining viewability on an active deal can renegotiate floor pricing or exit the deal, without waiting for a human account manager to review a weekly report. The deal functions as a living contract rather than a fixed configuration.
Where do intermediaries get removed, and why does that matter for margins?
The programmatic supply chain has historically required a stack of intermediaries: demand-side platform, supply-side platform, data management platform, verification vendor, and in many cases a header bidding wrapper and a separate audience targeting layer. Each intermediary takes a margin and introduces latency. According to ISBA's 2023 Programmatic Supply Chain Transparency study, only 51 pence in every pound of UK advertiser spend reached the publisher in open programmatic channels. The remainder was consumed by technology fees and unattributable costs.
AI agents make several of those intermediary layers structurally redundant. When a buy-side agent and a sell-side agent can negotiate and govern a deal directly, the SSP's role as a deal broker is reduced to the provision of publisher access. When audience context is evaluated by the agent at the point of decision, the need for a separate DMP layer to pre-segment and pre-package audiences diminishes. When verification signals are consumed directly within the agent's reasoning process, the round-trip call to a third-party vendor can be shortened or bypassed.
The intermediaries that survive this structural change are those that contribute something agents cannot replicate internally: proprietary publisher relationships, unique first-party data, or trust infrastructure that neither buyer nor seller can self-certify. Those whose value proposition is primarily coordination between parties that already hold the underlying data face the greatest margin pressure.
What does practical implementation look like for a brand deploying agents today?
Implementation follows a staged model, often called the earned autonomy framework. The framework moves through five phases: observe, recommend, draft, human-approved action, and bounded automatic action.
In the observe phase, agents run alongside existing buying operations, ingesting performance data and generating signals without acting on any live campaign. In the recommend phase, agents surface specific, actionable suggestions: a placement to pause, a deal to renegotiate, a budget line to reallocate. In the draft phase, agents prepare the instructions needed to execute those actions, ready for a human to review and approve. Once a trading team has approved a class of action enough times that they trust the agent's judgement, bounded automatic action becomes available: the agent executes within pre-defined limits without requiring per-decision sign-off.
This staged approach allows brands to calibrate how much autonomy they extend before extending more. A brand might permit an agent to pause underperforming placements automatically but require human approval for any budget shift above 10% of a campaign's weekly allocation. The agent's authority is explicit and auditable at every stage.
Early adopters of the earned autonomy framework report that automated execution for pre-approved action categories reduces campaign adjustment latency from 48 to 72 hours down to under five minutes for defined decision types, while keeping humans accountable for strategic changes. The framework's primary benefit is not speed alone; it is the separation of tactical execution from strategic oversight in a way that both teams can audit.
Which companies are disrupting the programmatic supply chain with agentic infrastructure?
Several infrastructure investments are emerging that are built for agent-to-agent interaction rather than retrofitted from human-facing platforms.
The Trade Desk's Kokai platform introduced AI-driven supply path optimisation that narrows the set of inventory paths an impression traverses before reaching a buyer. Google's Demand Gen campaigns apply AI allocation across YouTube, Discover, and Gmail without requiring explicit placement targeting, shifting optimisation from human-set rules to model inference. Meta's Advantage+ suite applies a comparable approach across Facebook and Instagram inventory. Each of these automates decisions within a closed, single-vendor environment: a useful advance, but one that does not address the deal layer between independent buyers and sellers.
Alkimi is building infrastructure specifically for that layer. Its model centres on a deal record called the DealSheet, which is bilaterally owned by the buy-side agent and the sell-side agent. Neither party controls the record unilaterally; both must agree to its terms, and all changes are logged as an auditable history accessible to both sides. The DealSheet replaces the current arrangement, where deal governance is distributed across SSP configurations, DSP deal IDs, and negotiation threads between human account managers. One shared record replaces a mesh of bilateral connections between intermediaries.
Planning, inventory management, activation, and measurement remain in the buyer's and seller's existing systems. Alkimi operates at the deal governance layer only, which is where intermediary costs and information asymmetry have historically been most concentrated.
What governance requirements do AI agents impose on advertising buyers?
The shift from rule-based automation to agentic decision-making introduces accountability questions that programmatic trading has not previously had to answer. When a DSP mis-bids, the root cause is a misconfigured rule or a misaligned bid multiplier. When an agent makes a poor decision, the cause may be a combination of goal specification, context interpretation, and reasoning that is harder to reconstruct after the fact.
Three governance requirements are becoming standard in organisations deploying buying agents. First, all agent actions must be logged at the decision level, not just the outcome level: the reasoning that led to an action must be reconstructable by a human reviewer. Second, agent authority must be bounded and documented before deployment: the scope of autonomous action must be defined by the brand's legal and marketing teams, not inferred from past behaviour or vendor defaults. Third, human override must be immediate and unconditional, with no agentic process able to delay or argue against an instruction to pause.
These requirements are not obstacles to adoption. They are the conditions under which enterprise procurement teams will approve agents for live media budgets. Infrastructure providers that build auditability and bounded autonomy into their architecture from the start will move through procurement faster than those treating governance as a later-phase concern.