Will AI replace your team? What IKEA and Meta show

On 26 August 2026 Reuters published how Meta's plan to hand much of its employees' daily work to AI agents fell apart in six months. That same week, the IKEA case went around again: the company used artificial intelligence to free up 8,500 customer service workers and laid off none of them.

The two cases get quoted as opposites: one failed, one worked. Except the difference between them is not the technology — both companies used the same generation of tools. It is what each one chose to measure.

What Meta tried, and why it stopped

The plan had an internal name: Project OT, for Organization Transformation. It came out of the January 2026 leadership retreat and described an “AI native” company where agents would take over much of the daily work, supervised by smaller human teams. Scenarios went as far as shrinking many teams by up to 60%, in two rounds of cuts: one in May, another in November.

On the evening of 19 May, hours before the first round, Zuckerberg cancelled the second. Meta still cut about 10% of its workforce the next day, but the transformation stopped there.

What surfaced in internal posts explains the brakes better than any statement:

• AI-made code changes across the company's platforms grew 220% year over year.
• Improvements that actually reached users grew 36%.
• Serious technical and security incidents rose 40%.
• Internal sentiment fell from 74% to 55% once employees concluded that the keystroke and mouse-click tracking software installed on their machines was training the agents meant to replace them.

In July, Zuckerberg admitted at an all-hands that agent technology had not sped up as much as he expected.

What IKEA did with the same problem

The IKEA case is not news — worth saying, because it circulates as if it were. The statement from Ingka Group, which runs most of the stores, is dated 29 June 2023.

The virtual assistant Billie came to resolve roughly 47% of the enquiries reaching customer service — 3.2 million interactions — saving about 13 million euros between fiscal years 2021 and 2023.

That is the point where most companies cut staff. Ingka did something else: it reskilled 8,500 call centre workers into remote interior design, digital selling and complex problem solving. The advisory channel born from that generated 1.3 billion euros by the end of fiscal 2022, equal to 3.3% of total sales, with a stated target of reaching 10%.

⚠️ One correction that usually slips past: that 1.3 billion is revenue from a sales channel, not “profit the AI generated”. People retelling the case tend to merge the two.

The difference was not the technology: it was the measurement

Put the two dashboards side by side.

Meta measured AI output: how many code changes the agents produced. By that ruler the project was a runaway success — 220% in a year. The company only found the problem when it looked two steps further down the chain: what reached the user (36%) and what broke along the way (+40% in incidents). Between producing and delivering there was a sixfold gap, and it was being paid for by the very people the plan meant to let go.

Ingka measured something else: what the bot could not resolve. Instead of celebrating the 47% it handled, it went and studied the remaining 53%.

What was hiding in that 53%

Reading the conversations Billie could not close, Ingka found a pattern: they clustered around one question — will this work in my living room? Size, colour, match, space. Questions that need taste and context, and that a scripted call centre agent never had the time or training to answer.

In other words: demand for interior design advice had been knocking on the customer service door for years, treated as a ticket to close quickly. AI did not create that market. It only cleared the noise sitting on top of it.

That is where the transferable lesson comes from, and it fits in one line: the log of what your AI cannot resolve is the map of what you should be selling.

The three measures missing from most dashboards

1. What reaches the customer, not what the AI produces. Volume of text generated, code written or tickets answered is not a result. The result is what got published, delivered or resolved without coming back. If you only track the first number, you will think you are winning while the correction queue grows behind you.

2. The cost of cleaning up. Meta watched incidents climb 40% while output climbed 220%. In a small company that shows up as rework: the text that had to be rewritten, the quote that went out wrong, the customer who got the wrong answer. That time is part of the cost of AI and almost never makes it into the calculation.

3. What the AI did not solve. It is the most neglected measure, and the one IKEA turned into 1.3 billion. Every customer service tool logs what it failed to answer. Almost nobody reads it.

The part nobody tells you about either case

Cases like these get sold too neatly. Both have rough edges:

• IKEA did lay people off later. In 2026 Ingka cut around 800 roles and Inter IKEA another 850 — roughly 1,650 in total, less than 0.5% of its 166,000 employees. They were not the reskilled workers, and the stated reasons were complexity, falling consumer confidence and tariffs. But it stands as a warning: reskilling is not a promise of a job for life.
• Meta did not give up on AI. It cancelled a restructuring built on the premise that the agents were already ready. That is different from concluding the technology is useless — and Zuckerberg forecast benefits within three to six months.
• IKEA's numbers come from IKEA. They are from a corporate statement, not an independent audit. I use them because they are specific and dated, but it is worth knowing where they come from.

How this applies to a small company

No small business is going to reskill 8,500 people. The transferable part is the method, and it is cheap:

If you already run automated support, take one month of history and read what it failed to resolve. Not the errors — the questions it could not answer. Group them by topic. If one group repeats, you have found either a hole in your material or a service the customer wants to buy and you do not sell.

If you do not have it yet, the order is the usual one: start with the low-risk repetitive task, measure what comes out the other end, and only then decide what to do with the time freed up. The decision about people comes after the measurement, never before.

And if you already run more than one agent at a time, read first what happens when they meet: the supervision perimeter Meta lacked is exactly the one almost no company builds before hiring the second one.

There is one question both cases answer together: what to do with the time freed up. Meta treated it as a cost to remove. Ingka treated it as capacity to reassign. The second reading produced a sales channel; the first nearly produced an aborted restructuring, falling internal sentiment and a queue of incidents.

The honest summary

There is no universal answer to “will AI replace my team?”. There is a better question: what will I do with the time it frees up — and how will I know it really freed anything?

Whoever answers the second half with a number that measures delivery rather than output decides well in either direction. Whoever does not measure will find out the expensive way, as Meta did: six days from cutting people based on a dashboard that was counting the wrong thing.

Sources

• Reuters — investigation into Meta's Project OT (26 August 2026).
• Ingka Group — “AI and Remote Selling bring IKEA design expertise to the many” (29 June 2023).
• Restructuring announcements from Ingka Group (March 2026) and Inter IKEA (May 2026).