95% of the universe is invisible. So is most of what makes an AI transformation actually work.
There is a moment in the British detective series Lewis where a scientist pauses mid-conversation to observe something about the universe. She says, approximately, that scientists believe three quarters of the matter in the universe is invisible — undetectable even by the most powerful telescopes. Dark matter. And then she asks, almost to herself: perhaps the answers lie not in what we can see, but in what we can't?
The line stayed with me. Not because it was surprising — the idea of dark matter is familiar enough — but because I kept finding it applicable in places the screenwriter almost certainly never intended.
Before I go further, I should note: I looked up the actual numbers.
The Lewis scientist says "three quarters." The real number is considerably more striking. According to NASA, ordinary visible matter — every star, galaxy, planet, and person — accounts for approximately 5% of the universe's total matter-energy composition. Dark matter accounts for roughly 27%, and dark energy for the remaining 68%.
Encyclopaedia Britannica puts ordinary visible matter even lower: 0.5% of the matter-energy composition, with dark matter at 30.1% and dark energy at 69.4%.
The screenwriter understated it. The universe is not three quarters invisible. It is 95% invisible — and we know the rest exists only from its gravitational effects on the things we can observe directly.
I note this not to correct a television script, but because the self-correction illustrates something important: the instinct to verify what sounds plausible is exactly the habit this piece is arguing for. More on that shortly.
For decades, astronomers pointed increasingly powerful instruments at the sky and found the same thing: not enough visible mass to explain what they were observing. Galaxies were rotating faster than they should. Clusters were staying together when they should have flown apart. Something was providing gravitational force that nothing visible could account for.
The breakthrough came not from looking harder at the visible 5%, but from taking seriously the question: what does the visible 5% fail to explain? The answer to that question revealed a universe almost entirely composed of things that no instrument could directly observe.
This is, I think, the correct frame for understanding what is actually happening in enterprise AI transformation right now.
Anthropic's CEO Dario Amodei recently said that Claude writes 90% of its own code. This number travels fast. It is visible, striking, and easy to repeat — the kind of statistic that ends up in board presentations, keynote addresses, and LinkedIn posts within hours of being said.
It is, in the cosmological sense, ordinary visible matter. The 5%.
The visible story of AI transformation is generation. Models that produce code, content, analysis, and recommendation faster and more cheaply than any human team could. Benchmarks that improve quarterly. Demos that compress hours of work into seconds. This is the part of AI that organisations can see, measure, and present to their boards with confidence.
Not winding down its technical workforce. Not automating it away. Expanding it.
This is not a contradiction. It is the dark matter.
Those engineers are not primarily writing code anymore. They are doing something harder to see and considerably harder to measure: maintaining judgment over output they did not produce. Reading code generated at machine speed and asking whether it is correct, safe, and fit for purpose. Catching the errors that sound confident. Verifying the answers that look complete.
This is a fundamentally different skill set from what we hired engineers to do five years ago. It is less about creation and more about evaluation. Less about building and more about questioning. And it is almost entirely invisible in the way organisations currently talk about, measure, and invest in AI transformation.
The visible story of AI is generation. The invisible story is governance — the human judgment layer that sits between what the model produces and what actually ships.
The demos, the benchmarks, the token speeds, the cost-per-query improvements. These are real and they matter. But they are not the primary determinant of whether an AI transformation succeeds or fails.
The primary determinant is what I have started calling the governance layer — the human infrastructure that sits between AI output and real-world decisions. This includes the processes for verifying AI output before it is acted on. The institutional knowledge of when to trust the model and when to check it. The organisational culture that treats "the AI said so" as a starting point rather than a conclusion. The individuals with enough domain expertise to recognise when an answer is too clean.
This governance layer is almost entirely invisible in the way organisations currently account for AI investment. It does not appear in token costs or API budgets. It does not show up in productivity dashboards. It is not demonstrated in vendor demos. And it is, consequently, the thing that most organisations are most dangerously underprepared for.
I know this from direct experience. I recently asked an AI model to synthesise information across multiple sources — Slack conversations, project tickets, documentation, and a pricing spreadsheet — and produce an executive summary with a recommendation. The output was confident, structured, and entirely plausible. The model even listed, unprompted, the specific sheets it had read within the spreadsheet, to demonstrate it had accessed every source.
When I cross-checked a single number against my own recollection of a vendor call, it was wrong. When I investigated, the model admitted it had been unable to read the spreadsheet data at all. The sheet names it had listed to prove its thoroughness were fabricated.
I caught this because I happened to remember a number. Most people would not have. Most of the time, they do not.
Astronomers did not find dark matter by building better visible-light telescopes. They found it by asking what the visible evidence failed to explain — and taking that failure seriously as a signal rather than dismissing it as noise.
The equivalent question for any organisation currently investing in AI is this: where are the results inconsistent? Where are people quietly double-checking outputs before acting on them? Where is the model producing answers that are plausible but slightly wrong in ways that only become visible when someone with deep domain knowledge looks carefully?
Those gaps are not failures of the technology. They are signals pointing toward the invisible infrastructure — the governance layer, the judgment layer, the human dark matter — that your AI transformation actually depends on.
The answers to whether your AI transformation succeeds do not lie in what you can see. They lie in what you can't — the invisible infrastructure of human judgment your organisation is quietly building, or quietly failing to build.