WHY ENTERPRISES MIGHT NOT CAPTURE THE VALUE
- Apr 13
- 8 min read
Earlier this year, KPMG tried to negotiate a fee reduction from its own auditor. The argument? AI makes the process cheaper. This saving should be passed on to the client.
Grant Thornton pushed back. ‘High-quality audits rely heavily on expert human judgment,’ they said, ‘so our fees reflect both the cost of our people and the cost of the technology that supports them.’ But KPMG’s CFO Michaela Peisger won the argument. Fees fell from £416,000 to £357,000. The paperwork is filed at Companies House.
This would be a footnote to Q1 2026 if it were isolated. But insiders at a third consultancy, PwC, told the Financial Times: ‘Clients know that AI is making us more efficient, saving manhours. And now they’re asking for discounts.’
The fee pressure carries a second consequence: if AI means fewer hours billed, firms need fewer people billing them. UK Big Four graduate accounting job listings fell 44% in a single year. KPMG cut its graduate intake by 29%. PwC US announced plans to reduce entry-level hiring by a third over three years. McKinsey, BCG, and Bain froze starting salaries for three consecutive years. Echoing the warning from ESRI and the Department of Finance last week that AI adoption is likely to lead to job losses among educated workers, with falls in income tax receipts and rises in welfare spending.
And there is little doubt that fee reductions are happening across the most prestigious knowledge-work institutions in the world. The conventional reading is straightforward: AI is making professional services more efficient, so firms need fewer people and clients pay less. Productivity up, headcount down, fees under pressure. The business case of AI is working.
The conventional reading makes sense as far as it goes. Aaron Levie, CEO of Box – a cloud content management company with a $5 billion market cap – describes his engineering team building five times faster than two years ago. Box doesn’t charge its customers more. The gains are packed into existing licences because competitors are doing the same thing. A KPMG auditor is asked to lower fees. A Box customer gets five times the engineering output for the same cost. The productivity gains are real. The efficiency is measurable. And the value is flowing somewhere specific: to customers, to clients, to the organisations buying these services.
But the firms that invested in AI, trained their people, redesigned their workflows – they’re watching their gains evaporate into someone else’s margin.
The business case is working. The question is: for whom?
When a technology makes something abundant, we assume it amplifies what we already do. More productive engineers. Faster audits. Better diagnoses. The assumption underneath every AI business case is that abundance equals more of the good things the scarce resource produced.
Evidence from economics, neuroscience, clinical medicine, and economic history points in a different direction. Abundance consistently operates as displacement. It makes the previously scarce thing less valuable while creating demand, costs, or consequences in directions nobody planned for.
This is worth stating early: what follows is a broadly pro-AI argument. Displacement can be enormously beneficial – consumers benefit when competitive dynamics drive prices down. We have far more sophisticated iPhones for the same price as the original because compute is a lot more abundant than almost two decades ago.
The question is whether organisations understand where the value actually flows to when they invest in AI abundance. The KPMG story suggests many don’t. And fee pressure may be only the most visible portion of a larger pattern.
Most commentators have noticed this dynamic in at least one dimension. Since DeepSeek’s release in January 2025, the Jevons Paradox has become the standard frame: AI makes knowledge work cheaper, so we do more of it. Satya Nadella tweeted ‘Jevons paradox strikes again!’ and a 160-year-old economic observation became the lingua franca of the AI discourse. Azhar, Brynjolfsson, Evans, Andreessen, Thompson, and Levie have all made the argument. And, as you’d expect given this veritable brain trust, they’re right.
The evidence is already visible. Sridhar Krishnan, CEO of an enterprise AI startup, told Azeem Azhar that most of his company’s token growth comes from queries they wouldn’t have bothered running before AI made them cheap enough to attempt. Levie describes the same pattern at enterprise scale – Box’s customers aren’t just doing old work faster. They’re attacking problems that were never economically feasible to solve before. Demand creation is real, it’s measurable, and for many categories of knowledge work it will dominate the picture.
But Jevons is conditional. It holds where demand is elastic – where there is latent demand waiting to be unlocked by falling costs. Agriculture is the counter-case. Farming productivity tripled across the twentieth century. Food demand didn’t triple. What happened instead was that farms consolidated, margins collapsed, and millions of agricultural workers left the land. The efficiency was real. The amplification wasn’t.
Some categories of knowledge work will likely follow the Jevons pattern. Others will be more agricultural.
But demand creation, even where it operates, is only one of the ways abundance moves value around. There are at least three others – less discussed, less visible in the near-term data - that do, we believe, also carry weighty strategic consequences for most organisations.
Competitive absorption - the gains flow elsewhere
KPMG and Box are telling the same story. In competitive markets, productivity gains are competed away. The surplus flows to customers, not to the firms that invested. The organisation runs faster on a treadmill that doesn’t move.
This has a precise economic history. William Nordhaus, the Yale economist who won the 2018 Nobel Prize, spent years measuring how much of the value created by technological innovation is captured by the innovators themselves.
His answer may surprise you. It was just 2.2%.
And this is worth rolling around your head for a few more seconds.
The work of a Nobel Prize economist revealed that innovators captured just 2.2 cents of every euro of value their innovations created.
The other 97.8% flowed to consumers, to competitors, to adjacent industries – to people who did not do the innovating. In any reasonable sense.
True, that number is a long-run average across all technologies. But for generative AI specifically, the ratio appears even more extreme.
Erik Brynjolfsson and colleagues estimated that Americans enjoyed roughly $97 billion in consumer surplus from generative AI in 2024. The producers of that value - OpenAI, Anthropic, Google and Microsoft - captured just $7 billion in producer revenue. Which you will realise, as you’ve already done the maths, is a ratio of approximately 14:1.
For every dollar the AI companies captured, consumers enjoyed fourteen dollars of value that doesn’t appear in anyone’s revenue line.
Sam Hammond, senior fellow at the Foundation for American Innovation, put it bluntly on the Cognitive Revolution podcast: ‘AI is a machine for converting GDP into consumer surplus.’
And here’s the real bummer. Consumer surplus is ethereal. It’s the value you enjoy but nobody monetises. It's your more sophisticated iPhone that you’re not paying any more for. It’s the extra features Box delivers at no cost to its customers. Its the fall in audit fees.
All these things do make the world better. We can even measure by how much. And yet they often make the revenue lines of the companies producing the value only marginally better, no better or even measurably worse. Abundant things, even when they are the most useful things, become the least profitable to sell. Water is far more useful and widely required than diamonds. And yet…
The agricultural treadmill might now be mapping onto AI adoption with uncomfortable precision. Willard Cochrane, the agricultural economist, coined the term in 1958 to describe the cycle he was watching destroy American farming communities. An early-adopting farmer invests in new technology – better seed, better equipment, better technique. For a season or two, the farmer enjoys higher yields at the same cost. Margins improve. The investment looks smart.
Then neighbours adopt the same technology. Yields rise across the market. Prices fall. The early adopter’s brief advantage disappears, absorbed into lower food costs for consumers.
The farmer is back where they started – but now with the capital cost of the new technology embedded in their cost base. Cochrane described the 1950s: ‘Agricultural output increased at a rapid pace, and all of it as a result of increased agricultural productivity. However, the decade is remembered as a time of hardship for most farming families. A million and a half farming families ultimately gave up farming.’
The data spanning seven decades is equally stark. US agricultural output nearly tripled between 1948 and 2021 while aggregate inputs remained essentially unchanged. Labour hours fell by more than 80%. The percentage of US disposable income spent on food prepared at home dropped from 22% in 1950 to 7% by 2000. American farms collapsed from 6.8 million in 1935 to 1.88 million by 2024. A study of 25 EU countries confirmed the same pattern internationally, describing it as ‘Cochrane’s Curse.’ The gains were enormous. They went entirely to consumers.
Warren Buffett saw the same dynamic in aviation. ‘If a farsighted capitalist had been present at Kitty Hawk,’ he wrote in 1999, ‘he would have done his successors a huge favour by shooting Orville down.’
Between 1945 and 2000, the US airline industry operated at a combined net profit margin of 0.8%. Extraordinary value created – the ability to fly anywhere on earth in hours – and virtually none of it captured by the firms that produced it.
The counter-argument is fair and should be stated plainly: competitive absorption is a long-run average. Individual companies can and do capture enormous value during transition periods. From which it is possible to create seriously profitable businesses. Amazon, Google and Apple built durable moats – through network effects, proprietary data, and integrated ecosystems – that allowed them to retain a larger share of the surplus they created than Nordhaus’s 2.2% would suggest.
Competitive absorption, therefore, operates on a spectrum. It bites hardest in commoditised, transparent markets with low switching costs. It bites least where regulatory barriers, brand strength or proprietary assets create genuine differentiation.
The question for enterprise leaders evaluating their AI investments is where on that spectrum their industry sits. Professional services, where the KPMG story is playing out, sits toward the exposed end. A KPMG audit and a Deloitte audit are substantially similar products in a market where clients can and do switch. An AI-assisted audit and a human-only audit are converging to the same output quality, which means the differentiator collapses to price. The treadmill is already turning.
Cornelia Walther, a visiting scholar at Wharton, identified the feedback loop in a piece for Knowledge at Wharton. She described a four-stage cycle: initial productivity gains create worker agency and capacity; managers recalibrate expectations upward; workers delegate increasingly complex tasks and skill atrophy begins; each improvement becomes the new baseline, locked in.
‘We are making ourselves ever more dependent,’ she wrote, ‘on the assets that are making us redundant.’
She calls it the AI Efficiency Trap, and the mechanism she described is competitive absorption playing out inside the firm – the gains are real and continuously captured by whoever sets the performance benchmark.
Competitive absorption is uncomfortable, but it’s navigable. It shows up in financial data. It responds to strategic positioning – moats, differentiation, proprietary assets, pricing power. A leadership team that understands competitive absorption can plan around it. The value flows to customers? Then position yourself as the platform, not the commodity. Build the seed company, not the farm.
The harder problem is what’s happening in the dimensions that financial data can’t see.
Those graduate roles that disappeared – 44% fewer in a single year – were the mechanism through which junior auditors developed professional judgement over years of manual, painstaking work. The fee pressure and the hiring collapse are the visible and invisible faces of the same dynamic. One shows up on a balance sheet this quarter. The other shows up a decade from now, when the profession discovers it no longer has enough experienced auditors to train the next generation of subject matter experts.
And that’s only one of the costs that don’t show up in the data. There’s another, quieter, operating at a different scale entirely – in the quality of the decisions organisations make with all this additional AI-generated information flooding through them.
These are the costs that make it a trap. They accumulate with the people and organisations least equipped to see them.
All of which is the subject of Part 2.
