The Ferrari with no roads
You have bought, or are about to buy, a very fast car. This page is about the roads.
What was bought
The models are extraordinary and they are getting cheaper. They draft, summarize, answer, search, write code, and talk in everyday language about almost anything. Every vendor you deal with now has one in the product, and your people have them open in a second window whether or not anyone approved it. That part is done; the car is in the driveway.
What has not been bought, because it is not for sale, is the thing that turns a fast car into a way of getting somewhere: a network of roads, with signs, that connects where you are to where you need to be. In a company that network is the set of requests and promises that run the place. Who asked whom for what, whether they said yes, what yes meant, by when, and whether it came back. The org chart is not that network. The project plan is not that network. The network is in the email and the chat and the hallway, and in most companies nobody can see it.
Why the car does not help by itself
A language model is a very fluent new employee who walked into every meeting at once with no idea what anyone needs to walk out with. Where the organization already speaks in commitments, that employee is a gift: hand it a clear request, with what done looks like and a date, and it will work inside that request and report back. Where the organization does not, the same employee produces more drafts, more summaries, more plausible replies, none of them attached to a person, a result or a date. More talk. Talk is cheap and now it is nearly free, and a company that was already drowning in it will not be saved by a faster tap.
The phrase everyone uses for this is "human in the loop." Ask the next person who says it which loop, and what the human does in it. The loop is the one on this page: someone asks, someone agrees, the work is done, the person who asked says whether it was what they needed. The human makes the request, accepts or declines what comes back, declares it done and assesses it. The agent works inside the doing. An agent cannot promise, because a promise needs someone who can be held to it. Draw that line and you know exactly where AI belongs in your company and where it does not, which is a plan, and which is what most AI deployments are missing.
What it is already costing
Ronald Coase explained in 1937 that companies exist because coordinating people by direction is cheaper than coordinating them by contract. The corollary is that a large share of what it costs to run a company is the cost of coordination itself: the status meetings, the chasing, the translation between teams, the finding out who promised what, and the work that was done twice or not at all because a request went into the air and a yes meant maybe. A good deal of middle management is this cost with a title. It is on your books already, under salaries and under the quarter that slipped, just not under the name coordination.
Michael Jensen, whose work on incentives shaped a generation of executive pay, spent his later years arguing that what makes organizations work is simpler than incentives: whether people honor their word, meaning keep it or say early that they cannot and clean up after. He called that integrity and insisted it is a factor of production, with effects on performance that do not depend on whether anyone approves of it. That claim is argued more than measured, and it is presented here as a position rather than a proven result. But the mechanism is not mysterious: a company whose promises are visible coordinates more cheaply than one whose promises are not.
What we would do
Not install anything, to begin with. Agent in the Loop is an engagement: interviews with the people who do the work, one at a time, about what they ask for and are asked for and how it goes; a map of each person's loops, validated with them; the maps merged into a picture of how the work flows in practice; and a review where the people in each loop say what should change and someone with authority declares it. The method is thirty years old and the maps used to be drawn by hand. What is new is that an agent now sits beside the interviewer, listens, and drafts the maps, so the engagement is a fraction of what it once cost.
At the end you have your own evidence of where coordination is costing you, in your people's own words, and a declared way of working that can be measured against. If you then want something to stay and keep the record, the Ledger is the next rung; it is described on the ladder, and it is built so that the people whose promises are in it can accept it as fair, because a record nobody trusts is one more place for the politics to happen.
What we would not do
Score people. Rank teams. Build a surveillance archive of everyone's conversations. Replace anyone. Promise that software will make your people keep their word; it will not, and anyone who says otherwise is selling talk. What a harness can do is make requests and promises visible to the people who made them, ask a better question at the moment it matters, and keep the book. The rest is yours.
Or start where everyone starts: see one of your own threads. It takes a minute and keeps nothing.
Numbers on this page are stated with a source and a date or not at all. The figures from the original Ferrari presentation will be added here once each has both.