The Thinking You Cannot Outsource

Most organisations still count brain capital as a single quantity, when the thing they are measuring has already split into capabilities that artificial intelligence is absorbing at wildly different speeds.

Every consulting deck about the future of work now leans on the phrase brain capital, and the framing stays useful right up until it starts implying that human intelligence is a single measurable stock rising and falling with headcount. An underwriter who can price a risk nobody has seen before, a designer who understands why a campaign will land in one market and quietly die in another, and an analyst who can assemble a clean cash flow model are all described as holding brain capital, while running entirely different cognitive machinery that artificial intelligence is reaching at entirely different speeds.

Five capabilities wearing a single label

Analytical and computational intelligence covers formal logic, structured synthesis and quantitative modelling, and it sits under the heaviest automation pressure, leaving people the narrower and far more consequential job of deciding which problems and which metrics deserve attention at all. Contextual intelligence, which governs informal networks, unwritten rules and organisational politics, together with the relational intelligence carrying empathy, negotiation and trust, sits almost entirely beyond what current systems reach, and whatever durable advantage you hold is concentrated there.

Models absorbed our intuition before they touched our reasoning

Daniel Kahneman divided human thought into System 1, which is fast, intuitive and driven by heuristic pattern matching, and System 2, which is slow, deliberate and effortful, and the early consensus in AI strategy assumed machines would claim the second while people comfortably retained the first. That expectation inverted, because a system predicting the next token performs something functionally much closer to rapid statistical association than to step-by-step verification, which explains why generative AI is so extraordinary at first drafts while remaining ungrounded in the world it describes. Human deliberate reasoning survived with a rewritten job description, shifting from performing the calculation towards auditing the assumptions underneath it, which is why Microsoft’s 2026 Work Trend Index found 86% of workers already treating model output as a starting point and half of them naming quality control of it as an increasingly important part of their role.

Everything you wrote down has already become training data

Michael Polanyi’s observation that we can know more than we can tell turns out to be the most practically useful lens available, because explicit knowledge — the procedures, playbooks and case libraries that entire professional service industries were built on codifying — is precisely the corpus frontier models were trained against. The tacit residue is what retains its value: the sense that a deal is drifting before the numbers say anything, the knowledge of which stakeholder needs telling first and which stated objection conceals the real one. Harvard Business School researchers found experienced practitioners identifying gaps in model output that novices were simply unable to see, which leaves you depending on a form of expertise that accumulates only by doing work you are initially bad at.

The apprenticeship layer is where the moat is quietly being spent

Tacit judgement accumulates when junior people grind through repetitive analytical work under supervision until they develop an instinct for when something is wrong, and that work is precisely what organisations are handing to models first, because it is the easiest thing to delegate and the saving appears immediately in this year’s numbers. Brynjolfsson, Chandar and Chen at the Stanford Digital Economy Lab found a 16% relative employment decline among workers aged twenty-two to twenty-five in the most AI-exposed occupations, concentrated in roles where the technology replaces the worker outright. The mechanism connecting that finding to your future capability is uncomfortably direct, since work that never passes through a person teaches that person nothing, and the senior judgement you will need in ten years is being quietly decided by hiring choices made this quarter.

Work that never passes through a person teaches that person nothing.

What survives is the part you were never able to write down

Thirty years of systematising and documenting professional judgement succeeded well enough to make all of it cheap, while the capabilities dismissed along the way as soft or unscalable or too dependent on particular individuals turn out to be carrying the value now. Article 14 of the EU AI Act makes one part of this explicit by requiring that high-risk systems be designed so a named natural person can interpret the output, decline to use it, and stop the system entirely, which converts moral stewardship into a documented control with an owner. So automate the pattern matching aggressively, and then spend the attention it frees on contextual judgement, political navigation, authentic trust and the willingness to be accountable when a confident model turns out to be wrong, because those are the only holdings in the portfolio currently appreciating.

All articles on this site are written by me. I use AI to assist with final formatting and editing before publication.