Why 55% of Companies Regret AI-Driven Layoffs (And What They Got Wrong)

May 22, 2026 — Wendy Kinney

Fifty-five percent of companies report regretting AI-driven layoffs. According to Forrester research published in April 2026, more than half of organizations that reduced headcount based on AI capability projections say they moved too quickly, cut the wrong roles, or both.

They weren’t bad leaders. They were working from bad data.

In early 2024, Klarna announced it had reduced its customer service workforce by roughly 700 people, citing the capabilities of its AI assistant. The announcement was cited widely as evidence of AI’s workforce displacement potential. Within months, Klarna was quietly rehiring customer service staff – a reversal covered by The New York Times, Forbes, and Bloomberg. The AI had handled the high-volume, transactional interactions. The judgment-intensive, complex cases that remained required exactly the institutional knowledge that had just been shown the door.

Klarna’s story isn’t an outlier. It’s a preview of what happens when workforce decisions are made from AI capability projections rather than from ground-truth data about what work actually requires human judgment.

Key Takeaways

55% of companies regret AI-driven layoffs – the primary cause is decisions made from benchmarks and projections rather than individual-level activity data

The Klarna reversal illustrates the core problem: AI handled the routine transactions, but institutional knowledge handling complex cases was cut along with the headcount

The three root causes of AI layoff regret are benchmark-based targets, invisible work left unaddressed, and institutional knowledge treated as headcount rather than workflow

Operations leaders who establish a ground-truth activity baseline before acting can identify which cuts reduce work versus which cuts just reduce capacity

If you haven’t acted yet, you have a window. The 90-day workforce intelligence assessment is designed specifically for leaders who need a defensible baseline before their next board review

What the 55% Regret Statistic Actually Means

The Forrester finding is specific: “55% of companies that reduced headcount based on AI transition plans report outcomes they now characterize as mistakes – either the cuts were deeper than the AI actually required, the wrong roles were eliminated, or service quality degraded in ways that weren’t anticipated.”

Operationally, “regret” shows up in three ways:

Service Level Agreement degradation. The work didn’t disappear. The people who handled it did. SLA performance drops because the remaining team is managing the same total work volume with fewer people.

Emergency rehiring. The most visible form of regret. The company publicly reduces headcount, then quietly opens positions for roles that turn out to be essential. The hiring cost, knowledge reconstruction cost, and reputational cost typically exceed what was saved.

Institutional knowledge loss. This is the hardest to measure and the most damaging. When experienced employees leave, they take with them the informal processes, exception handling patterns, and relationship knowledge that held operations together. That knowledge doesn’t show up in any job description or process documentation – which is also why it doesn’t show up in the consulting analysis that justified the cuts.

The companies in the 55% aren’t failed companies or reckless leaders. Many of them followed the best available guidance from reputable consulting firms. The problem wasn’t the leaders. The problem was the data those leaders were working from.

The Three Root Causes of AI Layoff Regret

Understanding why these decisions go wrong requires looking at how the targets get set in the first place.

Cause 1: Targets built from benchmarks, not from actual activity data

Most AI workforce reduction targets start with an industry benchmark. A consulting firm identifies that similar organizations achieve X% efficiency through AI. The board adopts the benchmark as a target. The operations leader gets handed the number.

The benchmark describes average performance at average companies. It doesn’t describe what work your specific organization does, at what volume, with what system constraints, serving what SLA requirements. A benchmark can tell you what the number should be. It cannot tell you which specific activities in your specific operation are the right ones to eliminate.

When the benchmark drives the decision, cuts happen to the visible, countable roles – not to the specific activities that actually drive cost. The result: headcount goes down, cost-per-transaction holds steady or deteriorates, and service quality suffers.

Cause 2: Visible work gets automated; invisible work gets ignored

Every operation has two categories of work: the work that’s visible in process documentation and org charts, and the work that’s invisible – absorbed into individual work habits, handled through informal channels, never named as a discrete task.

The visible work is what consulting analysis identifies. The invisible work is what gets left behind when headcount is reduced. And invisible work is frequently the high-volume, high-cost activity that was actually driving unit costs in the first place.

Here’s what invisible work looks like in practice: Claims adjusters spending 90 minutes per day manually cross-referencing a claims database with a policy system that should be integrated but isn’t. Loan processors who’ve built personal workarounds in Excel because the loan origination system doesn’t communicate with the compliance database. Operations coordinators who act as informal bridges between field teams and back-office systems because no one ever built the integration.

None of this shows up in a job description. None of it would appear in an interview-based time study. All of it disappears – along with the people doing it – when headcount is reduced.

Cause 3: Institutional knowledge treated as headcount

The third cause is the most consequential. When experienced operations employees leave – voluntarily or through reduction – they take with them knowledge that isn’t documented anywhere. Exception handling patterns developed over years. Relationships that enable fast escalation when systems fail. Understanding of why certain informal processes exist that would look inefficient to an outside observer.

This knowledge is operationally critical and structurally invisible. It’s not captured in any reporting system. It doesn’t show up in productivity dashboards. And it’s precisely what AI cannot replace – because AI can only replace work that’s been defined, classified, and scoped. Undefined, undocumented knowledge can’t be automated; it can only be lost.

David, a VP of Operations at a mid-sized utilities company, described the experience this way: “We modeled the reduction carefully. We thought we understood the roles. What we didn’t understand was what those people actually spent their time on – specifically, the coordination work they did between the field dispatch system and the back-office billing systems. That coordination was 40% of their day. Nobody had ever written that down. When they left, the failure rate on billing reconciliations tripled within 60 days.”

What “Bad Data” Actually Means in This Context

When operations leaders say they acted on bad data, they don’t mean the data was fraudulent. They mean the data was structurally incapable of answering the question it was being used to answer.

The question is: “Which work in our operation can AI replace?”

The data consultants typically use to answer this question includes:
– Interviews with team leaders and managers
– Time studies based on employee self-reporting
– Application usage data from IT systems
– Industry benchmarks for role categories

This data answers the question “what do our employees say they do?” It does not answer the question “what do our employees actually do, at the activity level, including the informal and invisible work?”

The gap between those two questions is where AI layoff regret lives.

Individual-level, activity-based data captured directly from the work environment – without relying on self-reporting or application-level summaries – answers a different question: “What specific tasks, in which workflows, at what frequency, with what rule consistency, are present in this operation right now?”

That data supports a completely different kind of cut. Not “we’re removing 15% of claims adjusters because the benchmark says we can” – but “here are the 8 specific activities that account for 34% of adjuster time, each of which scores high for automation potential, and here is the implementation sequence that removes those activities without reducing the human judgment capacity the operation requires.”

One of these approaches produces the Klarna story. The other produces a defensible roadmap.

What Would Have Prevented the Regret

The companies that got AI workforce decisions right – and there are examples – shared a common approach. They didn’t start with a headcount target. They started with an activity baseline.

An activity baseline is a complete, individual-level, daily view of what work is actually happening in the operation: which tasks, how frequently, in which applications, with what inputs and outputs. Captured automatically, not reported by employees or estimated by consultants.

From an activity baseline, three things become possible that aren’t possible without it:

Identifying the right work to eliminate. Not the visible roles – the specific activities within roles that are high-volume, low-judgment, and clearly automatable. The baseline shows exactly where those activities exist and what volume they represent.

Preserving what can’t be replaced. The baseline also surfaces the informal, invisible work – including the institutional knowledge that tends to disappear in headcount reductions. It makes that work visible, documentable, and preservable before the reduction happens.

Building a defensible roadmap. When every cut is tied to a specific set of activities that AI will replace, the workforce decision is defensible to the board, to remaining employees, and to regulators. “We reduced headcount by X% and unit costs dropped by Y% because we eliminated these specific activities” is a fundamentally different narrative than “we reduced headcount by X% and we’re monitoring the impact.”

See how the Summit Trails approach establishes this baseline – specifically the Capture to Classify to Insight pipeline that produces consulting-grade output automatically, for every employee, every day.

If your AI mandate has a deadline, the time to build the baseline is now – before the cuts, not after.

Book a 30-minute strategy call to discuss what a workforce intelligence assessment would look like for your specific operation and timeline.

What to Do If You Haven’t Cut Yet

If you’ve received a mandate but haven’t acted, you have a window. The 90-day workforce intelligence assessment is designed specifically for this moment: enough time to establish a complete activity baseline, identify the right automation targets, and build a defensible roadmap before your next board review.

The sequence is:

Deploy the Ground Truth AI2 Platform on the target workforce (50+ employees in the relevant operations area)
Capture 30 days of individual-level activity data across all workflows
Receive quick-win findings at weeks 3-4 – the first ground-up view of the operation before the full report
At day 90, receive the full Ground Truth AI2 Report: activity breakdown, automation scoring, prioritized roadmap, and staffing model analysis
Take that roadmap into your board review

What you take into the board review isn’t a consulting estimate. It’s your own data. The activities your operation actually performs. The volume. The automation potential. The specific implementations that will reduce unit costs and the ones that would damage SLA performance.

That’s the difference between a defensible plan and a benchmark.

What to Do If You’ve Already Made Cuts

If your organization has already reduced headcount and is managing the consequences, the data is still useful – and available.

An activity baseline established post-reduction tells you:
– Which activities were lost when people left (by identifying gaps in the current operation’s output)
– Where the remaining team is carrying disproportionate load
– Which of the remaining manual workflows are highest priority for automation to prevent further SLA degradation

The regret statistic is about decisions made without data. It’s not a verdict on companies that are working to course-correct. The data that would have prevented the problem can also diagnose it and point toward solutions.

See what the platform output includes – and whether it addresses the specific gaps you’re managing now.

The Broader Context: Why This Is Happening Now

The 55% regret figure is not a stable number. As AI capability projections continue to outpace actual organizational readiness to implement, and as board-level pressure to show AI-driven cost savings intensifies, more organizations will face this pattern.

80% of enterprises will deploy AI agents by 2028, according to Gartner. 30% of work hours are projected to be automatable by 2030, according to the McKinsey Global Institute. 92% of large businesses say they’re already prioritizing AI. And yet only 29% of CEOs say they’re confident in their current AI strategy (PwC 2025 CEO Survey).

The gap between “we have AI ambitions” and “we know what work AI can actually replace in our specific operation” is the data gap. It’s not a technology gap. The technology exists. It’s a ground-truth data gap – and it’s the specific data gap that Summit Trails was built to close.

The companies that navigate the next 36 months without becoming part of the regret statistic will be the ones that built their workforce intelligence baseline before their board handed them a number. Not after.

For operations leaders preparing for that conversation, the May 14 webinar “Workforce Intelligence for the AI Transition” covers exactly this territory: what the data foundation looks like, why current approaches fail, and what a defensible AI implementation blueprint requires. Register here.

FAQ: AI-Driven Layoffs and What Goes Wrong

Why do 55% of companies regret AI-driven layoffs?
The primary cause is that targets were set from industry benchmarks and consulting estimates rather than from individual-level activity data showing what work AI could actually replace in that specific operation. When cuts happen to visible headcount rather than to specific automatable activities, unit costs hold steady and service quality often degrades.

What happened with Klarna’s AI layoffs?
In early 2024, Klarna reduced its customer service workforce by approximately 700 employees citing AI assistant capabilities. Within months, the company began rehiring for customer service roles. The AI effectively handled high-volume transactional interactions but the institutional knowledge needed for complex, judgment-intensive cases had been reduced along with the headcount.

What data would have prevented AI layoff regret?
Individual-level activity data showing which specific tasks – not which roles – are high-volume, low-judgment, and automatable. This data needs to be captured at the activity level (not application level) and classified by type, frequency, and automation potential before workforce decisions are made.

Is it too late if we’ve already made cuts?
Not entirely. A ground-truth activity baseline established post-reduction can identify which activities were lost, where remaining teams are carrying unsustainable load, and which remaining manual workflows are highest priority for automation to stabilize operations.

How can operations leaders defend workforce decisions to their board?
By grounding every decision in specific activity data: “we removed X% of headcount because these specific activities – representing Y% of unit cost – are being automated by these specific implementations.” That narrative is defensible. Benchmark-based projections are not.