By Mike Conway, Field CTO, AdTech at Ocient
Walk the floor at any industry event this year and you will see the same thing I saw at Cannes and heard again over breakfast in London. Every vendor on stage has an agent. NBCUniversal is talking about automating live sports inventory with them. Amazon DSP is orchestrating full-funnel spend autonomously. The IAB Tech Lab and the AdCP camp are already arguing about how agents should transact with one another, before most agents are transacting at all.
The demos are everywhere but the production deployments are not. And the gap between the two is almost always an infrastructure problem, not a model problem.
The conversation I want to have is not just another round of “AI will change everything,” but the harder and more useful version: what does it actually take to move an AdTech agent from a controlled demo to live, autonomous spend, and what does that cost you underneath?
The tax nobody prices in
Here is the part that gets lost. When people describe agentic systems, they reach for the word “retry,” as if the agent is fumbling and having another go which undersells what is happening. An agent working a task is not failing and retrying. It is reasoning. It replaces a single human query with a loop of successful, sequential, dependent queries, 30 to 50 of them for a task of any real complexity, each one shaped by the answer to the last.
Every one of those queries is a real query. In AdTech, that means a high-cardinality join across identity, bid-stream, and outcome data, often over tables measured in trillions of rows. A human analyst asked that question once a day. An agent asks a version of it 30 times before it decides anything, and it does that across thousands of concurrent tasks.
Now put that on a consumption-priced stack. The systems teams run today, like Snowflake and BigQuery, bill you for the work. They were architected for a human cadence of queries, and their economics assume it. Point an agent at them and the loop that makes the agent smart is the same loop that compounds your bill, on exactly the workloads AdTech already runs hardest. That is the infrastructure tax. It is not the token count on the model. It is the physical scan that fires underneath every step of the agent’s reasoning, over and over, at machine speed.
The teams who have actually put agents into production have all discovered the same thing. The model was never the hard part. The economics of the data layer underneath it were.
Measurement’s reckoning is the same problem in a different suit
There is a second thread running through every upfront conversation right now, and it looks separate until you look closely. Outcomes-based measurement has become the price of entry. NBCUniversal is rolling out a unified view of delivery and full-funnel performance across linear and streaming. Disney is synthesizing attention, brand health, search, and attribution into a single brand impact metric. Retail media networks are racing to fuse panel and purchase data with media data and call it closed-loop.
Almost nobody says out loud what that promise requires. To close the loop you have to join panel, purchase, and bid-stream data at trillion-row scale, and you have to do it fast enough that the answer still matters. The closed loop is only ever as fast as the slowest join in it, and most of those joins run on data whose latency and cardinality you do not control. When panel-purchase fusion moves from a slide to production, that is where it strains. The measurement reckoning and the agentic tax are the same infrastructure question wearing two different suits. Time-to-insight is the constraint in both.
And then, the part we do not talk about until it is too late
Move agentic buying from demo to live spend and trust quietly changes definition. For fifteen years, trust in this industry meant brand safety, the question of where an ad ran. When the buyer is a machine transacting on its own, trust becomes a governance question instead.
If an agent transacted against a bad signal at two in the morning, could your organization reconstruct why in under an hour? Which signal, which join, which version of which model, which threshold it crossed.
That is not a question a standards body is going to answer for you. Auditability at machine speed is an infrastructure decision you make, or fail to make, when you design the layer the agents run on. It is a great deal cheaper to build in than to bolt on after the first incident you cannot explain.
What I want to ask the industry
I do not think the winners over the next few years will be the teams with the best agent demo. I think they will be the teams who treated the layer beneath the agent as the actual product, and who priced the tax honestly before they scaled into it.
Feel free to disagree with me. A few of the questions that I’ll be asking below if you’d like to debate:
- Whose agents are actually transacting in production today, and whose are still a very good stage demo?
- When an agent needs an answer in milliseconds, what breaks first: data latency, identity match rate, or trust in the output?
- Is agent readiness being funded as infrastructure in your organization, or shipped as a feature?
- How much of your closed-loop promise depends on data whose latency and join cardinality you do not actually control?
- If an agent transacted against a bad signal at 2am, could you reconstruct why, in under an hour, this quarter?
The best conversations we have had this year started with someone in the room saying “that is not quite how it works for us,” and whether I see you at one of the upcoming AdTech industry get-togethers or if you’d like to reach out to me via email, I am counting on you to do exactly that.