The Orchestrator

Now the same company, eighteen months later, with the same returns problem.

Wednesday night, after the humans have gone home, the system does the preparation. It has read-only access to anonymized support conversations, return events, website analytics, and the refund-status database; it can build prototypes in a test environment; it cannot change the live site, issue a refund, or message a customer. She gives it the product question: find where customers abandon the return process, propose ten improvements, build prototypes of the three most practical ones, test them against recent cases, estimate the effects, and show the evidence and the assumptions — with no changes to production.

Thursday the room is smaller — six people — and shorter: twenty-five minutes. The pack is already on the screen. Its first finding overturns what the old meeting would have taken as a given: customers are not abandoning at the photo upload at all. They abandon after the return is accepted, when the site tells them to print a label without ever confirming that a label was created. People leave the page and call support because they cannot tell what to do next.

The conversation is different in kind, not just in length.

“That matches what my team hears on the phone,” the support manager says. “They think the label page is broken. They keep asking whether we even received the return.” Then she adds the thing no database contains: oversized furniture uses a different label provider, so the fix has to be tested against that flow before it goes near those customers.

The engineer — the one who suspected the label step last year — looks at the API events in the pack and finds the defect in about a minute: the returns API reports a label as created before the shipping provider has accepted it, so the page has to distinguish “label requested” from “label ready to download.” A small change plus a test. Last year she sat on that suspicion because proving it required data she did not have in the room. This year the data is in the room.

Finance challenges the revenue estimate, and for once the challenge is a real conversation, because the pack shows which cases went into the estimate and which claims the data cannot support. The team picks the first improvement for a limited release, asks that the second be reworked with the corrected shipping status, and defers the third until a support permission is reviewed. Twenty-five minutes, one decision, a named owner for each consequence. No homework. The system implements what was approved, and the room’s next contact with the problem will be measured results.

That is the after. Notice what it feels like from inside. The support manager finally gets to contribute the thing only she knows, instead of defending her department’s backlog. The engineer reviews instead of investigating, which is a promotion in everything but name. The finance lead argues with numbers instead of with the absence of numbers. Nobody performs status for anybody. The meeting is shorter, and it is also the first time all week these six people have actually decided something together.

Amazon proved half of this argument years before agents could assemble evidence. Amazon meetings open with up to half an hour of silent reading of a six-page narrative memo — no PowerPoint — a practice Jeff Bezos has called “the smartest thing we ever did,” because a meeting without a shared artifact produces bluffing: “executives will bluff their way through the meeting as if they’ve read the memo because we’re busy and so you’ve got to actually carve out the time for the memo to get read.”² The six-pager moved the preparation of the argument to before the meeting. Orchestration extends the same discipline from the argument to the evidence: the investigation itself now happens before the room convenes. GitLab, one of the largest all-remote companies, runs its collaboration model on the same instinct — handbook first, meetings as a last resort — because synchronous time is the most expensive time a distributed company buys.³ What orchestration adds is the part neither memos nor handbooks could solve: the evidence no longer has to be gathered by humans at all.

And the rest of the calendar? The standup is gone; there is a board that is always current, because the systems update it as they work. The status meeting is gone; the system reports its own state, with the measurements attached. Sprint planning for delegated work is gone, because there is nothing left to estimate in human terms; what replaced it is a shorter conversation about goals, which is the subject of Section 10.6. What remains is the meeting this chapter has been describing: rarer, shorter, and heavier than the ones it replaced.