In May 2023, I asked a question that felt like provocation for its own sake: will AI become the new McKinsey? ChatGPT was four months old. Most companies were using it to write emails. The idea that a language model might replace strategic advisory sounded like a stretch. It wasn’t a stretch. It was a signal. And by March 2026, the data arrived to prove it.
The trendslop study
Researchers at Esade Business School tested seven leading language models across seven core business tensions — the real strategic choices every organisation faces: differentiate or compete on cost, augment or automate, play the long game or the short one. Choices where the right answer depends entirely on who you are, what market you’re in, and what you can actually execute.
Every model gave the same answer. Not similar. The same. Across 15,000 trials, every LLM recommended the fashionable option — differentiate, augment, invest in innovation — regardless of whether the company was a startup or a conglomerate, cash-rich or drowning in debt.
Better prompting improved results by less than 2%. Feeding the models detailed organisational context — revenue, competitive position, real constraints — shifted the bias by only 11%. The researchers had a name for what came out the other end: trendslop. Strategy-shaped content that sounds authoritative and says nothing, because it isn’t reasoning about your situation. It’s reflecting the average of everything ever written about strategy.
Why this matters more than it looks like it does
The obvious takeaway is “AI is bad at strategy.” The real one is bigger: the value of intelligence is collapsing toward zero. What’s left standing is context. Not the model. Not the data. Not the algorithm. Context — the accumulated operational knowledge, the institutional memory, the years of paying attention to your specific situation that no model was trained on because it doesn’t exist anywhere except inside your organisation.
That’s why “which AI model should we use?” is the wrong question. It’s the McKinsey question — it assumes the value lives in the tool. The right question is: what context do we have that would make any model useful? Ask the best model in the world what your company should do about AI, and you’ll get a competent, generic answer: cross-functional task force, pilot programme, governance framework. True. Useless. It doesn’t know your data infrastructure is a swamp, that two of your executives haven’t spoken in six months, or that your engineering team quietly built a shadow AI system last quarter.
That’s context. No model has it — unless you build it.
Context rot
I’ve started calling the alternative context rot: the slow decay of operational knowledge inside an organisation. It happens when people leave and take what they knew with them. When processes live in someone’s head instead of a system. When the meetings and conversations that actually generate understanding get replaced by dashboards showing numbers without meaning. Context rot is worse than technical debt, because technical debt is visible to engineers. Context rot is invisible until the moment you need it — and discover it’s already gone.
Strategy is human
The 11% figure from the Esade study is the most damning number in it. It means even your best effort to feed a model context still leaves the output 89% generic. The remaining gap — the specific, courageous, contrarian call that actually moves an organisation — is human. It’s institutional memory. It’s knowing why the last three initiatives failed. It’s the judgement built from paying attention, year after year, to signals everyone else filtered out.
AI can accelerate every part of that process. It cannot replace any of it.
