Anthropic, Meta, Block and Klarna show why the main barrier to AI productivity is organisational design, not model capability. Executives have largely stopped asking whether their companies should use AI. The real question is how aggressively they should deploy it — and, more importantly, how quickly they should remove people once AI appears capable of doing part of their work. The evidence so far points to a clear answer. Companies should move aggressively to redesign workflows around AI. They should be much more cautious about removing human capacity before the replacement system has proved that it can maintain output, quality and reliability. The danger is not moving too quickly or too slowly in isolation. It is doing things in the wrong order. That distinction helps explain a widening gap between companies reporting large productivity gains and the majority of enterprise AI initiatives that produce little lasting impact. Anthropic reported in June 2026 that its Claude models were writing more than 80 per cent of the code merged into its own codebase. It also said that the typical engineer was merging around eight times as much code per day as in 2024. Those figures are company-reported and come from the most favourable environment imaginable. Anthropic builds the model, employs engineers who understand it, controls the software environment and can change the tool and the workflow together. It is therefore not a normal enterprise benchmark. But it shows what becomes possible when work is genuinely rebuilt around AI. Across the wider corporate economy, the picture is very different. Also Read: The accordion effect: How AI follows the rhythm of expansion and compression MIT’s Project NANDA found that only about five per cent of enterprise generative-AI pilots reached successful implementation with a sustained operational or financial impact. The figure is directional rather than an audited industry statistic, but the pattern is clear: companies are experimenting widely, yet very few are turning those experiments into dependable production systems. This is not mainly because the models cannot perform useful tasks. It is because performing a task is not the same as operating a workflow. An employee using a general-purpose AI tool supplies the context, checks the answer, handles exceptions and decides what happens next. The human is still the workflow. For an enterprise system to replace that person, it must do much more than generate an answer. It must reach the correct data, remember previous decisions, apply company policy, recognise exceptions, assign responsibility and pass the result into the next process. Most AI demonstrations prove that a model can complete one visible part of the job. They do not prove that the surrounding organisation can safely remove the person. This is where the largest strategic error is emerging. Some companies are beginning to cut labour before they can demonstrate that AI has absorbed the underlying work. Block is the clearest example of the risk. In February 2026, the company announced that it was reducing its workforce by about 40 per cent, taking it from more than 10,000 employees to fewer than 6,000. Its chief executive explicitly linked the decision to AI, arguing that a smaller team using the company’s tools could accomplish more. But Block had also expanded dramatically during the pandemic. Its workforce rose from roughly 3,900 in 2019 to around 12,500 by 2022. Part of the reduction may therefore be correcting previous over-hiring rather than demonstrating pure AI substitution. That ambiguity is important. AI now gives management a justification to remove capacity before its contribution can be measured. If performance improves, the smaller workforce will be presented as proof that AI worked. If performance deteriorates, it will be difficult to separate failed AI conversion from ordinary restructuring, weak demand or the consequences of earlier expansion. A headcount reduction does not prove that the work disappeared. Also Read: AI slop is a strategy problem, not a content problem Meta sits in a different position. It has the models, capital and technical talent to build an AI-centred company, yet its 2026 restructuring shows that even the best-equipped firms face organisational friction. Meta reduced its global workforce by about 10 per cent, moved thousands of employees into AI-related initiatives and then acknowledged that the transition had produced mistakes. Management spans had become too wide in parts of the organisation, and the company had to reconsider how reassigned employees would fit into the new structure. Meta’s financial performance remained strong. This is not evidence that AI reduced its measured productivity. It is evidence that access to advanced models does not remove the conversion problem. The company must still finance the new system, keep the existing business operating, redesign management, reassign people and decide which human capabilities remain necessary. Klarna provides the clearest example of what happens when cost reduction moves ahead of quality. The company leaned heavily on an AI assistant for customer service and said that, at its peak, the system was performing work equivalent to around 700 full-time agents. It handled a large share of customer interactions and reduced response times and transaction costs. Those were real gains. But by 2025, Klarna had started to reverse course. Its chief executive acknowledged that the company had focused too heavily on efficiency and cost, while quality deteriorated in more complex customer interactions. Klarna resumed recruiting human agents and moved towards a hybrid model in which AI handled routine work while people managed exceptions and escalations. The problem was not that the AI failed at the task level. It was that the company optimised the wrong measure. The system performed well on visible volume metrics: conversations handled, response time and cost per interaction. But those figures did not capture the declining quality of the smaller number of interactions that mattered most. Also Read: Singapore, AI, and the rise of emotional outsourcing This pattern produces what economists call a productivity J-curve. When companies adopt a general-purpose technology, productivity can initially remain flat or even fall because the firm must invest in complementary assets before the technology produces its full value. The old system cannot stop while the new one is being built. Companies must run both. They must preserve existing customers, revenue, regulatory compliance and service levels while cleaning data, documenting processes, redesigning controls and training the organisation to operate differently. At the same time, work removed from one role does not necessarily disappear. Exceptions, judgments and relationships often move to managers and remaining employees. Payroll may fall while coordination costs rise. Institutional knowledge can also be lost before it is captured. Much of what makes organisations function is undocumented: which data cannot be trusted, which customer exceptions matter and which apparent shortcuts failed in the past. If the people holding that knowledge leave before the new system has absorbed it, the company may destroy part of the context AI needs to operate successfully. The correct strategic distinction is therefore not aggressive versus cautious AI adoption. It is aggressive conversion versus aggressive cost-cutting. Companies should be aggressive in redesigning workflows, integrating data, building evaluation systems and changing accountability structures. They should be conservative in removing people until the new workflow demonstrates three things at the same time: stable quality-adjusted output, improved cycle times and lower total cost after integration, supervision and capital expenses are included. The minimum test is simple. Management should be able to identify exactly what the AI system now completes, what humans still supervise, how exceptions are handled, what has happened to error rates and whether the full cost of deployment is included in the claimed saving. Without that evidence, a smaller workforce is not proof of AI productivity. It is a bet. Anthropic shows the upside when the organisation is rebuilt around the technology. Meta shows that conversion remains difficult even with enormous resources. Block shows the risk of cutting ahead of proof. Klarna shows how task-level efficiency can damage the wider product when quality is measured too narrowly. The companies that win will not necessarily be those using the most AI or removing the most people. They will be the ones who redesign the organisation quickly enough for AI to become a production system — without destroying the human capacity they still need to build it. — Editor’s note: e27 aims to foster thought leadership by publishing views from the community. You can also share your perspective by submitting an article, video, podcast, or infographic. The views expressed in this article are those of the author and do not necessarily reflect the official policy or position of e27. Join us on WhatsApp, Instagram, Facebook, X, and LinkedIn to stay connected. The post How to capture AI’s gains without wrecking your company appeared first on e27.
Author: John Millar
Source: e27