Who's Capturing the AI Efficiency Dividend: Clients, or Agencies?
The conversation your AOR's execs don't want to have along the Croisette next week.
Next week, the world’s largest communications agencies (and the holding companies that own most of them) will descend on the sunny shores of France for the Cannes Lions International Festival of Creativity. In splashy corridors and aboard rented yachts, they will hail their own brilliance. And for all the sturm und drang about the decline of the agency, it remains a good enough business to pay the champagne bill.
The story they’ll tell about why it stays that good (why the retainer is still worth it) is, this year more than any other, artificial intelligence. Every firm now has an AI story: a proprietary platform, a named model, a head of AI on a panel promising the work is faster, sharper, and more effective than it’s ever been. And there’s truth there. The agencies have built stuff over the last year, even if there’s plenty of hype heaped on top of it.
But if the AI is real and the efficiency is real, who pockets it: the clients funding the retainer, or the agencies billing it? Those are very different answers, and only one of them is the story you’ll hear next week.
Let’s start with product
Edelman built Archie, a proprietary model trained on 20 years of its Trust Barometer data, and wrapped it last fall in a new operating layer it calls ECOS, the “Edelman Central Operating System.”1 Burson shipped the Decipher suite, cognitive-AI models that claim to predict how a message will land, nested inside WPP’s holding-company AI platform. Weber Shandwick signed a multimillion-dollar, three-year deal with Google for HALO, an agentic platform with a roster of named data partners. FGS Global runs an agentic tool called Fergus across 1,500 consultants. Golin has publicly committed to becoming “the first fully AI-integrated global PR agency” and hired a chief AI officer to get it there. Ruder Finn says custom AI already touches 88% of its core U.S. account work.
Agencies have spent two years building, buying, and partnering their way to an AI story, then telling it at full volume. But listen for the one claim conspicuously missing from all that noise: that any of these proprietary platforms actually beats the off-the-shelf models and data you could license yourself. That’s the line you’d put on a billboard if it were true, and it’s the one line none of them will say. What they sell you instead is a frontier model you could rent direct, wired to the same third-party data feeds you could subscribe to yourself, with their name on it.
But it doesn’t have to be better than those things. I run my own practice on these models, and I can tell you plainly: they’re already good enough to seriously lift productivity, so good that not even a big global agency can screw it up. The efficiency is real. The proprietary layer you’re paying a premium for mostly isn’t where it comes from.
If the gains come from a model your own team could rent for a fraction of the retainer, then the question isn’t whether the AI works. It’s: real for whom?
The scorecard is self-graded
Edelman’s model claims “trust recommendations” with “up to 97% accuracy,” and the two words doing the work in that sentence are “up to.” That’s the phrasing you reach for when you can’t quote an average.2 FleishmanHillard says more than 1,000 of its strategists are on its AI platform and points to “a Fortune 100 client” across “six markets.”
These are the agencies’ own scorecards: unaudited, self-reported, and where there’s a client at all, conveniently anonymous.
Then there’s what’s missing from the marketing. These firms are machines built to publicize a win: the moment a client sees a real result, the case study writes itself, the logo goes up on the site, the happy CMO gets booked for the panel. That’s the reflex for anything that works. So where is it? Where’s the parade of named clients crediting an agency’s AI for a gain it delivered? For an industry that never met a win it wouldn’t trumpet, the empty space where those testimonials should be is the loudest thing in the room.
I did this exact exercise once before, for the GEO racket: a fully built service category, five-figure monthly retainers, and not a single public case study showing the underlying thing works. Same result here, gone industry-wide.
This is not evidence that the AI doesn’t work. It’s evidence that the agencies can’t, or won’t, show you that the working part reached you. Which raises the obvious question. If the efficiency is real, and it isn’t showing up on your side of the ledger, where did it go?
It went upstream, by design
This is the part nobody on a Cannes rooftop will volunteer, and it’s the part you should care about most. The AI efficiency is real. It’s just accruing to them.
Start with cost. Forrester and the 4As found that 75% of U.S. agencies are absorbing their generative-AI costs rather than passing them to clients, a share up sharply from a year earlier. Forrester named it the “AI Cost Center Crisis,” which tells you whose crisis the industry thinks it is.
Now follow the productivity. When AI lets an agency do the same scope in half the hours, where does that time go? Not into a lower invoice. The retainer is a fixed monthly number set to capture effort, and it doesn’t move when the effort drops. The margin improves; the retainer doesn’t. Muck Rack’s State of AI in PR found AI use in the field is overwhelmingly internal: the gains stay inside the shop. The efficiency becomes margin accruing to the P&L, not to you.
The holding companies are explicit about this when they’re talking to investors instead of to you. Omnicom, fresh off swallowing IPG, has told the market it expects roughly $1 billion in savings from labor, headcount reduction banked as merger synergy, with AI named as what lets the cuts go deeper. That efficiency is being paid out to shareholders and merger math, not to the people whose retainers funded the capability in the first place.
And the people who broker these deals see the split plainly. Lucinda Peniston-Baines, who advises brands on what they pay their agencies, says agencies treat AI as “an internal efficiency tool … to regain some of that margin erosion they’ve suffered over so many years,” even as their clients expect those savings now. Same technology, two ledgers: one side booking margin, the other still waiting for a discount.
Building > renting
Here’s the tell. When clients have asked to share in the savings, the answer has been no. Roughly a third of agencies have already fielded an explicit request for an “AI discount.” Meanwhile one analysis of the survey data pegs the premium agencies actually command for AI work, where they command one at all, at 20% to 30%: not the doubling the conference stage implies, but a surcharge all the same. Read the sequence: you ask for the discount and you’re declined; the surcharge is what’s on the table next.
So when an agency asks you to pay more to fund its AI, be clear about the trade: you’d be subsidizing the tool that fattens their margin and lengthens your invoice, in exchange for an outcome nobody will put a number on.
You don’t have to take this from me. Take it from the analysts paid to advise the buyers, on the record where the agencies’ happy clients are not. Gartner’s Jay Wilson has watched this cycle and called it: the honeymoon around AI and agencies “is essentially over.” His read is the whole argument in miniature. Clients asked whether the AI would make the work cost less. Agencies said not yet. And the savings still aren’t showing up in the fee. Not can’t. Aren’t.
Gartner is also telling CMOs not to lock into agency AI platforms at all, and forecasts that half of those proprietary platforms will be obsolete by 2029. The people paid to advise buyers are coaching them to treat the “proprietary” tool as a depreciating rental on a frontier model (which is what it is), not something worth a premium or a long contract.
Here’s what to ask
If you’re doing an agency search any time soon, here are three things I recommend you bring up in your process to cut to the chase.
Ask for the client, not the tool. “Show me a named client and the outcome the AI delivered.” A demo, a roadmap, or a Fortune 500 logo with no number attached is your answer. A real result has a client’s name on it.
Interrogate the price in both directions. If their AI makes them faster, ask where the efficiency went. And don’t accept “innovation investment” as the reason it went to them. And if they ask you to pay a premium, make them itemize exactly what you’re buying that you weren’t already getting last year.
Keep the exit. Take Gartner’s advice over your agency’s: any AI commitment gets a pilot and a clean way out. The firms confident in what they’ve built won’t flinch. The ones who can’t will suddenly discover all the reasons it won’t work.
But don’t ask these questions too aggressively at Cannes, because urine stains on white linen outfits are a pain to get out.
Full disclosure, since I’m naming names: I have history with two of the firms here. I worked at Edelman in 2020 and 2021, and I was a senior advisor to Weber Shandwick in 2025 and the first couple months of 2026.
How do you prove a trust recommendation accurate? It’s like Sex Panther cologne in Anchorman: “They’ve done studies, you know. 60% of the time, it works every time.”