AI Layoffs Create Budget Room But Not Returns

Gartner published a finding in May 2026 that I have been thinking about since I read it. The claim is simple and uncomfortable: autonomous business and AI layoffs may create budget room, but they do not deliver returns. That sentence, from an organization that advises the largest enterprises on the planet, should be stopping board meetings. As far as I can tell, it is mostly not.

The pattern Gartner is describing is one I recognize from the IS literature on technology-driven labor substitution. An organization looks at an AI system, estimates the labor it can replace, cuts the headcount, banks the savings, and then waits for the ROI to arrive. The AI system, now responsible for work that used to involve human judgment, context, and informal knowledge, starts producing results that are inconsistent, incomplete, or wrong in ways that are hard to detect until a downstream process fails. Then comes the realization: the system needs more human oversight than expected, not less. And the humans who would have provided that oversight are gone.

McKinsey's 2025 State of AI report found that 88 percent of organizations are using AI in some form, but only 7 percent have fully scaled it (McKinsey, 2025, https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai). The gap between "using AI" and "fully scaled AI" is exactly the gap where the labor substitution fallacy lives. Organizations that are in the first category, which is nearly all of them, have AI systems running on favorable data, in controlled use cases, with human teams close by who can catch errors and fill gaps. "Using AI" looks like it works. Then the headcount reduction happens, and the conditions that made it look like it was working disappear with the people.

This is not a new dynamic in the IS field. The labor substitution debate has been around since at least the early days of enterprise automation research. The consistent finding is that substitution is harder than it looks because jobs are bundles of tasks, and the tasks that are hardest to automate are often the ones that make the automatable tasks possible. A customer service agent who handles routine queries also knows when a routine query is actually not routine. They know the context. They know the customer's history in a way that goes beyond the data in the CRM. Automate the routine queries, and you are fine, until the edge cases pile up with no one to catch them.

I think there is a deeper governance failure underneath the budget argument. When organizations cut headcount based on AI productivity projections, they are treating a forecast as a fact. The forecast is that the AI will handle the work. The fact is that the AI can handle some of the work, under conditions that resemble the conditions in which it was deployed, as long as the data it depends on remains clean and relevant and the processes around it remain stable. None of those conditions are guaranteed. All of them erode over time without active management. And the people who would have managed them are the ones who were let go.

Gartner's predictions for 2026 also note that AI investments are growing at more than 35 percent year over year, and that CIOs need to shift from calendar-based planning to trigger-based planning (Gartner, 2025, https://www.gartner.com/en/newsroom/press-releases/2025-10-21-gartner-unveils-top-predictions-for-it-organizations-and-users-in-2026-and-beyond). The trigger-based framing matters here. If the trigger for maintaining or expanding AI deployment is evidence of actual returns, rather than the passage of time or the existence of a pilot, then organizations would know much earlier when the returns are not materializing. Instead, most organizations are still treating AI deployment as a one-time decision rather than an ongoing management problem that requires continuous measurement.

What does "returns" from AI actually require? My read is that it requires three things that headcount reduction cannot provide and in fact actively undermines. First, process redesign. The work the AI takes over needs to be redesigned around AI capabilities and limitations, not just handed to the system as-is. AI does not slot into existing processes cleanly. The process has to change, and changing processes requires the people who know how the current process works. Second, data quality. AI systems are only as good as the data they run on, and data quality is a continuous operational problem, not a one-time fix. Someone has to own it. Third, change management. The people who work alongside AI systems need to understand what the system can and cannot do, develop judgment about when to trust its outputs, and build new routines for acting on what it produces. That is a human capability that takes time to develop and requires the right people to be present.

The labor substitution framing treats AI and human labor as direct substitutes, where one can simply replace the other at the ratio the productivity estimates suggest. The augmentation framing treats them as complements, where the combination produces something neither can produce alone. The IS research literature has been working through this distinction for decades, and the augmentation hypothesis has the better empirical record. When organizations use IT to augment human work rather than replace it, they tend to see sustained improvements. When they use it to replace human work wholesale, they tend to see the returns plateau or reverse as the hidden coordination costs of the automated system accumulate.

I wrote about why AI pilots rarely become products in an earlier post, and the mechanism is related. The conditions under which AI looks productive are often not the conditions under which the organization actually operates at scale. The pilot uses clean data, has motivated humans watching closely, and runs in a forgiving environment. Production uses whatever data exists, has reduced human oversight by design, and runs in an environment where errors compound. Cutting the humans who provided the oversight before the AI has proven it can work without them is the exact sequence that produces the result Gartner is describing: budget room, but no returns.

The IS research question that I keep coming back to is whether this is a governance design problem or an adoption sequencing problem. My instinct is both. Governance design, because the decision to reduce headcount is being made without adequate measurement frameworks for whether the AI is actually handling the transferred work. Adoption sequencing, because the right order is probably to demonstrate that the AI works at full scale first and then reduce the headcount that is genuinely redundant, not to cut first and hope the system performs. The organizations that will actually see returns from AI are the ones that resist the temptation to cash in the labor savings before the returns are real.