The ROI Calculation Your Board Approved Does Not Include the Real Costs

Meta's AI transformation numbers looked right on paper. The gulag, the petition, and the memo were not in the model.

Sign up for my Substack daily AI newsletter here.

See my AI Training course portfolio for corporate Business Leaders here.

Article link: https://open.substack.com/pub/johanosteyn/p/the-roi-calculation-your-board-approved?r=73gqa&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true

Follow me on LinkedIn: https://www.linkedin.com/in/johanosteyn/

On 12 June 2026, Mark Zuckerberg sent an internal memo to Meta’s employees. Reuters obtained it. In it, the CEO of one of the world’s most valuable companies acknowledged that Meta had made mistakes in the AI transformation of its workforce and would almost certainly make more. He promised stability. He said he did not want to overpromise because the world is changing in ways that are out of his control. The memo was unusual in its candour. It was also, in a very specific sense, the bill arriving for a business case that had never included the items it was now paying for.

The infrastructure was costed. The licensing was negotiated. The headcount reductions were modelled. The gulag was not in the spreadsheet.

CONTEXT AND BACKGROUND

The specifics of what happened at Meta in the months before Zuckerberg’s memo are important because they are not abstract. In May 2026, Meta laid off approximately 8,000 employees — roughly 10 per cent of its total workforce. A further 7,000 employees were forcibly reassigned to a three-month-old unit called Applied AI Engineering, a division of roughly 6,500 engineers and product managers assembled to generate training data for Meta’s AI models. The work assigned to these engineers was not software development. It was puzzle generation and coding problem creation — tasks designed to teach AI models how to think, performed by people who had previously been architecting social media infrastructure at a global scale.

The response from inside the unit was unambiguous. Employees described it as literally the gulag. Others called the work soul-crushing. They called themselves draftees — people who had been given no real choice but to join or leave the company entirely. During a live-streamed internal presentation this week, an employee hijacked the feed with an expletive-laden outburst demanding that attendees tell a senior Meta AI executive precisely what they thought of him. Meta’s Chief Product Officer Chris Cox acknowledged the environment had been brutal. More than 1,600 employees across the company signed a petition against a separate programme monitoring their clicks and keystrokes to generate AI training data. Meta scaled the programme back slightly, allowing employees to pause data collection for up to 30 minutes — a partial concession that satisfied almost nobody.

The Applied AI unit is led by Maher Saba, former vice president of Reality Labs — the division that spent 83 billion dollars on the metaverse before Meta shifted focus. That detail is not incidental. It is a pattern.

INSIGHT AND ANALYSIS

The financial model that justified Meta’s AI transformation included infrastructure costs, compute spend, licensing arrangements, and projected productivity gains from a workforce restructured around AI. It did not include the cost of treating 6,500 highly skilled engineers as data labelling conscripts. It did not include the legal exposure of a keystroke monitoring programme that 1,600 employees considered invasive enough to petition against. It did not include the productivity loss of a demoralised workforce, the institutional knowledge that walked out with 8,000 redundancies, or the reputational cost of a gulag comparison appearing in Wired. These are not edge cases or unforeseeable consequences. They are the predictable human costs of a transformation strategy that calculated everything except the people it would be executed by and on.

The unit’s management architecture made that indifference structural. At inception, up to 50 employees reported to a single manager — against a standard software engineering ratio of one manager to six or eight contributors. At fifty to one, the unit was not a team. It was a processing queue, and the engineers inside it knew it.

This is the structural incompleteness at the heart of most AI ROI models, and Meta has just provided the most public and detailed accounting of what that incompleteness costs in practice. When boards approve AI transformation investments, they model what AI will do to the business. They rarely model what the transformation will do to the people, and they almost never model what the people will do in response. The petition, the gulag, the hijacked livestream — these are not failures of communication or change management. They are rational responses by skilled people to a strategy that treated them as implementation variables rather than as the human architecture the transformation required.

I have previously written about the pattern of worker resistance to AI automation, and the long historical argument — from the Luddites of 1811 to the draftees of 2026 — that what looks like resistance to technology is almost always resistance to a specific distribution of technology’s costs and benefits. The Meta story is the same argument with a new cast. What the Applied AI unit introduced was something the Luddites would have recognised immediately: algorithmic Taylorism — the application of industrial-era time-and-motion logic to cognitive labour, reducing engineers who once designed systems to workers who generate two tasks per week on a quota. The puzzle generation work was not merely dull. It was deliberately atomised, stripped of context, and disconnected from outcome — digital piecework for minds that had been hired to build things. The engineers who signed the petition were not opposed to AI.

They were opposed to a model in which their skills, their data, their working hours, and their professional identities were extracted for the benefit of a system whose productivity gains they would not share and whose decisions they had no part in making.

The ROI model that approved Meta’s AI transformation did not include that argument. It should have.

IMPLICATIONS

The lesson for boards and executives is not that Meta spent too much. It is that Meta calculated the wrong things. The variables missing from most AI transformation business cases are the ones that determine whether the transformation actually produces the capability it was designed to create. A workforce that has been demoralised, surveilled, and treated as a data source does not produce the quality of output that a willing, engaged, and purposefully directed workforce produces. The 83 billion dollars spent on the metaverse did not fail because the technology was impossible. It failed because Zuckerberg could not get enough people — employees, developers, consumers — to genuinely want what he was building. The Applied AI unit is the same pattern applied internally.

Before any AI transformation investment is approved, boards should be able to answer four questions in plain language. Who will execute this transformation, what will their working conditions look like, and what genuine choices — not join-or-leave ultimatums — are being offered to the people whose roles are changing? What monitoring or surveillance does the transformation require, and what is the legal and employee relations exposure of those programmes? And critically, what feedback mechanism exists to surface problems before they become petitions, viral moments, or CEO memos acknowledging mistakes that were entirely predictable? Most AI transformation business cases cannot answer any of those questions. Meta’s June 2026 memo is the consequence of that omission — not a confession of unexpected failure, but a reckoning with costs that were always present and simply never accounted for.

For South African boards and executives, the Meta case lands in a specific and legally consequential context. South Africa’s Draft National AI Policy, gazetted in April 2026 - then withdrawn - explicitly positions employment decisions and workforce restructuring as high-risk AI applications requiring stringent governance. More immediately, Section 189 of the Labour Relations Act imposes a joint consensus-seeking consultation process on any employer contemplating retrenchments or significant changes to operational requirements — and the forced reassignment of thousands of employees to fundamentally different roles, under the guise of an AI pivot, is precisely the kind of operational shift that triggers that obligation. An employer who unilaterally reassigns staff into algorithmic Taylorism without meaningful consultation does not simply risk low morale. They risk procedural unfairness claims at the CCMA or the Labour Court, filed before the transformation has produced a single productive output. The costs that were not in Meta’s model are not only human and reputational. In the South African context, they are statutory.

CLOSING TAKEAWAY

Zuckerberg’s admission that Meta has made mistakes and will almost certainly make more is more honest than most CEO communications about AI transformation. It is also more honest than the business case that approved the transformation in the first place. That business case was structurally incomplete — not because the numbers were wrong, but because the variables that determine whether a transformation succeeds were never included in the model.

The organisations that learn from Meta’s experience will not simply add a workforce management line to their AI transformation budgets. They will ask, before the investment is approved, whether the people who will execute the strategy have been treated as participants in it or as variables within it. The answer to that question is the most reliable predictor of whether the transformation will produce the capability the board approved, or the memo that Zuckerberg sent on a Friday in June 2026.

Johan Steyn is a prominent AI thought leader, speaker, and author with a deep understanding of artificial intelligence’s impact on business and society. He is passionate about ethical AI development and its role in shaping a better future. Find out more about Johan’s work at https://www.aiforbusiness.net