Talk is cheap.
Speech is powerful.
Teaching computers to talk in everyday language was a good first step. Making them useful to an organization means exercising the value of speech, and talk and speech are not the same thing.
Paste a conversation thread. The Diagnostic reads it against the loop below and shows what was asked, what was promised and what was left unsaid. Nothing is kept. How it works.
The difference
Talk describes. It reports, suggests, summarizes, explains, and when it is over nothing has to happen. Speech, in the sense that matters to an organization, does something. A request creates an expectation. A promise creates an obligation. "Done" changes what is true. Most of what passes between people at work is talk, and that is fine, because talk is how people think together. But every piece of work that gets done started with speech: somebody asked, somebody said yes, and both of them could later be held to it.
A computer that can talk in everyday language is a convenience, and a large one. An organization that can speak, where requests are clear and promises are real, is one that keeps its word. Adding the first does not produce the second.
Why now
Companies are adding AI to their conversations faster than they are deciding what those conversations are for. Each new tool is a very fluent new talker with no idea what anyone needs to walk out of the room with. Where the organization already speaks in commitments, that talker is useful: it can be given a clear request and asked to report back. Where it does not, the tool multiplies cheap talk, more drafts, more summaries, more plausible replies, none of them attached to a person, a result or a date. The gap is not in the model. It is in the organization's ability to make and keep agreements, with people and with machines, and most AI deployments have no plan for it.
What speech does
The idea that saying something can be doing something is old and well worked out. Austin and Searle described it in the 1950s and 60s; Fernando Flores and Terry Winograd brought it into the design of work in the 1980s. Every piece of work travels the same loop.
Someone asks for something. Before that, they have to work out what they need, from whom, and by when; that is the preparation, and it ends in a request.
Someone agrees, declines, negotiates, or asks what exactly is wanted. That is the negotiation, and it ends in an agreement, which is the moment a request becomes a promise.
The work is done. Often the person doing it has to ask other people for things along the way, which opens smaller loops inside this one. It ends when they say it is finished.
The person who asked says whether it was what they needed. That is the assessment, and it is the move most often skipped, so that work ends without anyone saying it has.
Four moves, each of them a sentence somebody said or wrote, and almost every breakdown in a workplace is a loop that got stuck at one of them: a request with nothing to hold onto, a yes that was really a maybe, a job finished that nobody declared finished, a result nobody assessed.
A request a stranger could check names who is doing it, what finished looks like, and by when. A yes that means something answers all three. "I'll try to get to it" answers none of them, and everyone in the conversation knows it. This is a way of speaking, not a personality, and it can be learned.
A company is a network of promises
The org chart shows who reports to whom. The work runs on who asked whom for what, and whether it came back. Seen that way a company is a network of agreements, negotiations and promises, some made and some not, and most of what it costs to run one is the cost of promises that were unclear, quietly broken, or fought over. Coase explained in 1937 that firms exist where coordinating by direction is cheaper than coordinating by contract; a company whose promises are visible coordinates more cheaply than one whose promises are not, and needs fewer people whose job is chasing down who said what.
Michael Jensen, who spent a career on how to make managers behave through incentives, ended up arguing something simpler: that integrity, meaning honoring your word, keeping it or saying early that you cannot and cleaning up after, is a factor of production rather than a virtue. Its effect on performance does not depend on whether anyone approves of it. Robert Solomon and Flores made the companion argument about trust: it is not a feeling but something people do, built by making and keeping commitments and by honest conversation about the ones that were not kept.
Where agents fit
An AI agent can do a great deal inside that loop. It can ask what exactly is wanted, draft, look things up, do the work, report progress and say when it is finished. It cannot make a promise, because a promise needs someone who can be held to it, who can come back and renegotiate when it will not hold, and who bears the consequences when it does not. When a person hands work to an agent the commitment stays with the person.
The phrase "human in the loop" is everywhere, and almost nobody using it can say which loop or what the human does in it. The loop is the one above. The human makes the request, accepts or declines what comes back, declares it done and assesses it. The agent works inside the performance. Draw that line clearly and you know where AI belongs in your organization and where it does not.
How we work with it
Everything here runs on invitation rather than enforcement, because trust cannot be forced and a tool that scores people turns into surveillance. The sequence starts small and gets larger only when the people involved ask for it.
It begins with an individual: a short explainer and a diagnostic that reads one of your own conversation threads against the loop and shows you what was asked, what was promised, and what was left unsaid. For those who want more, a private coach in your inbox asks a question when a request or a yes is thinner than you meant it to be. When several people in one company have looked at their own position in the network, a team report shows the patterns, never the individuals. From there an engagement, Agent in the Loop, maps how the work flows in practice, interview by interview, and the people in each loop decide what should change. What stays installed afterward is a harness that keeps the record nobody had time to keep: what was requested, promised, done and assessed, confirmed by the people involved and visible according to role. The five rungs, one by one.
Nobody is replaced and nobody is made to talk differently. The harness listens to how people already talk and keeps the book.
Going deeper
The ideas above come from a small number of sources, and the reading list gives them from the most readable to the most demanding. The glossary joins the plain words on this page to the proper terms, so a reader who starts here can reach Winograd and Flores without the terms changing underneath them.