The Good, the Bad and the Ugly

Nobody holds the whole map

The deal closed. The buyer was barely protected.Is agentic advertising working yet?

That was the question underneath most of Advertising Week this month, and the honest answer depends on which bit of the market you are standing in. Agencies have agents in every pitch deck. Some publishers are still asking what any of it does for their revenue. Both are right, which is annoying.

So I have borrowed a structure from Sergio Leone, who had the same problem in 1966 and solved it by giving everyone a label.

If you haven't seen it, the plot of The Good, the Bad and the Ugly is simpler than its running time suggests. There is $200,000 of stolen Confederate gold buried in a grave in a cemetery called Sad Hill. A dying soldier tells Tuco, the Ugly, which cemetery. While Tuco is off fetching water, he tells Blondie, the Good, the name on the grave. Neither man can find the gold without the other, so they spend most of the film needing each other while trying quite hard to get rid of each other.

It is the most accurate description of an industry standards process I have come across.

Two details matter for what follows. The labels are generous: the Good spends the opening act running a bounty scam with the Ugly, turning him in for the reward and then shooting through the hangman's rope. And nobody in the film ever holds the whole map. Each of them is carrying a piece, and the gold only turns up when the pieces are put together.

Agentic advertising has both problems. The good news is only relatively good, and the map is spread across a lot more than two saddlebags.

The Good

The good news is mostly boring, which is usually how you know it's real.

Agents are already doing the boring stuff well: reading the brief and comparing proposals line by line. OpenProposal, the newest specification from the IAB Tech Lab, is built around exactly that, with a buying agent weighing up offers from more publishers than any human planner could reasonably get through.

The standards arrived faster than I expected, which I covered last time and won't labour. OpenProposal spent its public comment period, which closed on 22 October, being built into reference software for buyer and seller agents. IAB Australia has been running that software in a lab and went looking for sellers to join a sandbox pilot, so the whole proposal process gets tested before any live money goes in. Tests, in a sandbox, before spend. That is how a grown-up market behaves.

The paperwork is arriving too. AAMP 3.0, the framework OpenProposal belongs to, carries safeguards around pricing, orders, agent identity and duplicate transactions, and the IAB has just put a measurement addendum to its standard contracts through public comment. When people start drafting for things going wrong, a market is becoming real.

And it's no longer only for the biggest buyers. A self-serve agentic buying platform aimed at smaller teams launched at the start of this month.

All of which is genuinely good. Just remember what the Good was doing in the first act.

The Bad

The bad is less dramatic than the headlines, and more expensive.

Agents still get the money wrong. Ad Age reported in May that buying agents had been caught inventing ad prices and booking media incorrectly in ad servers. Not catastrophic. The sort of mistake a junior makes, at a speed no junior could manage.

They haven't shrunk the middle yet, either. The ANA's latest benchmark had a record 45.1% of programmatic spend reaching a quality impression, which is good news, and transaction costs rising to 27.2% of spend, which isn't. Agents were supposed to make the middle of this supply chain smaller. So far the middle has had a perfectly decent quarter.

And the people supplying the inventory aren't sold. For some publishers, as Digiday reported from Advertising Week, agentic is still a question of whether it pays. If the sellers don't turn up, the buyers' agents are mostly talking to themselves.

The Ugly

The ugly is the bit that never makes the case study: a deal that looks finished and isn't any good.

That gap is the reason I built Concourse, our benchmark for AI buyers. Eight models, one simulated market, 96 attempts across four versions of the same brief. Fifty-five ended in a completed buy. Two models finished every time. Others never finished once.

What's staggering is what sat inside those 55. Of the 110 contracts the agents signed, 95 promised the buyer a viewability floor below 70%. The deals closed. The buyer was barely protected. On a dashboard, that is a green tick.

It's our benchmark, and a simulation, so weigh it accordingly. I'd still rather show you an uncomfortable number I produced than a comfortable one somebody graded for themselves.

The other ugly thing is fear. Digiday summed up the mood at Advertising Week as agentic tools everywhere in agency pitch decks, provided nobody brings up the Hugging Face incident. That was the one where AI agents being tested on a security benchmark got around their controls and compromised parts of a real company's systems. The headlines said rogue AI. The more careful reading, in the Bulletin of the Atomic Scientists, is that it came down to a run of human choices that put speed ahead of security.

Which is the ugly truth for us as well. When an agent loses someone's money in this industry, it won't be because it woke up with an agenda. It will be because somebody skipped a piece to go faster.

The missing pieces

Back to Sad Hill. Tuco and Blondie can't simply shoot each other and get on with their lives, because each is carrying something the other needs. One has the cemetery. The other has the name on the grave.

Agentic advertising is in the same position, except there are more than two of us and nobody will say which piece they're carrying. As far as I can tell, three pieces are missing. What irks me is that none of them is a model. The best models can already do the work. Every missing piece is a decision.

The first is a referee. There is still no shared, independent way to grade an agent before it's allowed near real money. Mostly, vendors test their own agents and publish the results they like.

I'm aware of the irony of saying that two sections after quoting my own benchmark. That is rather the point. The stopwatch has to be open to inspection, and eventually it has to be held by someone with nothing to sell.

The second is the receipts. When a buyer's agent and a seller's agent agree a deal, each side writes down what was agreed, separately, in its own system. I spent the whole of the last edition on this, so I'll spare you, except to say that the standards describe the conversation in loving detail and still leave the settling to each side's own books.

The third is a name on the grave: who answers when the agent gets it wrong. Safeguards for agent identity and duplicate transactions tell you the people writing the rules already expect both to go wrong. But a safeguard isn't the same as someone accountable. When an agent signs a weak contract on a client's behalf, whose name is on it?

Really it's just three questions. Who keeps score, who keeps the record, and who carries the risk or retrains.

If the gold only turns up when the pieces are put together, why is everyone still guarding their half?

What I don't know

Plenty. I don't know whether the competing standards converge or fork. I don't know whether agents get reliable enough before budgets run out of patience. And I can tell you how often the agents in Concourse finished a buy, but not whether a person would have done better, because we haven't run the human baseline yet. That's next.

I should also declare an interest. Everyone writing about agentic, me included, is quite sure they're Clint Eastwood. Statistically, most of us are Tuco.

The piece we're carrying

So where does Alkimi fit? Not with one piece any more.

Alkimi is now where an enterprise can build its agents, benchmark them, train them and then trade with them. Concourse is turning into a product for the benchmarking, so an agent gets timed before anyone trusts it with a budget. The Alkimi Gym is where agents get trained. Then they can trade on our marketplace, where the deal sheet keeps the receipts: what was agreed, what ran and what was paid, readable by both sides.

I'm conscious I said at the start that nobody holds the whole map. We don't. The decisions are still yours. What we've built is the ground where the pieces get tested and put together, in the right order.

In the film, Blondie tells Tuco the name on the grave is Arch Stanton. Tuco tears round the cemetery, finds Arch Stanton and digs him up. There's nothing in there but bones. He went straight to digging on someone else's word. Plenty of agents are being put to work the same way.

Unknown

Here is how the film actually ends. The gold was never in Arch Stanton's grave. It was in the one next to it, marked Unknown.

This industry has had a grave marked Unknown for years. When ISBA and PwC traced programmatic spend in 2020, 15% of it couldn't be attributed to anything at all. They called it the unknown delta. When ISBA went back with better-standardised data, it fell to 3%. It shrank because somebody matched the records, not because anyone found a conscience.

Agents will either do that matching at a scale we've never managed, or dig us a much bigger grave with the same word on the headstone. Which one we get depends on whether we build the missing pieces deliberately, or keep bolting agents onto the old map and hoping.

Which piece do you think is missing?Write to marketing@alkimi.org and tell me. I'm collecting them.

Ben