I have spent the better part of a year telling anyone who would sit still that the agentic future is coming for advertising. So it is a slightly awkward thing to admit, but here it is: the people rubbishing that future are right about almost everything. Not most of it. Almost all of it. I have read the sceptics properly, the way you read something you are hoping to disagree with, and I came away nodding.
The trouble is they have drawn the wrong conclusion from a set of facts I happen to think are correct. And the gap between the facts and the conclusion is the whole reason I still get up in the morning.
The bear case, put fairly
Let me build their argument, because a bear case you can flick away in a sentence was never worth writing about.
Start with the one that stung, because it came from someone who was actually in the room. Seb Joseph at Digiday walked the floor at CES this year and reported back that the entire event was agentic AI and little else, every stage pitching autonomous media buying, and that the buy side was conspicuously cooler than the marketing. His phrase for how buyers actually feel was that they find it more interesting than urgent. Sit with that for a second, because "interesting, not urgent" is about the most damning thing a buyer can say about a technology you are trying to sell them. It is not a no. It is worse than a no. It is a shrug.
Then the adoption numbers, which back the shrug up. McKinsey's work, reported through MarTech, put 62% of organisations still in the experimental phase with agentic AI and only 23% actually scaling it. That is a broad cross-industry read rather than an ads-specific one, so take it with the appropriate pinch, but the shape is unmistakable. Everyone is having a go. Almost nobody has handed over the keys.
And while everyone is having a go, the walled gardens carry on winning regardless. US programmatic spend is forecast to clear 220 billion dollars this year, and Google, Meta and Amazon are projected to take somewhere north of 80% of it. So the cynic's version runs like this: agents are a rounding error at the edge of a market that keeps concentrating without them, and you are asking me to get excited about the edge.
The last point is the sharpest, and it is the one the serious operators actually make. The teams genuinely winning with these tools are not the ones chasing full autonomy. They are the ones building what one write-up called governed autonomy: spend caps, approval gates, audit trails, a human hand hovering over the off switch. The reason they build all that scaffolding is that agentic systems make chained decisions, one on top of another, and an error early in the chain compounds through everything downstream before a human notices. You do not build a kill switch for a tool you trust.
Four points. I concede all four. If that is where the argument ended, I would be telling you to wait too.
What they are actually describing
Here is where I get off the bus.
There is a line from a Forbes piece that I have not been able to put down. It was about commerce agents rather than ad agents, but it travels. The author's point was that AI agents can discover and they can buy, but they cannot verify. They inherit whatever data they are handed, good or bad, and they turn it into a confident recommendation in a single pass. The confidence is the problem. The agent does not know what it does not know, and it will spend your money with exactly the same certainty whether the underlying facts were true or invented.
Read the bear case again with that sentence in your hand and something shifts. Every one of those four objections is really the same objection wearing four different coats. Interesting not urgent, still experimenting, winning only with a hand on the off switch: none of it is a complaint that agents cannot act. They can act. The complaint, underneath, is that nothing can yet prove what they did. You cannot check their working. So you keep a human in the loop, not because the human is faster, but because the human is the only audit trail you have got.
I have used this line before and I will use it again because it keeps coming true: when agents can negotiate anything, every auction quietly becomes a deal. The old world was standardised. A PMP configured once, buying across twenty sites, every parameter set in advance, and if you wanted to know what happened you read it off afterwards because it was all the same shape. The new world is an agent striking a bespoke arrangement with specific sites, in specific regions, at specific moments, for specific spend, on specific formats, and then doing it again, differently, a million times over. Same goal. Completely different machine.
And the machine has a hole in the middle of it. When every transaction is a one-off negotiation, somebody has to remember what was agreed. Not roughly. Exactly. Who offered what, on what terms, at what price, against what promise about the inventory. The standardised auction never really needed that because standardisation was its own kind of record. The bespoke deal needs it desperately, and right now it mostly does not exist. That is not a technology failure. That is a missing layer. The agents work. The ledger underneath them has not been built.
The bit I will not wave away
I want to be careful here, because this is the point where a founder usually reaches for the reassuring noise, and I would rather not.
The strongest thing the sceptics say is that vendor claims and independently verified results diverge sharply, and they are completely right, and it is embarrassing for my whole industry. Almost every autonomy number you will see quoted comes off the vendor's own best-performing account, dressed up as a general result. A self-reported figure on somebody else's account mix tells you almost nothing about yours, and the moment anyone measures the same tools across a real spread of advertisers rather than the showcase deployment, the number drops. Not by a rounding margin. By enough that "we deliver X" and "you will get X" turn out to be very different promises.
You already know the feeling, because you have sat in the meeting. The vendor puts up a slide with a number on it that would change your quarter if it were true. Someone asks how it was measured. The answer is a sentence that sounds like an answer and isn't, and everyone writes the number down anyway because the deck has moved on. I cannot stand here and tell you agentic ad buying is different, because the honest answer is that most of the numbers you will hear are self-graded homework. Mine included. When we ran our own study for the whitepaper Chandru and I put together with WPP, we said in writing that we had run it ourselves and it had not been independently verified, because that felt like the only defensible way to put a number in front of people. I would rather flag the limit than dress it up. So no, I am not asking you to trust the agents today. Trusting the agents today is precisely the mistake the sceptics are right to warn you off.
But look at what that concession actually is. "You cannot yet prove what the agent did" is not an argument for staying out of the water. It is a specification. It is a list, written by the sceptics, of the exact thing that has to get built.
The question
So if the one thing standing between an agent and your budget is the ability to prove what it did, why is the entire industry building the agent and almost nobody building the proof?
Where this lands for us
Because that is the part we decided to build.
Strip out the machinery and the argument for what we do is just this: when two parties strike a deal, both of them should be able to open the same record afterwards and see the same thing. What was agreed, what actually happened, at what price, against what promise. Not a report the seller writes and the buyer files unread. The same account, visible to both, that neither has to trust because both can check. We did not bolt that on when agents got fashionable. It is the thing the exchange was built around from the start, which turns out to matter now that the rest of the industry has arrived at the problem it was built for.
I will not put a self-graded number next to that, for the reason I just spent a section on. What I will point at is the work Chandru and I did with WPP, "Agree. Transact. Verify.", which is the argument set out in full with the mechanism attached rather than my say-so.
That is the whole thing. Governed autonomy is not a brake on the agentic future, whatever the framing suggests. It is the infrastructure the agentic future has been missing, and it is the reason the sceptics keep reaching for caps and gates and a nervous hand on the off switch. You keep a human in the loop when you have no other way to know what happened. Give the machine a record it cannot quietly rewrite, and the human stops being the audit trail and goes back to being the strategist, which is what you were paying for in the first place. The agents were never the hard part. Proving them was.
Where it is all heading
Last time out I argued the AI bears were digging in the wrong grave. Their case was really a capital-markets story wearing a technology costume: the fibre glut, the writedowns, the confident obituaries, all of it aimed at the wrong target, because the cable stayed in the ground and the internet ran on it for twenty years. The timing looked like failure right up until it very obviously wasn't. Consider this part two, because there is a second grave they are standing over.
This one is a layer along. The sceptics have looked at an agentic advertising world with no verification underneath it, watched a sensible operator refuse to hand over the keys, and concluded the technology is not ready. But the technology is fine. What is not ready is the thing that makes it trustworthy, and "the trust layer hasn't been built yet" and "this doesn't work" are very different sentences, however similar they sound in a hurry.
The bear case is not wrong. It is a punch list. And I would much rather be the one working through it than the one waiting for someone else to.
If you want the longer version of the argument, it is stitched through the last few LabTalks editions, and I am genuinely curious which of the four objections you find hardest to answer, because I am not certain I have answered all of them myself yet. Reply and tell me. That is the useful conversation.
Ben