· Charlie Holland · Leadership · 11 min read Every enterprise technology programme for forty years has been the same project: getting what people know out of their heads and into a system. AI is just the first tool sharp enough to reach the knowledge we always called too skilled to automate. Here's what the rising water does when it finally reaches the high ground, and why I help build the thing that worries me.
In the last post I gave you the map I use to deliver AI. Where it goes, and how you climb from assisted drafts to autonomous operation without landing in the 40% of projects that get cancelled. Useful, practical, the sort of thing a client pays for.
I left out the part that keeps me up at night. Not how this work succeeds. What happens when it does.
So let me start with a picture instead of a framework.
For forty years the water’s been coming in. Every wave of enterprise technology raised it a little. The first ones drowned the low ground: the typing pools, the filing clerks, the ledger rooms, the data-entry floors nobody ever called skilled. We watched it happen and called it progress, because mostly it was. The people on the high ground watched too, and felt safe, because their skill was their elevation. The water was for other people. It was never going to reach them. The water’s reaching them. And the uncomfortable thing I’ve got to tell you, as one of the people who spent a career building the tide, is that it was always going to.
The job was never the technology #
Here’s the part that never makes the brochure, because no one’s worked out how to point a smoke machine at a spreadsheet. Every enterprise technology programme of the last forty years has been the same project wearing a different suit.
ERP was never really about software. It was about taking how a business actually runs, the ordering and the invoicing and the thousand local arrangements that lived in people’s heads, and forcing it into a system that would execute the same way every time. CRM did it to how you sell. The document management platforms, including the one I once wrote a rather long book about, did it to how a business handles its own paperwork. Data warehousing did it to what the numbers mean. Expert systems in the eighties tried to do it to professional judgment directly, and mostly broke their teeth on it. Robotic process automation did it to the repetitive clicking the earlier waves left behind.
Different decade, different brochure, identical work underneath. Take what people know. Write it down in a form a machine can run. Hand it to the machine.
We’ve always been in the business of getting it out of their heads. The tools changed every ten years. The job never changed at all.
And none of this is technology being the villain. I want to be clear about that, because I build the stuff for a living and I’m not going to romanticise the filing cabinet. The point isn’t that automation is wicked. It’s simpler than that, and harder to argue with. A commercial organisation exists to create value, and the most dependable way to create value is to take cost out. McKinsey built a global empire singing that one song in a hundred local accents. Knowledge trapped in one person’s head is a cost and a risk. Get it into a system and you’ve turned a salary into an asset you own. Nobody in that chain is twirling a moustache. They’re doing precisely what a business is for.
Why it keeps failing the same way #
If that’s the job, then the reason these programmes fail is the same every time, and it’s never the bit anyone blames. Michael Polanyi put his finger on it in 1966 with one sentence: we can know more than we can tell. The master joiner can’t fully describe the cut. The senior support agent can’t fully enumerate why this customer gets the refund and that one doesn’t. The knowledge that actually runs the place is tacit. It lives in the hands and the judgment of the people doing the work, and a good chunk of it has never been written down anywhere. Research keeps landing that figure near 42%, expertise held by exactly one person and recorded nowhere.
That’s the real backlog of every one of these programmes. Not the software. The 42%. Forty-two. The answer to life, the universe and everything, and also the share of your business that lives in one person’s head. Douglas Adams would’ve enjoyed that.
And it’s miserable work to clear. It’s slow. It’s political. The person whose head it lives in understands perfectly well that the moment it’s written down, they’re a little less essential than they were the day before. So they’re helpful in the meeting and vague in the detail, and who could blame them. Brynjolfsson’s Productivity J-Curve is the academic version of the same story: bolt the new technology onto the old workflow and you get almost nothing, because the value was never in the technology. It was in the painful organisational work of writing down what you do, and most organisations would rather buy a tool than do that work. So they buy the tool, skip the codification, blame the tool when it underdelivers, and wait ten years for the next one with a better logo.
I’ve watched this across enough waves to find the new one familiar. The current wave calls itself agentic AI, and Gartner already expects 40% of those projects to be cancelled by 2027. Same failure. New acronym.
Today’s map, same old territory #
The way I deliver it now has two halves, and I covered them last time, so I’ll be quick.
There are three places AI actually goes in a business. Talk to your customers, build your things, make sense of your data. And there’s one honest path to get there: start with the human in charge and the AI assisting, earn your way to the AI executing under review, and reach real autonomy only when the measurement says you’ve earned it. Earn autonomy, don’t grant it.
It’s a good map. I stand by it. But look hard at what the first rung actually is. In the Assist stage the AI drafts, the human corrects, and every correction is captured. That’s not a side effect. That’s the point. The system isn’t there to help the expert. The expert is there to train the system. Every edit they make is another piece of the 42% finally coming out of their head and into a form the machine can keep.
We’ve built, at last, a beautifully efficient machine for doing the oldest job in enterprise IT. It plays the dutiful apprentice, all “yes, Master.” But this is the Sith version, where the apprentice exists to learn everything the master knows and then replace him. Always two there are. Never, in the end, both.
You’re probably doing it right now. Ever wonder why Claude Code’s two-hundred-dollar-a-month plan lets you burn through many times that in metered terms without anyone trying to stop you? The top model runs at five dollars per million tokens going in and twenty-five coming out on the open tariff, and yet heavy subscribers chew through thousands of dollars of equivalent usage a month for a flat two hundred. Some of that’s a land-grab, the familiar bet that the below-cost developer of today is the locked-in enterprise of tomorrow. But some of it’s the point of this essay. The tokens aren’t the product. You are. Every correction you make to the model’s code, every suggestion you reject, every plan you rewrite, is you doing the unpaid and unglamorous work of getting the last of it out of your own head and into theirs. The best-instrumented codification project in history, and the experts are queueing up to pay for a seat.
The one thing that is genuinely different #
So far I’ve told you nothing’s new. Now the part that is, because it’s the whole reason I’m writing this instead of billing.
Every previous wave codified work that nobody defended as a skill. The clerk, the ledger, the data-entry floor, the repetitive click. We drained the low ground, the people on the high ground stayed dry, and the high ground was made of judgment. The reason the analyst and the developer and the lawyer and the architect got paid was precisely that their knowledge was tacit. It couldn’t be written down, so it couldn’t be copied, so it had to be rented from them by the hour. The 42% wasn’t a gap. It was the moat. It was the thing that made a professional a professional.
This is the first wave with a tool sharp enough to reach across the moat.
Not perfectly, not everywhere, not yet. But the direction isn’t in doubt. The water that drowned the typing pool is now lapping at the kind of work we built whole identities and mortgages on, the work we told our children to train for because it would keep them safe. And the question the entire forty-year history was always going to arrive at, the one nobody on the project wants on the slide, is finally on the table.
What’s the person for, once you’ve got it all out of their head?
The two endings we were promised #
Popular culture sold us two answers to that, and we should be honest that both are already in the room.
There’s The Matrix, where humanity is kept alive in vats as a power source, pacified by a comfortable simulation while the machine quietly harvests what it needs. The detail that always struck me is the interface. A jack, in the back of the skull. The literal picture of getting it out of their heads.
And there’s Wall-E, the clean bright one, where the machines do all the work and humanity reclines in floating chairs, every need met, soft and entertained and no longer able to stand up. No war, no exploitation, just a comfortable and total obsolescence.
Here’s the part the films got wrong, and it’s the part that should worry you. These were never two futures we choose between. They’re one future, and it’s divided. For most people, the codified-out and economically surplus majority, it’s the first film. Your skill is now worth what the copy of it is worth, which is nothing, and you carry on not because you’re fooled but because, as with so much of this, the alternative was quietly removed. For the few who own the models, the data, and the pipelines that did the copying, it’s the second film. The chair. The captured surplus of everyone else’s knowledge, flowing to whoever owns the machine that now holds it.
Because that’s the mechanism, stated plainly. Automation doesn’t destroy the value of a skill. It moves it. It always has. From the person who holds the skill to the person who owns the thing that copied it. Every previous wave did this at the bottom of the income scale, and we called it progress and found those people other jobs. This wave does it at the top, to a class large enough and expensive enough that I genuinely don’t know where the other jobs come from.
I don’t know how a society stays stable when most of its people have had their economic value extracted and banked upward into very few hands. I’m not predicting the barricades. I’m noticing, the way you notice a number that doesn’t add up, that this one doesn’t, and that nobody selling the tools seems keen to do the sum.
The bit I have to own #
This is the bit where a sensible essay reassures you. New jobs will appear, human ingenuity always wins, it all works out in the end. I’m not going to, because I don’t believe it, and the industry hasn’t given me a reason to.
I make my living doing this. The map in the last post is a real map, I’ll deliver it well, and the better I do my job the more efficiently I get the 42% out of people’s heads and into a system that no longer needs them. I’m very good at building the thing I’ve just described. The tide tables in this essay aren’t something I observed from the shore. I drew them. I’m the surveyor the estate hires to map the common ground before it gets fenced.
And the moat, the tacit judgment that can’t be copied, is the one I’ve been renting out by the hour for my whole career. The architect’s judgment. Mine. There’s no high ground in this piece from which I get to point at the water.
Forty years ago the job was getting it out of their heads. It still is. The only thing that’s changed is whose heads, and the answer, for the first time, includes the people who were sure they were safe.
The water’s still coming in. It always was. The difference is that now I can see my own door from here.