Why the pilot failed and what to do next

A post mortem that lands on “more data would have helped” and “it did not suit the use case” has given you two true things about symptoms and nothing about the cause. If you suspected that while you were reading it, you were right to. Both are what people write when nobody in the room can point at a day the thing broke, because on no day did it break.

So this route is built differently from the others through the Cabinet, ArkOne’s reference design for an executive agent. The others teach the design, a part at a time. This one is mostly case notes: single accounts of single things going wrong, because a chief executive whose pilot has just been cancelled is not looking for a curriculum. You are looking to recognise what happened to you, and a page describing a control working as intended will not help you do that.

The opening account of that eleven-minute meeting comes first because it gives you the distinction the route runs on. A failure that announces itself gets fixed. A failure that produces confident, well-formed, wrong work does not, because nothing in it looks like a failure.

One thing to be plain about, because it bears on whether you trust any of this. Of the eighteen accounts in the case notes, sixteen are written as illustrations and two quote a real record. Each says which it is on its own exhibit, so read the label there rather than taking mine. Nothing here reports on your pilot or on anybody else’s, and none of it is evidence about a running system.

The nine accounts after it are grouped by what somebody in your company would have noticed at the time rather than by which part of the machine broke, and that grouping is the argument here. A vendor sorts failures by mechanism, because a mechanism is what gets patched. You experienced them as four kinds of nothing.

The first pair is the purest kind, which is nothing at all: a report that arrives on schedule and confidently describes a company nobody has loaded yet, and an answer that lost half of its own working before it reached you. Each is read, believed and filed.

The second pair is a number nobody could get twice, and it turns a pilot review into an argument about whether the finance lead read the invoice correctly. One is a bill many times the forecast, produced by a single helpful line of text. The other is two identical runs of the same work priced differently, with nothing to investigate either time.

The three after that are the group most likely to be yours, an answer that was confident and wrong. A forum figure quoted in the same voice as your own approved number. A check that found a problem that was not there, then counted it twice. A company-sized decision answered by the cheapest model on the list because the question was short.

The last pair is a message that should not have gone: a half-answer read as your approval, and a whole function running for months on an authority nobody chose. They sit last because they outlive a cancelled pilot. The others cost you usage; these leave you with something sent in your name and nobody to name as having decided it.

The hinge is the eleventh page, and everything after it is about the next attempt rather than more ways to lose the first. One page is how a change is examined before it reaches a company, with the limits of that examination stated in the same breath as its use. The other is the appointment nobody made the first time, which is the difference between an owner and whoever happened to reply to the vendor’s email.

For a shorter route, read the opening account, the forum figure and the mandate nobody set. That is about twenty minutes, and it leaves you able to ask whoever ran the pilot the one question with a real answer: what in it would have made a wrong answer visible while it was still cheap to catch.

Left out are vendor selection and the bill. A second attempt that opens by reopening procurement spends its first month there, so the vendor side waits in the Questions for the Vendor cluster and the cost mechanisms in the Cabinet Papers. Two accounts here touch the bill only because the bill was the thing nobody could reproduce, which is a diagnosis rather than a budget. If reading them tells you the diagnosis was the bill after all, what it will cost takes the same two accounts and builds the arithmetic around them; if it tells you a control was missing rather than mispriced, the governance hour walks every place a refusal can sit.