The AI Gap in Your Own Organisation

Two debates are running side by side right now, and they barely touch. On the business pages, the topic is the AI bubble: Ray Dalio has been warning for months that the AI boom has the hallmarks of a classic, debt-financed bubble. Last week Fundstrat's Tom Lee pushed back, arguing that we are closer to 1997 than to 1999 and that he does not expect the peak until around 2033. At roughly the same time, Andrej Karpathy, formerly of OpenAI and Tesla, shared a picture that points in a completely different direction: the problem is not that AI promises too much, but that hardly anyone sees what it can already do.

The funnel

On 11 October, Karpathy re-shared and sharpened a post from April. His estimate, which he himself calls napkin math: around six billion people, about three quarters of the world's population, have barely come into contact with language models. One to two billion use free chat products occasionally. Roughly twenty million work professionally with current frontier models, mostly in coding and mathematics. And about five thousand people have internal access to systems that are not yet public.

His point: anyone who bases their view of AI on a try-out of a free version from a year ago is talking about a different tool than someone who works with agents every day. The numbers are not measured, and I would not quote them as if they were. As a picture they are still useful, because they describe something that repeats itself in miniature inside organisations.

The same funnel, in miniature

In a biotech or medtech company with a hundred employees, it often looks like this: two or three people use current models intensively, usually in bioinformatics or IT. A larger group has licences but uses them for emails and summaries. The rest tried it once and concluded it is no good for their own work.

The problem is not the uneven usage. It is that decisions about budget, policy and priorities are often made by people who sit higher up the funnel than those whose work would change the most. I have written elsewhere about why AI adoption in life sciences is a change management task rather than a technology rollout. The funnel explains why the same licence in the same organisation produces such different results.

Why the bubble debate does not help here

Whether the valuations of chipmakers and hyperscalers are overblown is a question for investors. For your organisation, the relevant question is a different one: what can the tool you can license today do for a specific process, and are you using it? Even if the bubble bursts, the existing models do not disappear. The gap between what is possible and what is used remains your problem, not the capital market's.

What I recommend

Measure usage, not licences. A licence count says nothing about where your people sit in the funnel. In our case study on Copilot adoption in a biotech SME, that was exactly the metric: active users, not licences handed out.

Let leaders do the work themselves. Before anyone decides on AI budgets, they should have completed a real task from their own area with a current tool, not watched a demo.

Make your power users visible. The two or three people who are already far along need time and an official frame to show others what they do. Otherwise they keep working in the shadows and the organisation learns nothing from them.

Remember Article 4 of the AI Act. Since February 2025 it has required providers and deployers to take measures on AI literacy. With the Digital Omnibus, in force since 27 July 2026, this has become a duty to support the development of AI literacy without having to guarantee a specific level. That is a relief, but not an all-clear: an organisation where most people know AI only from hearsay will struggle to show that it supported anything.

If you want to know where your organisation sits in the funnel and what it would take to close the gap without putting your compliance at risk, let's talk. No theory for theory's sake.