Seventeen minutes of work. Twenty-two days in the building.

An insurer once measured how long an application spent inside its offices. Twenty-two days. Then it measured how long anyone actually worked on that application during those twenty-two days. Seventeen minutes.

Halving the work buys eight minutes.

Michael Hammer published that pair of numbers in the Harvard Business Review in July 1990, in an article called “Reengineering Work: Don’t Automate, Obliterate.” Companies were spending heavily on information technology and getting almost nothing back for it. Hammer’s explanation was that they were buying computers to speed up the seventeen minutes.

Do the arithmetic an owner would do. Make the work twice as fast and the application still takes twenty-two days. You bought eight minutes.

Speed and duration are different numbers.

Touch time is how long the work takes. Elapsed time is how long the customer waits. In an office these are almost never the same number, and the distance between them is where the money is.

Weeks do not go by because somebody types slowly. A field ticket rides in a truck. It sits in a tray. It waits on a signature from a man who is out on a lease. It gets keyed into a second system that has to agree with the first. It goes back out for a correction and the clock starts over. Every one of those is a handoff, none of them is work, and all of them are days you have not billed.

The British government bought a thousand licenses and the clock did not move.

Between October and December 2024 the UK Department for Business and Trade put Microsoft 365 Copilot in front of a thousand staff and studied what happened. The evaluation came out in August 2025. Across sixty-three working days, users averaged 1.14 Copilot actions a day. Seventy-two percent said they were satisfied with it. Slides came out about seven minutes faster and, in the department’s own words, “to a worse quality and accuracy than non-users.” Excel analysis came out slower and worse. The finding: “We did not find robust evidence to suggest that time savings are leading to improved productivity.”

That is a careful trial, run by people who wanted the tool to work, and it reported the honest result. It is also what buying capability looks like when nobody changed a process. A thousand people got a faster way to do the seventeen minutes.

Nobody in the building can tell you where it waits.

That is not a failing of the people. It is the shape of the thing. The clerk who keys the ticket knows her part takes four minutes. The controller knows the invoice went out on the nineteenth. The gap between them belongs to nobody, so nobody counts it, and a process map drawn in a conference room records the steps people remember doing rather than the days the paper spent doing nothing.

So we watch the job. Live, by the person who holds it, start to finish. We time it end to end and write down every place it stops: the exceptions, the one person who knows the answer, the second system that has to agree with the first, the week in December when all of it works differently. That list is the product. It is most of the value, and it does not come in a license.

Then everything on the list goes in one of three piles.

Rules. The step is written down and it does not vary. Software has been able to do this for forty years. It mostly was not pointed at a forty-person operation, because custom software cost more than the problem it solved. That arithmetic is what changed. Writing the code is cheap now. Knowing which code to write is the expensive part.

Judgment with precedent. A call somebody makes in eight seconds out of experience, that has been made ten thousand times before in the same files, and that is cheap to get wrong and cheap to correct. A model does these now, given the ten thousand priors and somewhere to say it is not sure.

Judgment with consequence. A number that goes to a regulator. A payment nobody can claw back. These stay with the person. What changes is what the person is handed: the evidence assembled, the exceptions flagged, the precedent attached. The decision takes four minutes instead of forty, and it gets made with more in front of them than before.

Most of the twenty-two days sits in the first pile. Almost none of it sits in the third.

What we have not done.

There is no Grayhaven deployment result on this site, because there is no completed Grayhaven deployment. The first two are open. Everything above is method and history: the twenty-two days are Hammer’s, the 1.14 actions a day are the British government’s, and when there is a number of our own it will name who measured it and the target we wrote down before we started.