I came across a piece in Digiday last week that reinforced something I’ve been pondering since I started writing on this blog.
Digiday’s executive news editor Seb Joseph and senior marketing reporter Sam Bradley wrote it, and at its core it is about AI marketing costs in the advertising industry, specifically the growing gap between how much agencies are spending and how well anyone can actually measure what it is delivering. The reason why this piqued my interest is that I have been watching this from a few different angles over the past year, running an agency, talking to people about marketing strategy, and teaching marketing at a polytechnic. The pattern the piece describes is one I recognise.
One detail that I found particularly grounding was the story of Laura Higgins, incoming chief brand and innovation officer at Dollar Shave Club, the US men’s grooming subscription brand. In her first month in the role, she ran everything through AI tools without paying much attention to what it cost. Claude, ChatGPT, Gemini, video tools, all of it, for campaign concepting, for routine questions, for just about everything. Within weeks, she had built herself a personal triage system with lighter models for routine tasks and heavier ones reserved for work that actually needed them. The shift from unmetered exploration to something more sustainable took about a month.
That is a fast cycle for one person. The industry is trying to make the same adjustment at scale, and the timelines are considerably longer.
Running without a meter
One of the more telling details in the piece is about PMG, the performance marketing agency. They rolled out a fifty-dollar-a-day token cap per user across the company. From the outside, that might read as a cost-control measure. The way the team described it is closer to governance. It simply means knowing they had capacity for when demand actually spikes, whether that is a major campaign launch or a busy retail period with multiple agents running at the same time.
The cap is not there to restrict. It is there so nobody finds out at the worst possible moment that the tokens ran out.
That distinction carries more weight than it first appears. Knowing what you spent is a different thing from knowing what it was worth, and the piece makes clear that the industry has got considerably better at the first while still figuring out the second. There is an analogy in the article about open-plan offices, which spread rapidly through corporate real estate on the strength of savings that were easy to calculate, without much evidence they improved the quality of work for the people inside them. Token caps carry a similar risk. They surface a number but do not tell you what that number bought.
Three pricing approaches and an unresolved problem
Agencies are currently taking three quite different approaches to AI marketing costs, and none of them are the same.
Dept, the digital agency network, will not pass token costs to clients at all, on their view that once you start billing clients for tokens, the focus shifts to the cost of the machine rather than the value of the work. Monks, the content and technology company under S4 Capital, has folded tokens into its subscription and technology fees. The large holding companies such as Publicis are largely absorbing AI marketing costs into broader commercial arrangements, which makes them harder to trace and also harder to audit.
All three positions are defensible, but none of them solve the underlying problem.
The harder problem sitting underneath all three approaches is that nobody has yet agreed on what AI work should deliver, or how you would measure it if it did. Without that, cost tracking becomes a reporting exercise rather than a management tool.
One of the things I teach in marketing is ROAI, or Return on Advertising Investment. The core question it asks is deceptively simple, which is how do we know we’re spending the right amount, and what are we actually getting in return?
What makes it tricky, and this is something I find my students often underestimate, is that measuring advertising effectiveness splits into two quite different problems. Communication effects, things like reach, recall, and engagement, are relatively tractable. You can run a study, pull a metric, show a chart. Sales and profit effects are a different matter. Sales and profit are influenced by so many other variables at the same time, product quality, pricing, competitor activity, seasonal shifts, that isolating what the advertising spend specifically contributed becomes very difficult.
The AI industry is now sitting in front of exactly the same measurement problem, dressed differently. Token spend is trackable. Agencies can tell you how many tokens their teams consumed on a given day, down to the user level, and what it cost. What they cannot easily show is what that spend contributed to the actual outcome, whether a campaign was better, a client was better served, or a margin improved because of AI rather than a dozen other things shifting at the same time.
Until there is an agreed framework for the uncomfortable question of what the AI spend actually delivered, three agencies can price AI marketing costs three different ways and none of them can prove which approach is right.
The margin numbers say it all
On a recent Publicis H1 2026 earnings call, the CFO pointed to a margin improvement of seventeen basis points in the first half of the year as evidence that AI marketing costs are being offset by genuine productivity gains. He also acknowledged that the improvement followed more than thirty basis points of efficiency gains that had already been reinvested into AI tools and staff training.
That is not a bad result. It may well be exactly what responsible technology adoption looks like during a transition period, phased, with gains reinvested rather than immediately extracted. Clients were told early on that AI would mean leaner fees and faster delivery, with no corresponding cost increase on their end. The numbers do not quite match that promise yet.
The gap between AI marketing costs and provable value is where most of the difficult conversations are happening right now.
The prompt count that is viewed differently now
Last year, it was reported that Coca-Cola’s Christmas advertisement required a team of five AI specialists to sort through more than 70,000 individual clips over a thirty-day span. At the time, that figure circulated as a sign of creative muscle, AI ambition and bragging rights, something worth mentioning.
I wonder how that number would land in a budget conversation today, when procurement teams have compute costs as a line item and CFOs are asking what the output was actually worth. The context around that kind of data point has shifted considerably in the span of twelve months.
When adoption outruns accountability
An industry analyst quoted in the piece described the current moment as the end of the honeymoon between AI and agencies. CMOs were told that AI marketing costs would be offset by efficiency gains and faster delivery. As contracts come up for renewal, many are still waiting for the line in the report that shows where those savings landed.
I do not think this means AI is failing to deliver. In a lot of the work I have seen, it is delivering. The more specific problem is that the frameworks for measuring it clearly have not kept pace with the speed of adoption. Every technology wave I have been part of in marketing has had some version of this. The implementation moves fast, the measurement catches up later, and the reckoning in between is uncomfortable but useful.
There is a principle from my AI governance programme that fits this moment well. The course makes the case that the most overlooked stage in any AI deployment is post-deployment monitoring. Most organisations put considerable effort into getting AI tools into people’s hands. Once the system is live, attention tends to shift to the next thing. The governance view is that deployment is actually when real oversight needs to begin, because that is when you find out whether the tool is doing what you intended, at the cost you expected, and producing the outcomes you were promised. It is a lesson Ford, Klarna, and Commonwealth Bank each learned the hard way.
The ad industry is now learning this at scale. The token caps, the pricing debates, the procurement conversations are all forms of post-deployment governance arriving late. The framework is being built after the tools have already been running for twelve months. Adoption moved faster than accountability could follow. That is fixable, provided the industry is willing to ask the harder question of what they actually got, rather than defending what they spent.
The AI marketing costs bill for the last twelve months has arrived. Working out what it actually paid for is the task ahead.
