Most discussions about agentic advertising start in the wrong place: execution.

This is not because agents cannot execute advertising. They increasingly can. Give an agent a brief, connect it to a demand-side platform, set some limits and let it build a campaign. It makes for a good demonstration, but I am less convinced it solves the most valuable problem.

Programmatic execution is already highly automated. People are not deciding whether to bid $3.72 for impression number 14,382,291. Machines have made those decisions for years, so putting another machine above the existing one may remove workflow without changing very much else.

The difficult decision comes before activation

Before anyone presses buy, an advertiser has to decide who it wants to reach, what those people are worth, where they can be found and how much suitable inventory exists. It also needs a view on price, likely outcomes, channel overlap and the compromises required by a finite budget.

Answering those questions across the open internet remains unnecessarily difficult. The information sits across agencies, publishers, DSPs, SSPs, measurement companies, identity providers and a long list of specialists. Each system knows part of the answer; very few understand the complete problem.

This is where large language models and agents become interesting. Their useful role is not to bid faster, but to interrogate many sources, compare conflicting constraints and make a complicated market easier to understand.

Take a conventional brief: reach environmentally conscious parents aged 30–45 across video and display, prioritise quality journalism and premium entertainment, deliver incremental reach beyond social and avoid excessive frequency. The budget is $4 million.

An agent could test whether the audience and inventory really exist, where they overlap, how much incremental reach is available and what happens when frequency changes. It could compare publishers, apply the client's quality rules, challenge the expected price and show where the plan is unlikely to deliver. By the time the campaign is ready for activation, most of the commercially important decisions have already been made.

Better decisions are worth more than faster set-up

Adtech has a habit of treating any automated workflow as agentic. A natural-language interface can be useful; so can automated campaign set-up or deal creation. These improvements save time, which is valuable, but they are still improvements to an existing process.

Saving a trader two hours creates operational value. Identifying a media plan that cannot deliver before $4 million is committed protects the investment itself.

I would place considerably more value on the second outcome. One reduces labour in a process we already understand; the other improves a decision that advertisers still struggle to make with confidence.

Agents weaken the value of the interface

For two decades, Adtech businesses have built interfaces around their own products. DSPs built buying tools, SSPs built curation tools, measurement companies built reporting products and data businesses built audience builders. Each company naturally expected the customer to log in.

An agent may not want to use the software at all. It may only need access to what the software knows and can do. That changes the product question from whether a person can use the platform to whether another intelligence system can understand its data, constraints and capabilities.

APIs, common semantics, structured data and interoperability then become more important than another dashboard redesign. That is a larger architectural change than adding a chatbot to an existing product.

A practical division of labour

Agents will execute parts of advertising. They will create campaigns, negotiate packages, establish private marketplace deals and move budgets. Execution will still be governed by latency, financial controls, brand rules, measurement and existing optimisation systems.

Real-time bidding operates in hundreds of milliseconds, and a language model has no reason to make every auction decision itself. A more practical architecture gives the agent responsibility for the strategy and passes its instructions to infrastructure designed to process billions of transactions efficiently.

The strategic contest therefore moves upstream. Audience evidence, supply intelligence, delivery forecasts, quality signals and independent comparison will shape the recommendation before execution begins. The company that influences that recommendation sits much closer to the advertiser's budget than the company that simply carries it out.

Agentic advertising's first genuinely valuable application may never place a bid. If it can tell an advertiser what to buy, what not to buy and why, with more confidence than the current process, it will have changed far more than another automated campaign builder.