Faster Horses, Again.

Two percent.

That is the share of major companies whose AI-driven headcount reductions were justified by actual AI performance, according to a Harvard Business Review study published earlier this year, titled (with some precision) "Companies Are Laying Off Workers Because of AI's Potential, Not Its Performance".

Sixty percent had already cut, or were planning to cut, in anticipation of AI's future impact.

Sixty cutting for potential. Two cutting for performance.

That gap is not an AI story. It is the most consequential strategy error of the past two years. And the marketing industry is beginning to reckon with it, quietly, expensively, and without the press releases that accompanied the original decisions.

Henry Ford observed that if he'd asked people what they wanted, they'd have said faster horses. The observation was never about horses. It was about the limits people unconsciously place on their own imagination when a familiar technology is working well enough. The horse is useful. The horse is understood. The horse is measurable. So when pressure arrives, the instinct is to optimise the horse rather than question whether the horse is still the right vehicle.

Here is what that actually looks like in practice: AI's earliest and easiest applications were never going to be the hardest ones. Content at volume. Research synthesis at speed. Presentations drafted in minutes. Creative variations multiplied at a fraction of previous cost. Faster. Cheaper. More of the same, delivered quicker. These were the obvious moves and organisations made them, announced them, and reported them as strategic progress (and sometimes, and here I'm looking at you WPP, they dress them up and call them 'agentic' hoping no-one looks behind the curtain).

But faster and cheaper are table stakes, not transformation. And even the cost argument is softer than it appears. The economics of AI that make it look cheap today are not market economics. They are subsidy economics. The major AI infrastructure providers like the compute platforms, the model developers and the API layers, are running at a scale funded by capital deployment at a pace that has never been sustained in technology before. The cost of intelligence is being artificially suppressed while the race for adoption is won. When that capital cycle matures, and the true cost of AI infrastructure normalises, the "cheaper" argument will look considerably different on the CFO's spreadsheet. Organisations that built their AI strategy around cost reduction rather than capability expansion will find the maths has quietly shifted beneath them.

The marketing industry has spent two years building faster horses and calling it transformation.

The horses are real. The productivity gains are genuine. I am not arguing against them. But there is a difference between a faster horse and an automobile. The early data suggests a significant number of organisations raced so hard in the wrong direction that they are now quietly turning back.

Fifty-five percent of employers who made AI-driven layoffs regret the decision, according to Forrester Research's 2026 Future of Work report. That figure is not a marginal finding. The Washington Times reported this week that companies including IBM, Salesforce, Google and Meta have been quietly rehiring for roles they eliminated such as content writers, strategists, customer service specialists, people whose judgment turned out to be harder to replace than the efficiency case suggested.

The Harvard study adds the structural explanation. Organisations automated the conclusion (fewer people) without having done the work to understand which human contribution was actually dispensable and which was not.

Gartner projects that half of the companies that cut for AI-related reasons will rehire for similar roles by next year. The "Boomerang", as analysts are now calling it, was entirely predictable. It is what happens when the faster horse gets mistaken for the destination.

The CEO of the industry's best-performing company has a precise explanation for why this happened. Arthur Sadoun gave a wide-ranging interview to Campaign Asia that deserves more attention than it received.

On AI and talent, he was direct in a way that no other major holdco leader has been. His competitors, he said, communicate to the investment community that they consider people a debt. He called this a dream. Specifically, the dream that tomorrow's marketing services business will function as a synthetic company. His position: Publicis will never be that. Not because it is opposed to AI, but because it understands what AI actually changes.

But the argument being read into his comments, that headcount should simply be preserved, misses what he actually said. Buried in the interview is a more precise observation about where the AI adoption story currently stands. Most companies, he said, are realising that leveraging AI is not that easy. It is difficult to scale. It is expensive. And so far it has not produced business outcomes at scale. He attributed this to a fundamental disconnect: consumer adoption of AI is accelerating, while company-level adoption is struggling to convert technology investment into commercial results.

This is not a technophobe speaking. Sadoun built Marcel, integrated Epsilon as CoreID, and positioned Publicis as the data-first holdco a decade before the rest of the industry caught up. He is not against AI. He is against a specific and widespread misreading of what AI means for a services business with the conclusion that because AI can automate execution, the human layer responsible for judgment and insight is the efficiency opportunity.

That misreading is what the Harvard data is now quantifying. Sixty percent acting on potential. Two percent acting on evidence. The gap between those numbers is where the faster horse lives.

Across thirteen years at Publicis and now building an AI company watched how organisations respond to new capabilities. The pattern is consistent. Every major technology wave (digital, programmatic, social…) generated the same initial response: use it to do what we already do, faster and more efficiently. The quick wins come first. They are measurable, defensible, and real. And they create the illusion of transformation while the actual transformation waits.

AI is following the same pattern, but at a pace that compresses the timeline dramatically.

The applications that moved first were the ones with clear, existing templates. If you were writing three briefs a week, now you write fifteen. If you were testing two creative variants, now you test twenty. If you were synthesising a research report in three days, you do it in three hours. These are genuine productivity gains. They are not transformation.

Transformation is when you do something that wasn't possible before. Not the same work faster. Different work entirely.

Here's where the industry is making its most consequential error. Not in adopting AI for efficiency… that is entirely rational. But in treating efficiency as the destination rather than the entrance. In cutting the human layer that was producing the marginal output without asking which part of that layer was producing the irreplaceable output. In building proprietary AI platforms, racing to announce headcount reductions, and telling investors the hard work is done, when the hard work, the genuinely difficult work of understanding what AI enables that was previously impossible, has barely begun.

We are only at the beginning of the beginning.

Gartner published research this month predicting that fifty percent of advertising agencies' proprietary AI platforms will either wind down or become obsolete by 2029. WPP Open. Omnicom's Omni. Publicis Marcel. Havas Converged.AI. Half of these, on Gartner's baseline projection, will not exist in their current form within four years.

The reason is structural, not circumstantial. Agency AI platforms are built to serve marketing and advertising functions. Hyperscaler platforms (Google, Amazon, Meta, Microsoft) are being built to serve enterprise functions across every department simultaneously. Gartner projects that open-source platforms will support more than seventy-five percent of enterprise AI deployments by 2028, cheaper and more customisable than anything agencies have built.

The CIO, not the CMO, will own enterprise AI strategy at the company level. The CMO who has embedded their team's AI capabilities inside an agency-controlled silo may find themselves excluded from that conversation when their CIO standardises on Microsoft Azure AI or an open-source alternative. The proprietary agency platform becomes not a competitive advantage but a compatibility problem.

This is not a warning about lock-in, though it is also that. It is a structural argument about where durable value lives. And Jay Wilson, VP analyst at Gartner, was explicit about it: What an agency provides that is not tech-dependent, that cannot be as easily commoditised, is the outside-in perspective on a client's business.

The outside-in perspective. The pattern recognition across categories and contexts. The creative judgment that operates above the execution layer. The human insight that doesn't arrive from the same statistical model that your three nearest competitors also queried this morning.

This is not the horse. This is the automobile. And it is where AI is liberating us to spend more time, if we choose to.

Gartner's companion research from February carries a number that should concentrate minds in a way the platform obsolescence story does not.

Only fifteen percent of CEOs believe their marketing leaders are currently AI-savvy in 2026. Gartner predicts that by 2027 a lack of AI literacy will rank among the top three reasons CMOs are replaced at large enterprises.

Sixty-five percent of CMOs believe AI will dramatically change their role in the next two years. Only thirty-two percent believe significant personal skills changes are needed to meet it.

The CMO who went downstream with everyone else, who used AI to make existing work faster and cheaper, who delegated AI strategy to the team or the agency or the IT department, is now facing a board that has a fifteen percent confidence level in their AI capability. The window is not three years. It is the next eighteen months.

AI literacy for a CMO is not knowing how to use the tools. It is knowing what the tools cannot do, where human judgment remains irreplaceable, and how to build an organisation that asks the automobile question rather than the faster horse one. That is a leadership question, not a technology question.

Henry Ford's point was never that horses are bad. Horses are useful, reliable, and understood. His point was that the people asking for faster ones had already defined the limits of their own ambition without realising it.

The faster horses arrived. They are real. They are delivering genuine value. And they are not the vehicle this moment is asking us to build.

The automobile question, or asking yourself what becomes possible that wasn't before, what work can we now do that we couldn't, what insight can only a human reach… that is where the value is going. It is harder than optimising the horse. It requires fighting the inertia of doing things differently rather than just faster. It requires tolerance for a measurement gap, because the genuinely new work doesn't fit neatly onto last year's investor slide.

Sadoun described his competitors as living in a dream. The dream is the synthetic company that is faster, cheaper, more automated as the destination. The industry is beginning to wake up.

The question is whether you're already building something that moves differently. Or whether you're still asking for a faster horse.

Brandflow is written by Justin Billingsley, who has spent his career on all three sides of the industry's table: senior client, global agency leader, technology founder. First published 16 March 2026 in the Brandflow newsletter on LinkedIn.