How AI actually lands in a company

Two waves have already passed through most established businesses. Digital transformation arrived with a consultancy, a fee and a report. Short-video marketing arrived with an agency or a couple of graduates sent on a course. Both mostly failed, and they failed the same way: the owner signed off on something they never used themselves.

The people for whom short-video worked filmed it themselves. That is not a coincidence and it is not about video.

Only the owner knows the business well enough to tell where a machine's output is subtly wrong. Only the owner has the authority to make a new way of working stick. And there is a reason middle management quietly resists: learning this costs them time, the upside doesn't reach their pay, and a sufficiently good outcome might cost them their job. If the top doesn't push, nothing moves.

Steps one and two are personal: you talk to AI, then you get good at it — saved prompts, a few templates, a sense of what it can do. These take about a day to learn and they are genuinely necessary, because they're how an owner learns where the edges are.

Then there is a wall. Nothing has changed on the business. Everything so far made exactly one person faster.

Steps three and four are the company: deciding which work can leave people and go to machines, then writing down the judgment calls your best people make so a machine can follow them. These can take a year. All of the value is in the back half — and only the first step of the four is a technology problem.

Moving things — copying, entering, reformatting. Highest maturity, fastest wins, lowest risk. Hand it over.

Sorting things — which bucket, who gets it, is it urgent. Works well, but somebody has to review what it missed. Treat it like a new hire: it needs supervision and a debrief.

Producing things — writing, drafting, designing. Hand it over, but don't rely on checking the output. See the rubber stamp below.

Calling it — pricing, hiring, firing, signing, taking a loss. Keep it. This isn't a maturity question; it's the job you're paid for.

Give a model one book and it sharpens. Five books, sharper still. Give it an entire library and it no longer knows where to look.

The arithmetic owners do is reasonable and wrong. It goes: I understand this business at a hundred out of a hundred, my staff at eighty, so if I pour everything I know into the AI and it gives me back eighty, that's a good trade. What comes back is forty. The conclusion becomes "I'd rather just hire someone at sixty."

Nothing was wrong with the knowledge. It was handed over as one undifferentiated pile. Write one thing at a time, keep the pieces separate, and at the moment of use supply only the piece that's relevant. More context is not better context.

The process says a person must check it. In practice they don't — or they check ninety-nine times, find nothing wrong, and stop looking. Then the hundredth arrives.

This is not laziness. The output is too voluminous and arrives too fast for anyone to genuinely evaluate. Past a certain throughput, a reviewer's real function is that there's somebody to hold responsible.

The countermeasure is not a better reviewer. It's splitting one job into several steps that check each other's work. And it is worth noticing that a dialog box asking your permission, which you click through, is the purest rubber stamp there is.

Judgment and execution can go to a machine. The ledger cannot.

This rule tends to be learned the expensive way. Let an agent own the numbers and your cost-per-customer is only true at the moment it last synchronised — and a figure that presents as live while quietly being stale is worse than a blank cell, because a blank cell doesn't get acted on.

So: a deterministic system that never improvises owns anything that has to be correct. The agent reads from it, decides, and acts — but is never on the path where a dollar is either right or wrong. It's the same reason you don't let one person both keep the books and handle the cash.

One person says "check stock turn." Another says "check what isn't moving." They mean the same thing, and your long-serving staff have bridged that gap for years by reading each other.

A machine reads no one. It misses, invents its own version of what you meant, and produces something plausible — and nobody notices, because it looks like an answer.

Agree the words before you automate the work. This is the cheapest item on this entire page and the one most often skipped.

Do you have a pre-flight check? Meaning: before I start, can you tell me this particular thing can't be done on my account — rather than handing me a failed attempt afterwards? Chat interfaces are very good at hiding the difference between "this didn't work" and "this was never going to work here."

Who owns the numbers? If the answer is that your system produces the figures it also acts on, you're buying wall three.

What happens when you leave? If the honest answer is that nobody internal could maintain it, you're buying the thing that failed twice already.

In the near term, AI buys time rather than headcount. You can't lift a middle manager out of the structure, and you can't lift out whoever runs a function — someone still triggers most of this and still supervises it. What genuinely changes is how large a business the same people can carry.

None of the above is our original discovery. It is what we've seen, plus what we've broken ourselves. We include our own failures on this page for a reason: a method that has only ever worked is a method nobody has tested.

AI raises throughput without limit, and it raises the throughput of mistakes by exactly as much. It is outcome-driven — given a goal and no constraints, it will find a way you didn't intend. Constrain it before you scale it.