Before You Cut Headcount for AI, Think About This

Somewhere in 2025, a manager was asked to let someone go. The reason given was efficiency. The real reason was a promise that AI would handle it from here.

This played out across hundreds of companies that year. The narrative was straightforward: AI was coming, it would absorb the work people were doing, and getting ahead of that shift meant cutting now rather than later.

It felt like a smart, forward looking move at the time. It did not stay that way for long.

The AI that was promised and the AI that actually showed up to work turned out to be two very different things. Here’s some data to back that up.

A 2026 survey of 600 HR professionals who had conducted AI-led layoffs found that only 8.4% said their plans delivered the results they promised. 

Forrester Research found that 55% of employers regretted their decision to lay off workers for AI-related reasons, and more than a third of companies had already begun rehiring for over half the roles they had eliminated, most within six months of letting those people go.

This is because when those people left, they took years of context with them: how clients liked to be handled, what workarounds existed, which processes only made sense if you'd been there long enough to understand why. AI couldn't replicate any of that.

The employees who stayed noticed, pulled back, and started doing the bare minimum. And customers, especially in service-heavy businesses, started feeling the difference before most companies were willing to admit there was a problem.

This is not an argument against AI, because AI is genuinely changing how work gets done. It's a look at what happens when companies treat it as a headcount calculator rather than a capability multiplier, and what the smarter path actually looks like.

Cost of letting people go: A case study

Klarna is the cleanest example of how this played out. The Swedish buy now, pay later company, used by roughly 150 million customers worldwide, became one of the most talked about case studies in enterprise AI when it decided to bet big on automating customer service. 

Between 2022 and 2024, the company eliminated around 700 positions, mostly in customer service, and replaced them with an AI assistant built with OpenAI. At its peak, Klarna said its AI was handling two thirds to three quarters of all customer interactions, and the company framed it as the efficiency story of the AI era.

For a while, the numbers looked good as the AI handled 2.3 million conversations in its first month alone. But the metrics that made headlines were the easy ones to win. Resolution rate and tickets handled per hour looked great, while quality on the harder, more nuanced interactions was quietly falling apart.

By mid-2025, customer satisfaction had dropped enough, and enough operational problems had surfaced, that Klarna's CEO Sebastian Siemiatkowski admitted the company had focused too much on efficiency and cost, and that the result was lower quality. 

So, Klarna began rebuilding its human support team, and by 2026 had shifted to a hybrid model where AI handles routine, high volume queries and humans take the escalations and the complicated, high value interactions.

The pattern at Klarna shows up across the industry too, just less publicly. Companies kept betting on a version of AI that could fully take over, and what actually showed up was a tool that was very good at the easy 80 percent of the work and not ready for the hard 20 percent that actually needed a person.

The unseen loss

When people leave, they take years of context with them. How clients like to be handled, what workarounds exist, which processes only make sense if you have been there long enough to understand why. 

AI models might possess vast amounts of data, but they lack understanding of a specific company's culture, its unwritten norms, and its history with clients. That kind of knowledge never makes it into a prompt.

The employees who stay behind notice, start to pull back slowly. 

Research firm Forrester has documented a growing group of employees termed as “coasters”, people who no longer feel their employer deserves their full effort, expected to make up 28 percent of the workforce in 2026. 

It tracks with what they watch happen around them: colleagues laid off for an AI that never fully materialised, entry level roles disappearing, and being asked to do more with less while being told to be grateful. 

And customers feel it before most companies admit there’s a problem. Frustrated customers eventually push ecommerce and fintech companies to rehire the content writers, engineers and service workers they had replaced with AI in the first place.

This effect only shows up months later.

Why it made sense on paper but not in practice

The logic behind most of these layoffs was never really about what AI was already doing.

Harvard Business Review surveyed 1,006 global executives and found that 60 percent had reduced headcount in anticipation of what AI might do in the future, while just 2 percent said the cuts were tied to AI that was actually deployed and working. 

That gap between potential and performance is the whole story. 

Cutting a role saves money right away, and that shows up on the books immediately. Whether the AI was actually ready to take over that work shows up much later, often after the people who could have caught the problem are already gone.

What the companies getting it right actually did

While all of this was playing out, a different group of companies were quietly proving the opposite point. 

PwC's 2025 Global AI Jobs Barometer looked at almost a billion job postings across six continents. It found that industries most able to use AI saw three times higher revenue growth per employee than industries least able to use it. 

PwC is careful to say this isn't proven cause and effect, but the timing lines up. Revenue growth in these industries picked up sharply in 2022, the same year ChatGPT made the world pay attention to AI.

The wage data tells a similar story. Wages in AI heavy industries are rising twice as fast as in industries with low AI use, and pay is going up even in the most automatable roles. Workers with AI skills now earn a 56 percent wage premium, up from 25 percent in 2024.

The companies seeing this kind of payoff are the ones using AI to help people create more value.

This rarely makes the same headlines as a layoff announcement, because there is no single dramatic moment to point to. It is a slower, less flashy strategy.

Three ways to use AI without losing your best people

The approaches of companies getting this right tend to fall into three categories. The first is redirecting people to higher value work instead of replacing them. When AI takes over the repetitive parts of a job, the person who used to do that work does not become unnecessary, they become available for the parts of the job that actually needed judgment in the first place. A support agent who no longer has to handle routine password resets can spend that time on the complicated, high stakes cases that Klarna eventually realized still needed a human.

The second is freezing hiring instead of cutting existing staff. If AI is genuinely going to reduce how many people a function needs, the lower risk way to find that out is to stop backfilling roles as people leave naturally, rather than cutting people who are already there and already carry context about the business. This gives a company time to actually see what AI can do before making a decision that is hard to undo.

The third is using AI to increase what the existing team can produce, instead of using it to shrink the team. This is the PwC finding in practice. The companies seeing the biggest revenue gains per employee are not running leaner teams, they are running the same teams at a higher output, because the people on those teams have been given better tools and the training to use them.

If there is one thing worth taking from all of this, it is to start asking what your best people could do if the repetitive, low value parts of their job disappeared. That question is hard to answer but it’s the question that the data is now backing.