If the AI becomes the business, what is the business?

In plain terms. If your competitors can rent the same AI, the time it saves isn't an advantage, and it isn't money until someone banks it. Delete work before you automate it, count only the savings that reach the accounts, and put your spend into what a competitor can't rent: your data, your decisions, your standards and your customers' trust.

The most dangerous number in AI is a big efficiency percentage on a process that should not exist

Most AI business cases I have read this year open with a percentage. Forty per cent faster. Sixty per cent of the effort removed. The number is usually true, and it is usually attached to a process nobody has asked a prior question about, which is whether the process should exist at all.

There is a simple way to ask it. Take any workflow and, for each step, ask what happens five minutes before it and five minutes after. A surprising share of steps exist because of a system boundary, a control added after an incident nobody remembers, or a report that somebody once asked for and nobody reads. Asking the before-and-after question sorts the steps into three piles. Steps that should be deleted, which costs nothing and needs no software. Steps that should be automated, because they are routine, verifiable and worth doing. And steps that should be augmented, where a person keeps deciding and the software drafts, because the step carries judgement or was how people learned.

Most of what is sold as transformation is the middle pile applied to the whole list, including the first pile. Automating a step that should have been deleted is the most expensive way there is to keep it, because it now has a vendor contract, a maintenance cost and a percentage improvement on the slide that makes it look like progress. Delete first. An afternoon with the people who do the work will tell you which pile most of your steps are in. Where the time actually goes is the longer version of how I run that afternoon.

Time saved isn't money until someone banks it

A team adopts a tool and reports that it saves each person four hours a week. Twenty people, four hours, fifty weeks: four thousand hours, and at a loaded cost the slide says a few hundred thousand pounds. A year later the finance director asks where the money went, and it isn't there, because nobody took it.

Time saved becomes money in exactly two ways. Either the freed capacity is redeployed to work that earns, in the same quarter, by a named initiative that couldn't have happened otherwise. Or the cost is removed. Everything else is a claim. A claim isn't worthless, since it may be what makes the redeployment possible, but it isn't money, and a buyer's diligence team will discount it heavily.

So keep two columns. In the first, everything the tools are said to save: the claimed column. In the second, only the gains that meet the test: the banked column, with the initiative or the cost line named against each entry. The rule for moving a line from left to right is that it happened in the accounts, in the same quarter, and someone can point to it.

Two things about that table. Most firms run only the left-hand column, and the first time they run the right-hand one the ratio is a shock. And the right-hand column is the only one a buyer will pay for, because it is the only one that shows up in the numbers they are buying.

If everyone rents the same intelligence, what do you still hold

Strip the labour out of a people business and what remains had better be worth owning. It is entirely possible to disrupt yourself enthusiastically, remove the work that made your judgement valuable, and end up cheaper and worth less: a smaller version of yourself with better margins for about two years, until the same tools reach your competitors and price follows cost down.

So ask the scarcity question. If intelligence can be rented by the month, execution is nearly free, and everyone's strategy deck comes out of the same models, what do you still hold that a competitor with the same subscriptions does not?

Draw a line from rented to owned and place your assets on it. At the rented end sit prompts and model choices; anyone can have them tomorrow. In the middle sit workflows and integrations, which take effort to build and can be copied by anyone who watches you for six months. At the owned end sit five things.

1. Your data, in context. The raw data is usually less proprietary than firms believe. What is proprietary is the data joined up and interpreted the way your business interprets it.
2. Your record of decisions. Who decided what, on what evidence, with what reasoning. Most firms keep facts and throw decisions away, and an agent can't be taught a decision nobody wrote down.
3. Your tests of what good looks like. Real cases from your own work with the right answer marked by someone qualified to know. This is your professionals' judgement, written down in a form that can be checked.
4. Your evidence that you are in control. For every piece of software acting in your name, who is accountable, what it may do, and how well it does it. The piece on earned autonomy is about this.
5. Your distribution and your customers' trust. Which you had before AI and which AI doesn't change, except that it makes the other four more visible.

Everything at the owned end comes out of your own operation. That is why no vendor can sell it to you, and why a buyer will pay for it. Competitors will match your automation savings within a couple of years. What you own is much harder to copy. The line is how you tell one from the other, and it is worth drawing for your own firm before the next planning round.

Push it to the limit: one person and a thousand agents

Take this to its end and you get a company of one person and a thousand agents. It is worth doing the thought experiment properly, because it settles the question of what the business still owns.

Strip out all the labour and what remains is the list above: data, decisions, tests, evidence of control, trust. But the limit case exposes something the list hides. Those tests were built from the judgement of people who learned the work by doing it. With no people doing the entry-level work there is no curriculum, and without the curriculum nobody new acquires the judgement. The test set is frozen on the day the last practitioner left, and from that day the firm is running on stored judgement. Stored judgement depreciates. Markets move, regulations change, the cases that arrive stop looking like the cases in the tests, and there is nobody left who would notice.

This is why the piece on people and this one are the same question. What the business owns is only worth something if it can keep producing the judgement behind it, and it can only keep producing that judgement if the software doing the work is held to a standard a human still sets and can still check. Lose the people who set the standard and the tests go stale; lose the tests and nobody can check the software. The end state is a company that has decided, on purpose, which judgement it keeps producing in people, and built its agents around that decision rather than in place of it.

Quality you can prove is worth more than time you save

Most AI business cases stop at the saving. The more valuable thing usually comes after it.

Once you have written down what good looks like, tested your software against your own cases and kept a record of who decided what, you can prove the quality of your work in a way you couldn't before. Decisions are signed by a named person, runs are logged and exceptions are explained. That proof is worth something to a client.

So I'd expect the value to arrive in three steps. First, more capacity per person, which is the saving. Second, quality you can prove, which keeps clients and satisfies regulators. Third, a service your clients will pay more for, because the proof is built in: the compliance work you already sell, with the intelligence and the evidence attached. It is the same order in a start-up, where proof is how you win the first large customer, and in a corporate, where it's how an internal function becomes a product.

It works the same way in a product business. An AI feature that saves customers time gets copied within a year. One that can show a customer why it gave the answer it did, and who checked it, is the one they renew.

Very few AI programmes are designed for the third step.

Three people will ask what this is worth, and they mean different things

A chief executive, a board and an investor will each ask you to justify the AI plan, and they are asking different questions.

The chief executive is asking: are we behind, and is this real? The honest answer is that what you are buying is the ability to answer, on any day, four questions about every piece of AI acting for the firm: what is it doing, who is accountable for it, how well does it work, and what did it cost. Few firms can do that today, and being able to puts you ahead.

The board is asking: what happens when this goes wrong? The answer they need is that every piece of software acting in the company's name has a named person accountable for it, a written list of what it may and may not touch, and a level of autonomy it earned against criteria published before anyone measured it, which it loses automatically when it slips. That is a control environment an auditor can test. Earned autonomy is how it is built.

The investor is asking: does this survive diligence, and does it lift the multiple? Two things do. Portability: if the firm's records of what runs and how well it works are its own, testing a new model is a day's work rather than a project, and the firm isn't hostage to anyone's pricing. And measurement: the banked column, because a buyer discounts what was claimed and pays for what was banked.

For the first few years a firm creating value and a firm disrupting itself into a smaller one show the same numbers. Measurement is what tells them apart: the banked column, and a test set that shows the quality held.

What to do with this

If you run the business. Run the before-and-after question on one workflow this month and delete what it finds. Start the second column. List what you hold at the owned end, and count how much of this year's AI spend lands there.

If you sit on the board. Ask for the banked column, with names against the lines. If the answer is a percentage, you have been given the claimed column.

If you're building a start-up. You have no legacy processes to delete, which is an advantage for about eighteen months. Keep the decision record from the first customer. It's the hardest thing to rebuild later.

If you invest. In diligence, ask for the two columns and the owned list. A management team that has them has been running the business through this. A team that has a percentage has been running a pilot.

What this rests on

The delete-automate-augment sort and the claimed-versus-banked distinction are both old ideas from operations and from finance, applied here to AI; I make no claim to have invented either, only to the observation that they are rarely applied to AI spend. The five owned things are my list, from years of seeing what a buyer's diligence actually looks at. The observation that models are commoditising and that context and operating records are the durable value is now widely shared, and I have credited it on the About page. The limit case is mine.

The strongest objection

"Models are commoditising and context is the moat. Everyone says this now. It is obvious."

Partly true. The direction is now consensus. What isn't consensus, and what few firms do, is the accounting: the two columns, the line, and the honest count of how much of the plan lands at the owned end. Agreeing that context is the moat and then spending the year's budget on the rented end is the normal case, and the objection does nothing to change it.