Layoffs Are the Wrong First Move on AI

A recent survey found that 69% of executives report their organization is making AI-related workforce reductions, while 39% admit their company doesn’t yet have a formal strategy for how AI will drive revenue. Read together, those numbers describe a decision most senior leaders are making right now: cutting permanent costs in exchange for a benefit that hasn't been defined.

AI-driven workforce reductions follow a reversal pattern: roles eliminated that get rehired inside six months, savings that don't hold, institutional knowledge that walks out and gets replaced at market rate. And the pattern isn't primarily a decision-making problem at the role level. It's a sequencing problem at the strategic level. Companies are making the workforce decision before they've built the operating model that would tell them what AI is actually producing.

Why the sequencing is wrong

The task-level business case for AI headcount reduction is often compelling. A model can now draft the first pass of a report, resolve a routine ticket, screen the initial round of applicants, or produce the standard analysis. If the work can be automated, the argument goes, the seat that used to do the work can go. On paper, the math works.

What's missing from this equation is the operating model. AI doesn't produce outcomes on its own. It produces them inside a system that has to decide how work now flows, who owns quality when the model gets it wrong, how errors surface and get corrected, what gets measured, and what happens when the AI produces something the business doesn't want. Those are the differences between AI that produces measurable value and AI that produces plausible-looking output nobody is accountable for.

Most companies making AI-driven workforce cuts have not yet designed that operating model. They have a tool, a pilot, and a business case built on projected performance. The reduction is happening against a benefit the system hasn't yet produced. When the system underperforms at scale, as it consistently does, the savings evaporate, and the rebuild starts.

Three questions to answer before the workforce decision

The senior leaders who navigate this well don't move more slowly. They sequence differently. Before the workforce decision, they can answer three questions specifically enough that the answers would survive a board conversation.

What outcome is AI meant to produce here, and how will you know? "Efficiency" and "productivity" name categories. They don't name outcomes. A defensible AI thesis names the specific business result that AI is expected to change: first-response time in customer service, cycle time for regulatory submissions, and quality of the first-pass draft in legal review. It also names the measurement that will confirm the change. If your executive team can't answer this in a sentence, the operating model hasn't been designed. Everything that follows is sizing against assumptions.

What operating model change enables that outcome, and is it in place? AI produces outcomes because a series of upstream and downstream decisions have been redesigned to let it. Deployment alone doesn't get you there. Who receives the AI's output? Who reviews it? Who owns the quality? Who catches the errors, and how quickly? What triggers escalation? If those questions haven't been answered with enough specificity that a manager could follow the answer, the AI is doing new work inside an old system, and any efficiency you're forecasting is speculative.

What triggers the workforce reduction, and what conditions reverse it? The reduction should be triggered by measured performance sustained over a defined period. Projected performance at pilot scale isn't the same signal. That means the operating model has to run long enough to know what it produces at scale, across a production workload, under the volume of edge cases the model will encounter in production. It also means naming, up front, the conditions that would reverse the decision. If the AI underperforms, if quality drops, if the remaining team can't absorb the exception handling, what changes?

What changes when you sequence correctly

The organizations we work with that get this right run their AI programs in the reverse order most competitors do. They design the operating model first. They deploy the AI inside that operating model and run it long enough to see what it produces at scale. They measure sustained performance against the outcome defined up front. Only then do they make the workforce decision, and when they do, it's grounded in what the system has already shown it can hold.

The organizations that skip the sequence end up in the reversal pattern: quick savings, followed by quiet rehiring, followed by the discovery that the rebuild cost more than the savings. That pattern comes from the absence of an operating model that could have told them whether the workforce decision was ready to be made.

Where the savings hold

The workforce decision is turning into a leadership signal. A company that announces AI-driven cuts without the operating model to defend them tells the market it has ambition ahead of discipline. A company that announces measured changes after AI has shown what it can produce at scale is proving it has both.

Over the next twelve months, that difference will sharpen. Boards will start asking which companies still have the savings they announced. The answer will trace back to what got built before the headcount moved.

P.S. On September 30, we’re walking through exactly how to build that sequence: a deliberate approach that moves a function from a broad AI mandate to a prioritized, defensible plan, starting with the operating model rather than the headcount decision. Join us: Leading a Life Sciences Function Under an AI Mandate: What It Actually Takes to Deliver.