AI's easiest business case is its weakest

Every wave of innovation arrives with the same slide. A chart of the current org, a smaller chart beside it, and an arrow labelled with a number of heads and a number of pounds. The same is now playing out with agentic operations. AI agents that watch the estate, diagnose the fault, and act are the newest technology to be pitched this way. The pitch is seductive: the saving looks certain, immediate, and easy to book. My last piece argued that these agents are an estate to be governed, not a toolbox to be filled. It ended on a warning I now want to make the whole argument: sold as a headcount play, agentic operations does not survive the first budget cycle.

Put headcount reduction at the core of the business case and the ROI you booked will not bank. Put operational performance at the core and the headcount falls out anyway — from a different layer, and more slowly.

This is not a soft plea to be kind to people. It is a claim about which numbers are durable.

The number that flatters year one

A headcount cut is a one-time step down in the cost base. It is real in the first year, then it laps. By the second year the like-for-like comparison shows nothing further, because the new baseline already absorbed it. To keep reporting a headcount-driven improvement you have to keep cutting. That is precisely how a programme starts eating muscle. Headcount is a level. It moves once.

Performance is a slope. Shorter cycle times, faster response, quicker pricing moves and product launches show up on the revenue and margin-quality side. They recur every quarter, and they can compound. You can cut a cost once. You can raise a capability every quarter.

Managers reach for the level anyway. The reasons are well documented, and rational at the individual level. Research on earnings-per-share targets finds that EPS is among the most pervasive metrics in executive pay, and that managers hit those targets with real operating levers — cutting R&D, capital expenditure and employment — even when it costs the firm later. And the stated reason conditions how the market responds: layoffs presented as efficiency measures are spared the penalty that falls on cuts blamed on weak demand. A permanent annual saving, capitalised at the firm's earnings multiple, is notionally worth many times itself in enterprise value. So the cut is chosen because it is controllable, measurable, and immediately rewarded — not necessarily because it is valuable. The headcount fixation is a measurement artefact, not an economic truth. People optimise what is easy to count, and heads are easy to count.

The trouble is that the durable evidence points the other way. In Responsible Restructuring, Wayne Cascio reported an eighteen-year study of S&P 500 firms. Companies that downsized were, as a group, no more profitable than peers that did not. There was no lasting advantage in profitability or shareholder return, and frequently a lasting cost in the capability, quality and morale of what remained. Layoffs are the lever that is easiest to pull and hardest to bank.

Where the AI earnings actually are

You might reasonably ask whether AI has changed this — whether, this time, the headcount saving is showing up in the accounts. The market data says the opposite of the pitch.

The earnings AI is producing are real. But they are concentrated in the companies building AI, not the ones adopting it to cut costs. Analysts expect S&P 500 earnings to grow around 32% in 2026, on a net profit margin running near 15% — among the highest FactSet has recorded since it began tracking the figure in 2009. Roughly half of the S&P 500's 2026 earnings growth is expected to come from AI-infrastructure beneficiaries, and the Magnificent Seven are set to grow earnings about 22.8% in 2026 against roughly 12.1% for the other 493. The ten largest US stocks are now about a third of the Russell 1000 — a level of concentration not seen in half a century. This is a picks-and-shovels story in the goldrush era: the people selling the AI are booking the earnings.

Among the adopters — the firms buying AI to run leaner — the return is largely missing. MIT's 2025 report, The GenAI Divide: State of AI in Business, found that 95% of enterprise AI pilots delivered no measurable P&L impact, despite an estimated $30–40 billion of investment. The few that worked were narrow back-office automations, not sweeping transformations, and budgets skewed to sales and marketing even though the better returns sat in operations. In the UK, Accenture's 2026 research makes the same point: the share of working hours that AI could potentially enhance jumped from 47% to 82% in two years, yet only 3% of UK organisations are fully ready for advanced, agentic AI and just 23% of employees have seen a major process redesigned around it — the gains confined to the margins rather than remaking the enterprise. The FTSE 100 tells the same story from the other side: it rose more than a fifth in 2025 and pushed to new highs, but on the back of banks, miners, oil and defence — technology is only about 3% of the index, against roughly a third of the S&P 500 — so its gains owe little to AI, and nothing to a cull.

Some firms do benefit at the coalface, to be fair to the counter-case: the Lloyds Business Barometer found that, among UK firms using AI, 87% reported higher productivity and 48% higher profits over the previous year. But those are self-reported gains from firms that chose to adopt — not the automatic headcount saving a slide promises.

We have run this experiment before

None of this is peculiar to AI. In 1987 Robert Solow observed that "you can see the computer age everywhere but in the productivity statistics." The IT payoff did arrive, but only in the 1990s, roughly a decade later. And it arrived only where the technology was paired with organisational change, rather than treated as a way to buy the same output with fewer people. The benefit lagged the spend. It accrued to the firms that redesigned the work around the new capability.

The most recent cycle is closer still. Robotic process automation was sold, a few years ago, on almost exactly today's headcount promise. Industry surveys bore this out: EY reported that 30–50% of initial RPA projects failed outright, and Deloitte's global survey found only about 3% of organisations ever scaled RPA past 50 bots. The programmes that stalled were typically undone by fragility and the cost of maintaining brittle bots. The organisations that treated a fleet of those scripts as free headcount soon found they had bought a maintenance liability instead. Agentic operations is more capable, but the failure mode is the same shape, and larger. That is the reason to govern the estate rather than simply count the savings.

The move: automate up, not down

If the headcount-driven level is the weak number, and the performance-driven slope is the durable one, then the restructuring inverts.

Start with an uncomfortable observation about the people you would cut first. When managers plan headcount reduction, the axe usually falls on the layer below them — the operators closest to the work. But in an organisation that has just automated itself, those operators are the people who built and deployed the agents, and who therefore hold the model of how the agents fail. Lisanne Bainbridge named the trap in 1983, in a paper now foundational to human-factors engineering. The more successful the automation, the greater the investment in human skill it requires, because the human is left to handle exactly the rare, hard cases the machine cannot. If that human has been deskilled or dismissed, no one can take it back when it matters. Cut your operators and you remove your recovery capacity in the same move you automate.

So do not cut them. Move them up. The operators who understand both the coalface and the agents governing it are the natural stewards of the estate. They can hold the accountability and manage the risk of an automated operation, because they are the only ones who know what "wrong" looks like. Their job changes from doing the work to owning the agents that do it.

The redundancy that automation actually exposes sits above them. A large part of middle management has always been an information-processing layer — collecting, aggregating, and relaying operational information up and down. That is not a slur; it is the documented history of delayering. Earlier IT waves already thinned that layer, letting senior managers oversee more people with fewer intermediaries, precisely because software took over the collecting and presenting of operational data. AI is the next turn of that same compression. What it removes is the relaying, not the doing. So the headcount does come down — but from the middle. The administrative content of that layer collapses into instrumentation and a thinner spine of genuine judgment, part of it now staffed by your elevated operators. You end with a real pyramid, thinner at every human level, but with value flowing up from the people who built the automation, rather than down from the people who allocated the heads.

Be precise about what survives the middle, or the argument becomes a caricature. The relay content is automatable. The judgment, the coaching, the cross-team negotiation and the accountability are not, and they consolidate onto fewer, better-paid people. The useful test on any middle role is simple: does this person produce a decision, or a relay? Relays automate; decisions stay.

Two tests before you touch the org chart

The reframe is only honest if it can tell a genuine case from a fashionable one. Two questions do most of the work.

The first is about the organisation. In a world before AI, was your headcount already a strategic issue — or has it appeared only in the context of AI? If the firm knew it was overstaffed relative to peers, and there is a dated rationale for it — a board paper, an analyst note, a benchmark discussion that predates the AI initiative — then AI is a legitimate lever for a decision that already existed, and a headcount objective can be defensible. If the first time "we are overstaffed" appears is inside the AI business case, the technology is being used to manufacture a rationale. That is the red flag. Even in the legitimate case, the cut is the enabling condition, not the objective. And it still cannot be done in haste: the estate has to be proven and the operator knowledge preserved first.

The second is about the adviser. Consultancies reach for the headcount story because it is the bookable, attributable number that gets them rehired — the same measurement artefact, one level up. But a consultancy is itself a people-hours business, which makes a firm that bills by the consultant-hour selling you headcount reduction a little like a locksmith recommending a house with no doors. The fair challenge is an honest, respectful alignment test. If AI genuinely delivers the headcount savings you are selling, show it first in your own delivery economics. Pass it to me as lower rates, shorter implementations, or both — and tie your fee to my durable performance metrics, not to how many of my people are let go. A firm that believes its own thesis will take that structure. A firm capturing the saving, or selling a thesis it does not hold, will flinch.

Slower than you can

The slower tempo matters, and not out of timidity. You can only remove a human loop after you have proven the agent loop and built the governance around it. You need your operators around to prove it, because they hold the knowledge. Move too fast and you have deleted the people who could take it back before the agent is trustworthy. Set the target as headcount reduction and Goodhart's law does the rest. The organisation optimises the number instead of the automation, and you get a fragile operation and a saving that quietly reverses. Fast looks better on the slide precisely because it books the ROI before it meets the reality that will claw it back.

The conclusion, put plainly

Automate well and you end up smaller. Set out to get smaller and you automate badly. The headcount business case is the seductive one and the weak one — a level that laps, a saving that is easiest to announce and hardest to keep, chosen because it is measurable rather than because it is durable. The performance case is harder to measure and worth far more: a slope that compounds, and the quality of earnings that a board or an investor actually pays a premium for.

So when the next slide arrives with its arrow and its number, the question to put to it is not "where is the saving?" It is: and what does this do to my operational performance? If the answer is "nothing, we just have fewer people," you were either overstaffed already or you are about to degrade the buffer and discover it in the fourth quarter. If the answer is that performance improves, you have a durable case — and the headcount falls out at the end, where it belongs, from the layer that was mostly relaying information around.

You can cut a cost once. You can raise a capability every quarter. Build for the second one, and the first takes care of itself.