AI pioneer Geoffrey Hinton said the quiet part out loud in a December interview with Fortune Magazine – that “companies investing trillions in artificial intelligence can only make their money back by eliminating human jobs.”

This stark assessment reflects the staggering levels of investment in generative AI infrastructure. To recoup this, a Bain study found, AI tech firms will have to find an additional $2 trillion in annual revenue by 2030. Given that the current technology spend of U.S. companies is $2.7 trillion, the percentage increase in companies’ tech budgets would be unprecedented and almost inconceivable without massive across-the-board layoffs.

The rapid adoption of ChatGPT, which grew to nearly 100 million users in its first two months, is certainly the most significant tech development in recent memory. Millions of workers claim significant productivity gains from the technology in tasks such as drafting routine emails, booking travel, creating financial reports, or producing training and marketing materials.

Nevertheless, tech companies will need a lot more than individual users to generate $2 trillion in revenue. Accordingly, they are aggressively promoting a concept that the MIT Industrial Productivity Center (IPC) calls zero-sum automation – tech deployments that are expressly designed to reduce costs by replacing people. Tech firms like Amazon, perhaps hoping to lead by example, have announced major layoffs which they claim they will replace at lower cost with AI.

Zero-sum automation, however, has not fared well. Since the 1980s, the tech industry has been dogged by concept called Solow’s Productivity Paradox – so named from a quip from economist Robert Solow that “you can see the computer age everywhere except in the productivity statistics.”

The problem is stubbornly persistent. IPC found in a 2022 study that despite significant tech investments in industrial automation in the 2010s, productivity actually declined. The reason cited, which was reported in a 2023 Harvard Business Review article, is that any gains made in productivity were erased by loss of flexibility – in the end, companies had to hire more people to adapt these automation systems to keep pace with evolving market conditions.

Recent generative AI deployments in white-collar environments are suffering from a similar pattern. According to The GenAI Divide: State of AI in Business 2025, a multi-year study by MIT, 95% of AI pilots failed to demonstrate a measurable ROI despite $30 – 40 billion invested in generative AI. Again, lack of flexibility is a major factor. “Most fail due to brittle workflows, lack of contextual learning, and misalignment with day-to-day operations,” says the study.

Companies aligned with the tech sector have been predictably more aggressive in their AI deployments, but even they are facing headwinds. Swedish fintech Klarna, once regarded as a flagship business case for AI deployment, had to walk back its plans to replace 700 customer service employees with AI amid declining quality and disappointing efficiency gains. Several other tech firms, including Salesforce, have recently revised their claims about AI-related productivity gains.

Deceptive viewpoints

White collar and factory environments alike have complexities that are not obvious to outside observers and don’t show up in the financials. Leaders, consequently, operate from a dashboard that grossly oversimplifies the dynamics of a typical work environment, and makes workplace automation seem much more straightforward than it is.

“We tend to reduce complex human organizations to mechanical models,” says Ken Eakin, senior consultant at Ottawa-based Lean Agility and author of the book “Office Lean.” “So, it seems a reasonable assumption that we can pull a lever by introducing AI and get rid of Joe because AI can do what Joe does. But when we consider all the interactivity and interdependence of the average office task, it’s far more complex than the sum of individual job responsibilities.”

This becomes clear when we look at work in the context of a value stream. “There are different ways of measuring work,” says Eakin. “In a very basic sense, you can look at the time somebody spends working on a file or project – fingers on a keyboard, thinking and analysis, etc. But in a white-collar scenario, there are lots of handoffs and back-and-forths, and people might have fifty things on the go. So, everybody’s extremely busy, but when you look at the timeline of a typical project, increasing the efficiency of the taskwork of the individual is worthwhile, but it only attacks 5% of the problem. Therefore, you really have to be thinking about the larger process.”

Taking a larger process view can’t be just about automation – it’s also about considering questions like ‘Are all these emails necessary?’ of ‘Why does it take so long for us to respond to customer requests?” But that takes time, experimentation, and a willingness to collaborate with the workers who live with the processes day-by-day – the very people large-scale AI advocates hope to eliminate.

A better approach

The good news reported by the above-mentioned MIT study is that there’s a booming “shadow AI economy” in which un-official AI projects are being successfully initiated at a workgroup level. The study suggests that this could be the key to solving the divide that’s stalling 95% of AI pilots.

“Forward-thinking organizations are beginning to bridge this gap by learning from shadow usage and analyzing which personal tools deliver value before procuring enterprise alternatives,” says the study.

Some experts in the AI industry subscribe to this problem-solving approach. “You should never ask, ‘Where can I use AI?’ says Sheldon Fernandez, AI strategist and founder and former CEO of DarwinAI. “You ask, ‘What processes do I have that are inefficient and where some level of automation could be useful?’”

The deployment of AI in the context of process improvement will be familiar to companies dedicated to lean practices. Toyota has always approached technologies as a problem-solving tools, and requires prospective implementers to follow a disciplined test and verification process, and to consider alternatives that don’t require technology.

It’s also important to bring the shadow AI economy into the mainstream, and to give project champions the resources to develop projects that could, if successful, scale to an enterprise level. “You should treat these initiatives as programs, not pilots,” says Fernandez, “programs with executive sponsors, product owners, governance, and infrastructure. That’s the only way I’ve see this work.”

Companies with strong lean teams and a track record in scaling kaizen improvements may have the ultimate edge when it comes to making significant productivity gains with AI. If AI is done right, it won’t be just about technology – it will be about taking a deeper look at the work processes that companies hope to improve.

The emergence of AI, accordingly, presents a golden opportunity. This was aptly expressed by ChatGPT 5.2 in response to a query about AI’s strengths and weaknesses:

“Artificial intelligence may be the tool that replaces repetitive work, but its greater gift is diagnostic. It shows us, with uncomfortable clarity, where our systems were wasting human potential all along.”