You Hired an Interpreter. You Needed a Negotiator.

There is a slide. You have seen it. A large language model sits in the middle of a diagram with arrows coming off it, and one of the arrows says "negotiation". I have seen it in more pitch decks this year than I can count and, in the interests of full disclosure, I have presented a version of it. It is not wrong, exactly. What irks me is that it is imprecise in a way that costs money, and I want to try to be precise about why.

Here is the picture I keep coming back to.

Say you are buying a flat in Lisbon. You do not speak Portuguese, the seller does not speak English, so you hire an interpreter, and the interpreter is excellent. Every word you say lands on the other side of the table exactly as you meant it. Every word that comes back arrives with the tone intact. To anyone listening in, it sounds like two people who know what they are doing.

Now ask what the interpreter does not know. They do not know your ceiling. They do not know which clause you would walk out over. They do not know what the flat is worth, and nobody has asked them to. If you have not told them any of that, and you have not brought someone to write down what was agreed, the interpreter will carry you into a dreadful deal in flawless Portuguese and go home feeling they have done a good day's work. Which, strictly, they have.

That is an LLM in a media negotiation. Really, it is just the interpreter.

The job the interpreter does do

I want to be fair to the interpreter, because the job is not a small one. Most of the friction in buying media has always been two humans failing to say precisely what they mean to each other. The brief that means one thing in the planning deck and something else by the time it reaches the trading desk. The counter-offer that comes back as a two-line email with the important bit left out. An LLM takes a brief and turns it into an offer a counterparty can parse, reads the counter and tells you what it actually says, and holds a conversation that sounds exactly like two people doing a deal. That removes a great deal of friction the industry has spent twenty years working around.

But removing friction and making a good deal are different jobs. A fluent negotiation and a sound one are not the same event. Sounding like you know what you are doing and knowing what you are doing are separate skills, and I have watched them come apart inside the same meeting.

What a negotiation needs that language cannot give it

A real negotiation needs things that outlast the conversation. A price. A floor. A record both sides can point at later. And someone who has to live with the result if it turns out to be wrong. None of those is a language problem. They are mandate, verification and consequence, and a model has no native concept of any of them. It will produce a perfectly articulate contract with a viewability floor that protects nobody, because grammatically there is nothing wrong with it.

I built Concourse, with a great deal of help from Chandru, who did the parts that work, to find out how big that gap actually is rather than keep asserting it on panels. Eight frontier models. One simulated market with three software-controlled sellers. Four versions of the same brief, three registered attempts at each, 96 attempts in total. Fifty-five ended in a completed buy, which on its own is not bad: more than half the time, the interpreter got to the end of the conversation and shook hands.

What is staggering is what they shook hands on. Across the 110 contracts signed in those 55 buys, 95 promised a viewability floor below 70%. The IAB set a 70% campaign threshold back in December 2014, calling it a reasonable goal for a year of transition. The agents put their names to something below a baseline that is more than a decade old, and every one of those deals closed cleanly. On a dashboard each one is a green tick. The dashboard was not the one whose budget it was.

The agents were not bad at communicating. They were fluent. They were just not negotiating in any sense that would survive a client asking whether it was a good deal.

So here is the question I would put to anyone with that slide in their deck. If the model is the interpreter, who in your stack is the negotiator?

Agreeing the dialect is not the same as knowing the price

The industry's answer so far has mostly been to standardise the language. AdCP and AAMP are both, at bottom, agreements about how one agent talks to another: what an offer looks like, how a counter is structured, what a signed deal contains. Useful. Necessary, even. Two interpreters who have agreed which dialect to use will get through the conversation faster. But a shared grammar does not come with a price in it. It does not tell either side what a bad deal looks like once the conversation ends. Nobody designed it to, and that is not a criticism. It is just not where the problem lives.

The problem lives in the part that happens before anyone speaks: the mandate the negotiator walks in with, and whether they have ever been in a room where they could lose. That is what the Alkimi Gym is for. It does not ask which protocol an agent speaks. It puts the agent in a market where it can fail, with a brief, floors it is supposed to hold, counterparties leaning on them and a record of what it signed, and it keeps doing that until the agent has a working sense of what a bad deal looks like before it signs one. Which floors to hold. When to walk. What "fine, grammatically" is hiding.

The model underneath does not change. What it walks in knowing does.

The bit I do not have yet

I would rather say this than have you work it out. Concourse is our benchmark and we ran it ourselves. Nobody independent has reproduced it, and until someone does I treat it as a strong signal rather than a settled fact. There is also one number I cannot give you, and it is the one that would make all the others mean something: how a trained human doing the same job would have contracted. We have not built that baseline yet. It is planned, and it is honestly the next thing this needs before a completion rate can be called good or bad rather than simply recorded. If the gap turns out smaller than I think, you will read that here too.

What to ask instead of which model

"Which model is it built on" is not the question. That is asking the estate agent which school your interpreter went to. A good one, probably. Lovely accent. Still cannot tell you what the flat is worth.

The better questions are the ones you would ask of a person before handing over a cheque book. Has this agent been tested somewhere it could fail? Has it been shown what a bad deal looks like before it touched a live budget? What is it allowed to agree without asking, and who signed off on that boundary? It is the earned-autonomy logic I have written about before: observe, recommend, draft, get it approved, then act on its own inside limits it has proven it understands. Nobody hands over the keys on day one.

Hire the best interpreter you can find. The models are extraordinary and they are improving faster than most of us can keep up with. Then make sure someone in the room knows the price, knows the floor, and knows when to stand up and leave. Put the two together and the transcript stops merely sounding competent. It becomes one you would put your name to.

If that slide is in your deck, I would like to know what sits on the other end of the arrow. Tell me. I am easy to find.

Sources

Concourse Bench v1, Alkimi: concourse.agency (8 models, 96 registered attempts, 55 completed buys, 95 of 110 contracts below a 70% viewability floor)
IAB, State of Viewability Transaction 2015, December 2014: iab.com
Ad Context Protocol (AdCP): ppc.land/ad-context-protocol
Agentic Advertising Management Protocols (AAMP), IAB Tech Lab: ppc.land/aamp