Skip to content
Intelligence Ownership
ArticlesAI translation· Intelligence Ownership · Teil 1· 12 August 2026

Intelligence Ownership Is Becoming a Business Case

My data, my AI – the follow-up

by 6 min read

Albert Einstein by Yayoi Kusama (not really 🙂)

Back in March I put forward a thesis here: the sustainable competitive advantage in the AI era lies not in the algorithms but in your own data. “It knows the internet. It doesn't know your enterprise” – that was the short version.

To be honest: at the time it was a well-considered but strategic bet. There were good arguments, but not enough hard numbers yet. Now they're on the table. And they're clearer than even I expected.

The analysis: the frontier plateau

Two publications from the past few weeks deserve attention.

Case 1: Bridgewater Associates. One of the largest hedge funds in the world wanted to automate a seemingly mundane task: filtering financial news for relevance. Trivial for experienced investors. Apparently not for frontier models: GPT, Claude and Gemini reached around 50 percent accuracy with simple prompts – a coin toss. After intensive prompt engineering: just under 78 percent. That was the ceiling, no matter how much they optimised or how new the model was.

Then Bridgewater trained an open-source model on labels from its own experts. Result: 84.7 percent accuracy, roughly 30 percent fewer errors than the best frontier model – at one fourteenth of the inference cost.

Case 2: Shopify e-commerce catalogue maintenance. A team built a digital twin of a marketplace workflow – product categorisation, brand verification, policy decisions – and let a small 9-billion-parameter model practise in it via reinforcement learning. Training cost: about $500 of GPU time, three and a half days. The result beat every frontier configuration tested by a good ten points – at a cost of $0.50 per 1,000 decisions instead of $34. At Shopify-scale volume, that's the difference between $7 million and $500 million a year.

The most interesting detail, though, is hidden elsewhere: the five frontier models tested – despite optimised prompts – all landed within a tenth of a point of one another. There is apparently a plateau that no general-purpose model breaks through, however expensive. The data suggest that this plateau isn't an intelligence problem but a context problem: the model knows the internet, it doesn't know your organisation.

The paradox: the prompt tax

Why do the smartest models in the world fail at tasks that are routine for a clerk?

Because a frontier model starts every episode from zero. It knows neither the shop's taxonomy nor the investment team's conventions, nor the silent discretionary calls I described in March as “quiet judgment calls”. Everything it is supposed to know about the company has to be crammed into the prompt with every single call. It's as if the clerk had to relearn his job every time he handled a case. Which, as we know, he doesn't: his context and his knowledge are instantly at hand.

A term is currently taking hold for this: the prompt tax. Anyone who injects their company knowledge via prompt is renting it back – token by token, call by call, forever. In the e-commerce case, the optimised instructions made every API call 28 to 55 percent more expensive. And here's the insidious part: the tax only buys the knowledge that can be written down explicitly. The corner cases that make the difference can't be enumerated by any prompt in the world.

The opposite of a prompt tax is Intelligence Ownership: the company's knowledge doesn't live in the prompt but in its own weights. It isn't rented back per call; it belongs to you – like owning a property instead of renting a flat.

Rented intelligence you pay for per call. Your own intelligence you pay for once – and from then on it lives in the weights.

This is exactly the asymmetric value creation I warned about in the spring – except now it comes with an invoice.

The uncomfortable consequence

For the Big Tech providers, the prompt tax isn't a bug, it's the business model. Every organisation that permanently channels its institutional knowledge through someone else's APIs is financing the scaling of those it depends on. Uber shows how expensive that can get: the company reportedly burned through its annual engineering budget for AI APIs within four months – tokenised into oblivion, you might say.

For European companies, something remarkable follows from this: the sovereign route is no longer just the more compliant one (think EU AI Act, data governance by design), but demonstrably the better and cheaper one. For a long time, digital sovereignty was an argument about values. Now it's a business case.

What does this mean for us? Three routes, no silver bullet

Intellectual honesty matters here: fine-tuning is not a universal tool. The success stories share two characteristics – high frequency and verifiable outcomes. A task that runs thousands of times a day and whose result can be checked objectively is a candidate. A rare decision that can only be debated is not.

That leaves three realistic paths:

Option 1: Rent frontier – deliberately and for a limited time. For prototypes and rare, creative tasks, the frontier model remains the right tool. The trick: treat every call as data collection. Log inputs, decisions, corrections – that becomes the training material for your own specialist later on. The risk: whoever stops here pays the prompt tax indefinitely.

Option 2: The managed route, Mistral Forge style. For organisations without their own ML teams, this is where we come full circle to my March article: fine-tuning in a controlled environment, weights in your own hands, EU AI Act compatible. The trade-off: less of a learning curve in-house, but a faster start.

Option 3: Build your own with open source and RL. The Bridgewater/Shopify route. Demanding – you need clean expert labels or a “digital twin” environment with scoring – but the barrier to entry has dropped dramatically: $500 of GPU time is no longer a conversation with the CFO, it's an expense claim. The real bottleneck isn't compute but expert time for data quality. Bridgewater's most important insight wasn't the training recipe; it was that cheap non-expert labels were worthless.

In practice it will be hybrid anyway: the frontier model orchestrates and your own specialist decides – depending on where Intelligence Ownership offers the most leverage at that moment. Or your own router LLM takes over the dispatching. Not either-or, but division of labour, because “It's the process, stupid!” (yes, I have to keep repeating it 😉)

But: it's also the data, stupid!

The common thread running through both studies is the same as in March – only sharper: the models have become interchangeable. Qwen today, a better base model tomorrow; the recipe transfers. What does not transfer, what nobody can buy or prompt, are your experts' labelled judgments, your taxonomies, your weightings and your reinforcement learning mechanisms. That's the institutional memory, that's Intelligence Ownership, and it is turning from an Achilles' heel into your greatest asset.

One possible consequence if this trend continues: in five years nobody will ask “Which model are you using?” but “Who owns your intelligence?”

Which brings us to the actual strategic question:

Which decisions in your company are worth renting – and for which are you currently paying an indefinite prompt tax on your own knowledge?

Live long and prosper 😉✋

▶Listen