Companies around the world are trying to use large language models as just another analytical tool, and running into the technology's fundamental limits. But these models weren't built for spreadsheets. They're designed for language. That means today's biggest AI opportunity doesn't live in BI warehouses. It lives in how you speak, write, decide, and communicate inside your company. There's far more untapped value in emails, meeting notes, presentations, and meeting transcripts than in most dashboards. And that's exactly where AI can start working immediately — no projects, no migrations, no waiting.
Most companies believe deploying AI first requires getting their data in order and building solid foundations. That sounds responsible, but it's a mistake. It doesn't mean investing in data infrastructure is pointless — it means it's absurd to wait another year or two for “real” AI deployment until big data projects get delivered.
Typical corporate IT, deep down, is still running on an old mental operating system that's been expanding since the nineties: big problems, big solutions, multi-year roadmaps, enormous budgets. It genuinely understands corporate transformation as a change in infrastructure, not a change in behavior. That worked in the cloud era, but it no longer fits the physics of today's AI so well.
Language models need context above all
Large language models don't need structured data. They need context. And enterprises are flooded with it. Every larger company today is sitting on an unmonetized asset: millions of words. Emails, presentations, documents, Slack threads, call transcripts — the living language of how the company is actually run.
Instead of waiting for your data to be perfect, feed AI the company's real language today, and add your strategy on top. Give AI raw, unedited meeting transcripts, presentations, old documents. AI will pull out the patterns and produce clear syntheses that can very quickly sharpen managerial judgment, refine not just the definition of strategy, but also ensure consistency in how it's executed.
That flips the logic of AI transformation somewhat upside down. For established companies, the bottom-up approach (clean data, models, insight, action) is now slower, more expensive, and less effective than the top-down approach (language, context, judgment). Companies are needlessly waiting for their data and systems to properly mature. They don't see that LLMs are already unlocking the organization's most valuable “soft” layer right now: decision-making, strategy, communication, and customer signals — everything expressed in human language.
Today's AI doesn't model the world
Why build language models a perfect data structure, when they're going to occasionally hallucinate over it anyway? LLMs aren't analytical machines. They're inference machines built on language. They don't model the world — they model how people describe it. A company's real readiness, then, doesn't lie in its data warehouse, but in its language: in how people talk inside the company, and therefore how they think. That language already exists — nobody's just tapping into it.
Some time ago, we started rigorously transcribing and analyzing every work conversation at our company. We got used to smart meeting minutes that don't just summarize the conversation, but also flag what we forgot and send people development feedback on how they communicate a given topic. Today, when I show up to a meeting without a recording running, it feels like heating a room with the window open. Words and ideas are money that companies are throwing out the window today.
Sometimes clients tell us openly: “We're not ready here for that level of transparency.” Top-down AI strips organizations of their favorite alibi — a lack of IT resources — and shows that the real barrier to change is cultural. When leadership can produce a first iteration of a strategy in 10 seconds, the bottleneck is no longer IT. It's people, their habits, and their identity.
That leads me to a maybe somewhat bold idea: the winning products of this generation won't be data platforms. They'll be cognitive infrastructure systems — systems that unify language, surface drift, align teams around shared context, compress ambiguity into clarity, scale up people's judgment, and ensure consistent interpretation of strategic decisions across the organization. This won't be infrastructure — it'll be a new layer of thinking. This will be the next McKinsey. Not as a consultancy, but as a permanent cognitive layer sitting on top of the enterprise.
What does top-down AI transformation actually look like?
Not as a project. Not as a roadmap. Not as a program.
It starts with five simple steps:
- Start with judgment, not data. AI should always produce the first draft of anything you need to discuss.
- Feed the model real language. Raw documents, meeting transcripts, presentations, and dusty old folders.
- Leadership has to lead by example. If leadership doesn't use AI every day, nobody will.
- Turn AI into a habit. Nobody starts from zero.
- Only expand where there's natural pull, and where you can see value right away.
This is the least ceremonial transformation in corporate history. And, at the same time, the most strategic.