9 Nov 2026 · 4 min read

AI in Programmatic Advertising: Automated, Assisted or Untouched

Inside a modern programmatic buy, artificial intelligence already completes several tasks start to finish without a person checking each action: pricing a bid on a single auction, pacing a budget across a flight, and generating creative variants for testing. Other tasks are AI-assisted, meaning a system proposes a decision and a person approves or overrides it before it takes effect, such as which supply route to send spend through or whether to grant a brand safety exception. A third group, chiefly contractual terms, final budget sign-off and dispute resolution, stays entirely manual. This matters because vendors routinely describe all three categories as AI-powered, and a buyer who cannot tell them apart cannot judge what a claim actually promises.

What does automated actually mean here?

An agent, in this context, is software that takes an action on behalf of a buyer or seller within rules it has been given, rather than simply surfacing information for a person to act on. The model behind it is the underlying system, typically patterns learned from past auction and performance data, that produces a price, a ranking or a recommendation. The useful test for whether a task is automated or merely assisted is whether a human reviews the output before it takes effect, not how sophisticated the model is. A model recalculating a bid tens of times a second, checked only in aggregate afterwards, is automating that task. A model ranking supply routes and waiting for a buyer to pick one is assisting it.

Which tasks are genuinely automated today?

Real-time bid pricing is the clearest case: a model sets a price for each auction request from signals such as the publisher, the device and recent performance, with no person reviewing that decision before submission. Pacing adjusts delivery speed across a flight so a budget neither exhausts early nor underspends, recalculated continuously. Creative optimisation, serving different ad versions and shifting weight toward whichever performs best, runs on the same logic. A human sets strategy and limits in advance, then reviews results after the fact rather than approving each action.

Which tasks are AI-assisted but still need a person?

Supply path optimisation, choosing which route an impression travels through intermediaries to reach a publisher so a buyer pays less and the publisher keeps more, is typically assisted: a model ranks paths by cost and fill rate, but a person decides whether to act on an unfamiliar recommendation or a flagged anomaly. Brand safety works similarly: automated scanning flags and scores risky content, but a person reviews borderline cases, particularly blocklist exception requests. Deal terms on larger private agreements follow the same pattern, a system proposes a price from historical benchmarks, but a person agrees final volume and price, because someone has to answer if the deal underperforms. This is where deal governance shows up: a timestamped record of what both sides agreed, held independently by each party's system, exists because a recommendation is not the same as an agreed deal. A human holds approval rights at that point.

Which tasks remain untouched by AI?

Legal sign-off, final budget authority, negotiating direct first-party data access with a publisher, and resolving a dispute when two parties disagree, all stay manual, requiring someone to carry accountability for the outcome in a way a model cannot.

Why does this split matter when evaluating a vendor's AI claim?

A protocol, a shared technical standard letting a buyer's software and a seller's software exchange structured information without a person translating by hand, only standardises how systems communicate. It does not determine which decisions get taken without review internally, that remains a choice each vendor makes. The IAB Tech Lab's AAMP framework and AgenticAdvertising.org's AdCP standard both address this communication layer, and neither answers where a given vendor draws its own automation line. The marketplace that Alkimi operates works within AAMP today and is being built toward AdCP compatibility, a reasonable position given neither protocol settles the automation question on its own.

The practical test for a buyer is simple: ask whether the vendor's system executes the final action on its own, or a person must approve it first. Benchmarking how an agent performs against defined tasks, rather than relying on a vendor's own description, is one way to check the answer before committing. Alkimi's Concourse is built for exactly that, a benchmark testing agent behaviour under live conditions rather than taking a vendor's claim at face value. The split between automated, assisted and untouched is the first question worth asking about any AI claim in programmatic media.

All articles

Speak to the team

Request a demo

You don't need to plug in your agent, the demo walks you through the steps.