AI Layoffs Reversed: Why Companies Are Rehiring

In July 2025, Commonwealth Bank of Australia announced 45 customer-service job cuts as it introduced an AI voice-bot. According to the staff union, call volumes were rising instead, and the bank was offering overtime and directing team leaders to answer calls. In August the bank reversed the cuts, apologized to the 45 employees, and said: "we should have been more thorough in our assessment of the roles required." (ABC News)

The takeaway: The AI bets that came undone were made from a task — a bot answering calls, an agent writing code — and applied to a whole job. The jobs held more than the task did. Where AI delivered, someone redesigned one process end to end and measured the outcome, not the usage. In the filings, cuts show up in costs; revenue changes have other causes. Prove the saving on one process before you change the org chart.

AI Replaced the Task. The Job Was Bigger.

Commonwealth Bank has company. In a 2026 Robert Half survey, 32% of U.S. hiring managers who eliminated positions after their company implemented AI later had to bring back the same or similar roles. Robert Half calls these "AI correction hires." The reasons managers gave for reinstating a role describe what the job held beyond the task (Robert Half):

• 40%: it required institutional knowledge or context AI couldn't replace
• 39%: it involved relationship management AI couldn't replicate
• 39%: business demand increased
• 38%: it required more human oversight and quality control than expected

That last one has a number of its own. A separate Robert Half survey found that employees using generative AI spend an average of 34% of their time on a task reviewing for accuracy and refining the final deliverable.

On paper, the math looks simple: if AI can automate a third of a team's work, you should be able to cut a third of the team. Reality doesn't work that way. Talented leaders, industrial engineers, designers, user experience experts, and human factors engineers all know that you can't simply swap out parts of an organization without looking at the bigger picture.

The remaining two-thirds of the work may not redistribute across the people who are left. The skills it needs may not be spread evenly across the team. The third that AI automates may not line up neatly with the human work around it. Or the human workers may have benefited from the fact that they carried a task from start to completion. An AI implementation of part of a job may not hand work cleanly into and out of a person's workflow. It's a risky assumption that making one part of a system faster won't slow down another part.

Whatever the reason, it's foolish to fire people before you have a working process that shows you don't need them.

Meta's Agents Wrote Code Faster Than Features Shipped

Meta made the largest bet. Under a 2026 plan called Project OT, reported by Reuters, Meta explored cutting some teams by as much as 60%, with smaller groups of staff supervising AI agents. Meta cut about 8,000 jobs, roughly 10% of staff, in May. The night before, Mark Zuckerberg called off planning for a second wave in November. (The Next Web, summarizing Reuters)

Meta's internal numbers show why output was the wrong scoreboard. According to a June post by chief technology officer Andrew Bosworth, code changes to internal platforms and infrastructure rose 220% year on year. Changes that reached users as new or upgraded features rose 36%. Major technical and security incidents climbed 40%, and time spent firefighting them rose 70%. Reuters could not establish what changed Zuckerberg's mind, and those figures don't prove AI caused every incident. At the least, it suggests that sweeping changes made in the name of AI can backfire.

At a later town hall, Zuckerberg said the "trajectory of the agentic development over at least the last four months hasn't really accelerated in the way that we expected." He still expected benefits within three to six months.

The Cut Took the People Who Knew the Work

Syndio's CEO, Maria Colacurcio, wrote in Fortune that in May 2026 Syndio eliminated a number of roles in a company redesign, and asked some remaining employees to take on more responsibility and a quota. Within a few months, several of the people she expected to keep resigned. "Suddenly, in our push to be more efficient in the age of AI, we lost some of the very talent we needed." In August, Syndio rehired a labor economist it had let go in May. Her conclusion: "AI can make experienced people even more valuable, if you give them the right tools to adapt."

Klarna didn't cut staff outright; it froze hiring after saying its AI chatbot could do the work of 700 representatives. In May 2025 it began a pilot to recruit customer-service staff again, so customers could always reach a human. Its spokesperson put it as "AI gives us speed. Talent gives us empathy." (CX Dive) Klarna has told the SEC that its CEO's comments on service quality were about its human customer service, not its AI, and were misreported. Klarna's customer-service and operations costs fell from $287 million in 2022 to $203 million in 2024 (prospectus), and they're rising again: up 18% from a year earlier in the second quarter of 2026, on an adjusted basis, though more slowly than revenue (Q2 2026 release).

Usage Is Not the Result

In April 2025 Duolingo said it would assess employees on their AI use. Nearly a year later, CEO Luis von Ahn let that metric go after employees asked whether the company just wanted AI for AI's sake: "the most important thing in your performance is that you are doing whatever your job is as well as possible. A lot of times AI can help you with that. But if it can't, I'm not going to force you to do that." In the same podcast interview, he described what happens when AI-written code fails: it's hard to debug because you don't know what it did, and "the unhappy path makes it so that it takes so long that you end up spending more effort on that than the time you saved."

Microsoft reached the same conclusion at a larger scale. Kathleen Hogan, Microsoft's chief strategy and transformation officer, wrote that "access and usage do not equal transformation: a tool licensed to over 200,000 people does not change how the work gets done." In sales, usage plateaued, and the impact did not materialize. (Microsoft)

A survey of nearly 6,000 senior executives in the U.S., UK, Germany and Australia, published as an NBER working paper, found nine in ten reporting no impact from AI on employment or productivity over the past three years. They expect it to come: on average, a 1.4% productivity gain over the next three years.

Where AI Delivered, Someone Redesigned the Work

After Microsoft's sales rollout stalled, the team stopped pushing adoption, started from the business goals, and mapped how account managers spent their week. Microsoft reports revenue per account manager 9.4% higher and close rates 20% higher in that group, but its footnote says the comparison is sellers who used Copilot daily against low-usage sellers in early 2024. Read that one as a usage comparison, not a before-and-after. The cleaner measurement is in supply chain, where experts and engineers mapped and simplified the workflow before adding agents: across five monthly planning cycles, average cycle time fell from about 10 business days to under 2.5. Hogan's rule for it: "Adding agents to a broken process still leaves a broken process — speeding up one step just creates a longer queue at the next."

The pattern at Microsoft: one process, mapped end to end, rebuilt with the people who run it, measured on a business outcome.

Block took the other route: it cut first and is measuring after. It cut more than 4,000 jobs, about 40% of staff, in February 2026, on the premise that "a significantly smaller team, using the tools we are building, can do more and do it better." CEO Jack Dorsey told staff, "I accept that we may have gotten some of them wrong," and Block has rehired at least four of the people it laid off, with no sign of broad rehiring (Business Insider). Its shareholder letter reports production code changes per engineer up over 2.5x since January — an activity number, like Meta's. It also reports a quality number that moved the opposite way from Meta's: incident rates after a production code change down over 70% from a year earlier. An internal agent makes 15% of production code changes nearly fully autonomously, "with people only intervening to make the final decision to push to production." Those are company-reported figures from about two months after the cut; they cover engineering, not every role the cut touched, and the reversals above took longer than two months to surface. The next few quarters will show whether Block's numbers hold.

Did the AI Layoffs Change Revenue?

Eight companies that cut staff, planned cuts or froze hiring on an AI premise publish their results. At about half of them, revenue kept the trend it already had. Where growth changed, something larger was happening in the same period.

The cuts show up in costs, but not as a smaller total. Three of the companies booked large one-time charges: $1.18 billion in severance at Meta in the quarter of its cut; $852 million in restructuring and other charges at Block, which also include legal accruals; and $2.5 billion in estimated severance at Amazon across the two quarters around its October cut. After that, the savings appear in particular lines. Block's product development costs fell 16% from a year earlier in the second quarter of 2026, which Block attributes to lower personnel costs, and its adjusted operating income rose from about 20% of gross profit to 25–27%. Amazon's general and administrative costs fell 6%, on lower payroll. Total costs still rose at most of the eight; Fiverr, whose revenue fell, is the clear exception. Part of the rise is AI itself: Meta's depreciation and amortization rose 46% and Amazon's 31% as both spent heavily on AI infrastructure, and Commonwealth Bank's technology costs rose 16%, with AI among the reasons the bank gives.

Fewer people didn't mean proportionally lower pay costs either. Klarna's headcount fell 49% from 2022 to 2025, while its total employee costs fell 10%. Its revenue per employee rose from about $344,000 to about $1.24 million, and Klarna credits its internal use of AI.

AI is moving some of these businesses: Meta and Amazon both credit it for growth. These results don't show the cuts moving revenue: where it changed, the filings name another cause. And where a cost line fell, that doesn't show AI is doing the work of the people who left.

Whether the cuts hold will show over the next year. Most of the roughly 8,000 people in Meta's cut leave its headcount by the end of September 2026, so its first full quarter after most of them leave is the last quarter of 2026. HP expects its full savings, about $1 billion a year, by the end of fiscal 2028. The table covers one to four quarters of company-reported figures, and none of them isolates the effect of a cut.

What to Prove Before You Change the Org Chart

The case is for proving the claim on one process before you spend headcount on it, and that is a case for moving. Before a staffing decision rests on AI, ask for these:

• The whole process, not the task. Use RACI: who is responsible, accountable, consulted, and informed? Map the exceptions, escalations, and handoffs, and the knowledge that lives in one person's head. Commonwealth Bank said it should have assessed the roles more thoroughly; the staff union reported call volumes rising after the bot arrived.
• A baseline, then a measured run. Compare end-to-end outcomes — cycle time, errors, customer results, revenue — against the old way. Lines of code, chats handled, and token counts are activity, not results; Meta's code changes rose 220% while shipped features rose 36%.
• Every hour of human work that remains. Count the review, the corrections, the firefighting, and the upkeep. If a third of the time goes to checking output, the plan has to say who does that.
• A change you can reverse. Prefer attrition, redeployment and a parallel run to a mass cut. Rebuilding that knowledge usually costs more than keeping it.
• End users involved in the design. At Microsoft, supply-chain experts worked side by side with engineers, mapping and simplifying the workflow before adding agents. Any new software, AI-enabled or not, needs to mesh with the humans who use it. Introducing a process improvement is a lot like launching a new external product; success depends on training, support and whether people choose to use it.

For a board, the questions are short. Which business outcomes changed because of AI, and over what period? Do the isolated productivity gains of an AI process translate into a productivity gain for the whole organization? Or do they just move the bottleneck somewhere else? If we cut workers now and quality slips in six months, what does rebuilding cost?

AI is the most useful new tool for running a company that I've seen, and Microsoft got real results by redesigning a process. Block cut first, and its results are still coming in. The companies that reversed their cuts made them before they had shown that AI could do the work. Show that AI can do the work on one process before you cut. More on separating activity from results in AI productivity isn't AI value, and on what changes for the people who stay in AI won't replace your team.

Starting January 19, 2027, I'm leading the Executive AI Build, a four-week live online program for CEOs and founders of small software companies. Each leader picks one business process and works on it with the person on their team who will run it. They finish with that process running with AI, owned by that person and measured against its old results, with a follow-up about six weeks later.