‘We’re starting to wonder’: Ad industry chases AI value as usage outpaces proof

“I have no idea what that bill looks like,” she said of that early stretch, chalking it up to an experimental phase. It didn’t last. Within weeks, Higgins had built herself a rough triage system: cheaper tools like Gemini for simple tasks, heavier models reserved for work that actually needed them. The shift from unmetered exploration to token discipline — or figuring out, in her words, how not to “burn up my tokens” — took about a month. 

It’s a small, personal version of a calculation now playing out across the ad industry.

The deeper AI gets embedded — moving past pilots and into the daily mechanics of a business — the more computing power it demands to keep running, and the more that shows up on the bill. Higgins hit that wall in a single month and changed course. The rest of the industry is working out how to do the same thing at scale.

Metering the machine

A month ago, PMG rolled out Alli For You to the full company. It pools staff access to every major model — ChatGPT, Claude and the rest — under a $50-a-day token cap per user. The cap followed months of alpha and beta testing that began in January, when testers ran on what one PMG exec called “free token range.” That pilot period was how the agency worked out what a sustainable daily allowance should look like. Before it existed, staff pinged LLMs directly, unmonitored, and the costs added up.

These days PMG says it rarely even hits the $50 ceiling. But that was never really the point. The cap exists for when it does.

“Because we’re starting to wonder, if it’s Black Friday and you’re launching ads, you’re doing reporting, you’re leveraging an agent, we want to make sure we have enough tokens to handle all that. Do we need to add more tokens?” said Kaitlin McGrew, head of SEM at PMG. “Thankfully if we need to, we can. But it’s really coming down to governance, making sure that we’re leveraging it in the right way.”

Reality check moments like this are par for the course with any meaningful tech shift. Take programmatic. It went through its own version of this. Years of uncritical spend before the industry sat down and actually audited where the money was going.

Tokenomics is a different kind of reckoning. Usage isn’t slowing down. But whether that usage is actually paying off is a lot less settled than the adoption numbers suggest. For every efficiency gain agencies can point to, there’s someone else questioning whether it holds up. That’s made harder by what clients expect AI to do in the first place. The assumption baked into most negotiations is that AI means lower headcount and lower cost, full stop, not better work worth paying more for. So agencies are stuck making a value argument to clients who’ve already decided the only acceptable outcome is a smaller bill – before anyone’s agreed on how to measure the work itself.

“The industry started at time saving,” said Caroline Giegerich, vp of AI and marketing innovation at the IAB. “Now it’s moving into what actual business impact do we have.” 

But for a business built for decades on billing by the hour, that’s a difficult sum to do and, as Giegerich put it, there’s still no way to know if an agency burning through tokens is driving results or is simply “the least efficient agency known to man”.

Because tracking token spend is easy. Knowing what it delivered is not. Call it the open-plan office problem: a cost-cutting move that spread through corporate real estate despite no real evidence it made anyone more effective, because the savings were easy to point to and the damage wasn’t. Token caps have the same blind spot. They tell CEOs how much was used, not whether it was worth using.

That blind spot is why agencies can’t agree on how to price this. Dept won’t pass token costs on to clients at all, arguing that itemizing them cheapens the value of the person using the tool, making the metric the tokens spent rather than the work produced. S4 Capital’s Monks has gone the other way, building tokens directly into its tech-and-subscription pricing. Others, the big holding companies especially, are trying to avoid the question altogether by folding AI costs into bigger commercial structures like principal media deals.

Three different answers to the same unresolved problem, and all three stand or fall on one thing. None of it matters without a way to define and measure what the AI actually delivered. Every fix on offer, insurance-style pricing, subscriptions and principal media bundling, is really a workaround for that missing measure. Without it, cost tracking can’t tell efficiency from waste.

“We’re moving all of our agencies toward output based pricing versus time and materials, all new biz is output based and 50% of existing clients have been transitioned over to output,” said Joe Maglio, CEO of Cheil Agency Network.

Maglio isn’t making that shift because he’s solved the measurement problem. He’s making it because standing still isn’t an option in a market where AI aptitude has become its own selling point. Most agencies are in the same position, forced to price outcomes they can’t fully account for, simply because the alternative is losing business to a competitor willing to try. Brands feel none of that urgency. Outside a handful with the scale to build AI operations in-house, most haven’t needed to get this disciplined, which leaves agencies carrying the risk of an unsolved problem largely alone.

It doesn’t help that procurement teams tend to see AI as a reason to pay less, not more. The expectation baked in from the start was that automation would drive costs down. An output-based fee tied to the revenue AI helped generate reads, to a lot of procurement teams, as the opposite of what they were promised, even when the number reflects real growth rather than padded hours.

“There is definitely a rebalancing between people cost and tech costs, and we’ve said it before, but we’re expecting it to stabilize now,” Publicis CFO Loris Nold said on the company’s earnings call last week. “What we are observing, if anything, is that the productivity benefits that those tools can generate do offset the cost, and it’s been clearly evidenced in our margin improvement.”

Nold was responding to a direct question about a 7% rise in other operating costs, partly driven by AI. He didn’t dispute the line item was growing. Rather, his case for why it’s manageable came down to three things: the spend itself is just licenses and usage, tracked daily down to the user level with caps and alerts: there’s the people-cost-to-tech-cost rebalancing, he flagged: plus, there’s a plan to cut the volume of certain manual tasks by an average of 25%, then scale that across the business — a target he said is “well underway,” not one already delivered.

The margin improvement he pointed to came to 17 basis points in the first half of the year, after more than 30 basis points of savings had already gone straight back into AI tools and staff training. In plain terms, the offset is spent before it ever reaches the bottom line.

Others, it’s fair to say, are still waiting for those gains, if the boozy, off-the-record chatter at Cannes is anything to go by.

“We had three conversations the last two weeks where they said it was cheaper to hire offshore engineers than to rely on a lot of code,” said one of those speakers. 

Put to a holding company executive, the same discomfort came out sharper. “Very few people are actually talking about the fact that the infrastructure has a real cost to it,” they said. “Tokens have a cost, compute has a cost — if you just go wild, the cost of the machines will quickly outpace the cost of humans.”

An industry analyst went further, arguing the goodwill that got AI this far in the door is already spent. “The honeymoon around AI and agencies is essentially over,” they said. “CMOs were sold on cost efficiencies and speed to market, and now, as they review and renew scopes of work, they’re still not seeing the cost savings in terms of fee.”

It’s a world away from what ad execs were saying this time last year when AI still looked like an unlimited resource. Coca Cola’s Christmas ad, for example, used 70,000 individual prompts. At the time, that was a stat to boast about. Try to imagine that figure going unquestioned now, when prompt volumes, compute costs and token caps sit squarely on the desks of CEOs and CFOs. 

The unlimited resource turned out to have a meeting running all along.