The ad industry still hasn't priced AI. One holding company found a place to put the cost where the client can't see it.
The pitch is easy to like. A holding company offers to eat the entire AI infrastructure bill — every license, every inference call, all of it. The client pays nothing for the technology. All they have to do is route 70 percent of their media budget through inventory the holdco owns.
Digiday reported the offer last week. It surfaced mid-renewal, when a CMO deep into a media account got terms nobody had put in front of them before. Service costs, discounts, and pricing were all folded into the same conversation, so it wasn't a clean trade. But principal media was the biggest chip on the table, and the AI giveaway was the sweetener.
One deal isn't a market. Digiday found two more executives describing versions of the same structure, including a proposal that opened at a zero fee for running all spend through principal inventory before both sides settled somewhere in the middle. The pattern is early. It's still a pattern.
The word doing the work in that pitch is free. Free is exactly what this isn't.
How the Margin Hides the Cost
Principal media is the mechanism, and it predates AI by years. In a normal arrangement, the agency buys media for the client and takes an agreed fee. In a principal deal, the agency — or a company inside the same holding group — buys the inventory itself, at wholesale and in bulk, then resells it to the client with data, targeting, and guaranteed placement attached. The gap between what they paid and what they charge is the margin. That margin is the point.
It's also where the AI cost now lives. The client never sees a token bill. They commit a fixed share of spend to principal inventory instead, and the markup on that inventory funds the technology. The bigger the allocation, the bigger and steadier the pool, and the more AI cost it can quietly absorb.
And the cost is real. PMG now caps each employee at $50 a day in tokens. A year ago, Coca-Cola's Christmas ad reportedly ran on 70,000 prompts, and that was a number to brag about. Nobody's bragging now.
None of the plumbing is new. Agencies were pouring money into principal media long before anyone worried about the price of inference, for the simple reason that it's the most profitable thing they do. Daniel Knapp, chief economist at IAB Europe, told Digiday that agencies have become good at running as futures markets — pricing risk on the media side, carrying it on their own credit lines. Absorbing an unpredictable AI cost, in his view, sits within the natural DNA of the agency.
So, there's a respectable version of this deal. The client gets predictable AI costs and discounted inventory. The agency finances a technology program before the market has decided what a token is worth. Both sides skip a tedious fight over every model call. If that were the whole story, this would be smart financial engineering and not worth an article. It isn't the whole story.
You're Paying Twice
Start with the people. For two years, clients have been told AI would lower their fees — less manual work, faster turnaround, leaner teams. A lot of that has happened. What hasn't happened is the fee coming down. Agencies still staff accounts with full human teams while AI absorbs a growing share of what those teams used to do, and the saving from that gap has not reached the client. It's been kept.
You don't have to take my word for the direction of travel. On Publicis's most recent earnings call, the CFO described a rebalancing between people cost and technology cost — people down, technology up — and named it as a source of margin improvement, 17 basis points of it in the first half. He was answering a question about a 7 percent rise in other operating costs driven partly by AI, and his case for why that's manageable was that the productivity gains offset it. Read plainly, that means the efficiency AI created landed on the agency's side of the ledger. Not the client's.
Now add the second charge.
The AI that produced the efficiency — licenses, inference, the engineering to make any of it usable — gets recovered separately, through the principal-media margin.
The client pays once by never receiving the fee cut the technology should have earned them, and again through a markup they can't see. "Free AI" is the second charge wearing a bow.
Robert Webster, a former WPP executive who now runs an AI consultancy, put the uncharitable reading on the record: agencies claim to have invested heavily, he told Digiday, but much of that is manufactured to justify skimming money out of media. I wouldn't state it that flatly. I can't see inside a staffing decision, and neither can he — overstaffing that pads a bill and overstaffing that reflects real caution look identical from the outside. But he has the incentive named correctly, and the incentive is the problem whether or not anyone acts on it cynically.
This is where the good-faith defense runs out. Knapp's timing argument — nobody knows what tokens should cost yet, the market needs a few years, agencies are built to carry exactly this kind of risk — explains why the cost needs a home right now. It does not explain why that home has to be a margin the client is structurally barred from inspecting. You can grant every word of the charitable case and still walk away holding a bill you aren't allowed to read. The timing problem is real. The opacity is a choice.
The Adviser Is Also the Counterparty
Principal media strains the relationship before AI touches it. An agency giving advice is supposed to recommend the media that serves the campaign. An agency acting as principal makes money when the client buys inventory it owns.
Those two jobs can live inside one company — but only if the client can see where one ends and the other starts.
The AI subsidy pulls that knot tighter. Once a holding company has agreed to cover infrastructure in exchange for a spending commitment, it has one more reason to point budget at its own inventory. A given recommendation might be completely sound. The client simply has less ability to tell whether performance or margin drove it.
The Association of National Advertisers put numbers to the unease in March. Ninety percent of the marketers it surveyed named uncertainty about whether principal-media recommendations serve their interests as their top concern, up from 79 percent two years earlier. Only 57 percent said their company had any guidelines governing the practice. That was measured before AI costs were folded in. The trust gap was already open; this deal widens it.
There's an irony worth sitting with. In the same survey, 76 percent of marketers named lower costs as the main reason they use principal media.
The mechanism they reach for to save money is the one they least trust to be straight with them — and it's now being asked to carry a second cost on top of the first.
Give Compute Its Own Line
The fix isn't crude token billing. Metering inference tells you what was spent, not what it bought. Fewer tokens might mean a cheaper model or thinner work; more might mean waste or a genuinely hard campaign. Caroline Giegerich of the IAB named the trap: without a measure of what the work delivered, you can't separate an agency driving results from the least efficient agency known to man.
But the answer to "tokens are a bad price" is not "so bury them in media."
The market is already testing other answers. Dept won't pass token costs to clients at all, arguing that itemizing them cheapens the work. S4's Monks does the reverse and builds tokens straight into its pricing. The holdcos are taking a third route — fold the cost into a bigger structure and stop itemizing anything.
The route worth arguing for sits between the first two. The client gets a separate AI schedule: what the agency provides, how the estimate was built, model access split out from implementation work, usage-based inference stated in bands or caps where an exact number is impossible. It doesn't need to be precise. It needs to be visible. When a model gets cheaper or the agency swaps in a lighter provider, the client should watch the economics move — not watch the saving disappear into a margin.
Two Tests the Seller Can't Grade
Whatever the pricing, the package makes two promises, and each needs its own proof.
Did the AI make the work better? Did the principal inventory beat what the client could have bought somewhere else?
A lower fee answers neither. Neither does a cheaper cost per thousand. AI can spin out more variants and faster reports without moving a single business number.
Principal inventory can carry an attractive unit price and still deliver an audience the client would never have picked in an open comparison.
Set the baselines before the deal starts. For the AI, that might be cycle time or production cost, and wherever you can manage it, some read on the campaign actually improving. For the media, test incremental results and quality against a credible alternative. One rule holds both together: the company selling you both cannot be the only one grading them.
The Renewal Is the Control Point
None of this is a conversation to have after the strategy is signed. It is the strategy. The terms decide whether the agency has a reason to steer your budget, and whether you can defend the allocation later to finance, to internal audit, or to a board.
Three things have to survive the negotiation. The contract has to state, transaction by transaction, when the agency acts as your agent and when it acts as principal — approval can't hide inside a blanket authorization. Audit rights have to reach the entity that actually buys and resells the inventory, not stop politely at the agency of record while the principal deal runs through another company in the group. And the AI commitment needs a reconciliation: if a fixed slice of media pays for a service whose cost keeps moving, there has to be a periodic reset and a way to shrink the allocation when the underlying cost drops.
The real enemy is inertia. A two-year media commitment struck at today's inference prices won't loosen on its own when the agency quietly moves to a cheaper model six months in.
Without a review mechanism, the client keeps paying through the margin and loses the evidence to challenge it. That exposure gets set at the renewal table, long before it ever shows up in a campaign report — which is the whole reason the renewal, not the report, is where a CMO has to win this.
Free AI is only free if you never learn what it costs. This deal is built so you don't.