YOUR TURN

In April, this series opened with auditor Grant Thornton agreeing to cut its fees. Its client, KPMG, argued that AI had made its audit cheaper. And that this saving belonged to the client. The bill fell from $416,000 to $357,000.

Eight weeks and one earnings season later, the question the story raised hasn't gone away. When AI makes knowledge work cheaper, who keeps the value?

Parts 1 and 2 described four ways AI abundance moves value around.

Competitive absorption: productivity gains flowing to customers rather than to the firms that invested.

Capability erosion: human skills degrading while AI-assisted performance metrics improve.

Sense-making collapse: information arriving faster than the institutions can filter it.

Demand creation: work that wasn't worth doing becoming viable once intelligence is cheap.

Different mechanisms, different domains. But they share a structural feature: each produces measurable gains in one dimension while degrading something in a dimension the gain metric doesn't capture.

So far this series has only offered a single strategy, in passing, to cope with the effects – build the seed company, not the farm. This final part looks at potential responses more seriously.

The seed company question

The agricultural treadmill we described in Part 1 produced an enormous transfer of value from producers to consumers. US agricultural output nearly tripled between 1948 and 2021. Food's share of household income collapsed. The number of American farms went from 6.8 million to 1.88 million. The technology that made this possible – hybrid seed, mechanisation, synthetic fertiliser, eventually GPS-guided precision agriculture – was available to everyone. The same input. Two radically different outcomes.

The farmers, in aggregate, paid for the technology and watched the gains flow through to consumers as lower food prices. Most of them left the land. The seed companies, the machinery firms and the agrochemical providers didn't have to be more efficient than the farmers. They had to be in a different position – upstream, selling the inputs, sitting where the value flowed.

The technology, in other words, wasn't the variable. Position was. That's the question for every organisation weighing its AI spend. Which side of each dynamic are you on? Most sit on the seed company's side of one dynamic and the farmer's side of another, usually without noticing. The job is to know which, and to invest accordingly.

Who's winning?

Five kinds of winners are emerging.

The first is the largest: customers. When the big providers compete on the price of inference, the savings don't accrue to the providers. They accrue to the enterprises buying it. The pattern Brynjolfsson and Collis measured at 14:1 – some $97 billion in US consumer surplus from generative AI in 2024 against $7 billion in producer revenue – is that dynamic at scale. KPMG's fee reduction is the dynamic in microcosm. So is every audit, legal review and analysis function quietly absorbing fee pressure as AI becomes routine. In any market with competition, the customer is the structural winner of AI abundance.

The second is the seed-company position itself: organisations that have placed themselves upstream of the absorption. John Deere is a classic case. In 2017 it paid $305 million for Blue River, a computer-vision startup, and turned it into See & Spray, a system that tells a weed from a crop in real time. Deere reports five million acres covered in the 2025 season, with non-residual herbicide use down nearly 50%. Deere charges for the machine, then by the acre for the technology. The farmer banks part of the saving; Deere banks the rest. The tractor company became a data company. But upstream isn't a free pass: the model laboratories sell an input too, and that 14:1 ratio shows how little of the surplus they've kept. The input only pays where something about it stays scarce – capacity, data, distribution, the cost of switching away.

The third is organisations whose moats predate AI – and who now point AI at them. Wingate, Burns and Barney named the mechanism in MIT Sloan Management Review: AI itself isn't a source of advantage, because it's widely available. Advantage comes from what they call residual heterogeneity – the proprietary data, regulatory standing, relationships and integration the AI is pointed at. Regulated industries are full of it: pharma's GxP regime, insurance's actuarial and licensing position, banking's prudential supervision. These were always costs. But they also work as filters on who can credibly enter. The regulatory burden, expensively maintained, is the seed-company position. Most organisations in regulated industries don't think of compliance this way. Perhaps more should.

The fourth is the inverse of Part 2's capability erosion story. Where the radiologists, the endoscopists and the pilots lost capability through routine AI use, others are keeping it on purpose. The mechanism isn't talent or seniority. It's a design choice: across the studies – suggestive rather than settled – AI used as a scaffold preserves the skill underneath, while AI used as a replacement consumes it. A few organisations now treat preservation as an explicit goal of deployment. It will be interesting to watch their paths.

The fifth is the response to the sense-making problem: organisations investing in filtering alongside production. TREWS, a sepsis early-warning system built at Johns Hopkins, shows the difference. In Part 2 we discussed clinicians overriding most decision-support prompts. But with TREWS patients whose alerts were confirmed within three hours had an 18.7% lower mortality ratel. The system was built for discrimination, so it earned attention, and attention saves lives. Much enterprise AI spend goes to production. Little goes to filtering. This asymmetry is the opportunity.

Who's paying?

Four costs are emerging, often in parallel.

The first mirrors the customers' win: where they kept the gains, the suppliers paid. BCG finds only 4% of companies creating substantial shareholder value from their AI investments. But, in the short-term at least, there are companies reaping the benefits. Payments company Block cut more than 4,000 roles in February – around 40% of its workforce – and raised its profit guidance the same week. It watched its shares jump almost a fifth.

Which leads us to the second cost, which seems to be landing on the young first. Stanford's 'Canaries in the Coal Mine' measured a 16% relative employment decline for workers aged 22 to 25 in the most AI-exposed occupations, after controlling for firm-level shocks. In the UK, accountancy graduate listings fell 44% in a year, well ahead of the graduate market overall. At the Axios AI Summit in March, Senator Mark Warner relayed that a major law firm had stopped hiring first-year associates. Juniors are how a firm grows its senior judgement. A thin intake now is a thin bench in 2036.

The third cost is harder to see. Organisations measuring AI-assisted performance without measuring unassisted capability are running up a debt that won't appear until the AI is unavailable. Part 2's colonoscopy finding is a leading indicator: a 20% relative decline in adenoma detection without AI, after routine AI exposure. Assisted performance had improved. Unassisted performance had fallen. Nobody was measuring the second number. By-mind skills – judgement, situational awareness, knowing when something has gone wrong – degrade silently while by-hand skills look stable. Most organisations are running this experiment on themselves without a control group.

The fourth cost piles up around decision-makers: more information than ever, less capacity to act on it. Part 2's figures – two-thirds of security alerts ignored, most clinical prompts overridden – show what this looks like where the consequences can be counted. More AI output, with no better way to sort it, deepens the problem. And in March it acquired a governance case to remember. A security firm's autonomous agent gained full access to McKinsey's internal AI platform, Lilli, in around two hours, through a 1990s-era exploit. A research exercise, responsibly disclosed, with no client data shown to have been taken – but the question 'who controls the AI in our workflow' got answered late, and by an outsider. There will be others.

The costs, across all four, accumulate with those least equipped to see them. Demand creation sits outside both lists – it's an opportunity, not a cost – which is why it waits for the choices.

The four choices

The map leaves four choices – one per dynamic. Most businesses are making them already, whether leadership teams have noticed or not.

The absorption choice – position. Where do you sit in the value chain? Deere chose seed company. Not all enterprises have Deere's options, but every enterprise can ask Deere's question: what part of our operation is becoming a commodity, and what part a platform? If competitors can buy the same model and replicate what you do with it – and for most AI applications they can – absorption is coming. But can it be held off long enough to fund the move upstream?

The preservation choice – capability. Is capability being built, or rented? AI deployed as scaffold appears to preserve the skill underneath; deployed as replacement, it consumes it. Most enterprises don't notice they're making this choice – the radiologists who can't read without AI, the juniors who never joined, the pilots who can fly but can't judge are paying for decisions nobody recorded as decisions. Tracking unassisted capability takes deliberate effort, and almost nobody does it – which means most organisations can't tell whether theirs is being preserved or consumed. Start there.

The filtering choice – time. What do you do during the institutional gap? Part 2 put the printing-press version of that gap at 150 to 200 years. We won’t have that long. The choice during the gap is between production – more output, faster – and filtering – better discrimination of what matters. TREWS earned answers because it filtered first. In the programmes we see, production gets the budget and filtering gets the leftovers. Reverse that.

The demand choice – direction. The fourth decides whether demand creation ever reaches you, and it's the easiest to misread. Call it efficiency AI versus opportunity AI: the first does the same work with less; the second opens work that wasn't worth doing – or wasn't possible – before. Efficiency compounds inside the competition you already have – exactly where absorption bites hardest. Opportunity builds value in territory the market hasn't yet standardised around, and choosing it means living with worse numbers for a while. It's a bet against your own scorecard.

The diagnostic

If you want to know where a business currently sits on the abundance trap map, there are five questions to consider. Knowing can, obviously, help leaders plan deliberately, so that the coming abundance does not become an actual trap for the business.

The absorption question. Where are AI gains staying with your firm, and where are they flowing through to customers? What evidence do you have either way?

The preservation question. Which capabilities are you keeping alongside AI use, and which are you spending without noticing? What would unassisted performance look like as a number – and have you ever measured it?

The filtering question. Are you investing in discrimination at anything like the rate you're investing in production? Where is your TREWS – the layer that decides which AI output deserves attention?

The opportunity question. What share of your AI spend is efficiency, and what share opportunity? Did you choose the split, or did the measurement system choose it for you?

The governance question. Who controls the AI in your critical workflows?

Answer these and you are already moving closer to taking as much control as possible.

The bottom line

The series began with the assumption that abundance means amplification – more of what we already do, faster and cheaper. The evidence pointed the other way, to displacement: the once-scarce thing loses its value, and the consequences turn up where no one was looking.

How much of this is already real? Enough to act on. The fee pressure and the consumer surplus are visible now. The erosion of judgement is real but early. The strain on sense-making is the thinnest of the three, a pattern rather than a proof. KPMG and Box. Accenture and Block. The Stanford canaries. The colonoscopy study. The alert dashboards. The honest answer to the question Part 2 left hanging is that more value often escapes than the business case assumes, and less than a pessimist fears.

Which returns us, finally, to the farm. The treadmill consolidated American farming to a fraction of its former numbers, even as it tripled what the land produced. Some operations became the seed companies. Most became smaller. Many disappeared. The technology was identical on every field.

That, in the end, is the abundance trap. The technology that creates the abundance doesn't decide who captures the value. Position does. And position, unlike the weather, is still a choice – whether an organisation makes it deliberately, or by default.