An advertising platform reported 303 conversions and $38,227 of value against $37,947 of spend in a single month. Read at face value, that account broke even. Following the same clicks through to work that actually got completed, 46 jobs could be traced.
Nothing was broken. Nobody entered anything wrong. The platform counts a conversion when the phone rings, which is the only thing it can see, and it has no way of knowing whether anyone did the work or what was paid for it. Then its bidding uses that count to decide who sees the ads and what each click costs. So the budget had been steering itself toward whatever produces phone calls rather than whatever produces paid work, every day, for as long as it had been running.
That's not a story about bad software. It's a story about what a number does when it arrives looking finished.
The answer arrives, and nothing about it looks wrong
Ordinary system failure announces itself. A workflow breaks, something stops arriving, someone complains, and within a day or two the problem has a name and an owner.
The failure described above has none of that. The number was formatted correctly. It carried a currency symbol and two decimal places. It appeared in the interface where numbers appear, on schedule, next to other numbers that were fine. There is no error state for a figure that is arithmetically correct and answers a slightly different question than the one being asked.
And the person receiving it has no independent way to check it. That is not carelessness. Not having a way to check it is precisely why they asked in the first place. Anyone who could verify the answer from memory would not have needed to ask.
This is the part of the current wave of business AI that gets the least attention, because it is the least demonstrable. You can demo a system answering a question. You can't demo one being quietly wrong for four months.
The connection stopped being the expensive part
Here is the thing that costs a sale to say out loud, so it's worth saying early.
Most business software now ships some form of plain-English access to its own data. Not as a paid add-on. As a feature, on plans companies already pay for. Turning it on is usually an afternoon of admin work rather than a project, and the list of tools that include it gets longer every quarter.
So if everything a business needs to ask lives inside one system, the answer is to use what that system already gives away and keep the money.
That was the honest answer for one prospect this month, a creative studio whose work all lived in a single project tool. Their vendor ships a connector. Half of what they described could have been running by the end of an evening on a twenty-dollar subscription with nobody hired. They were told that, they did the arithmetic against the cost of a part-time manager, and the manager won. The arithmetic was correct and there was nothing to argue.
What did not get cheaper is everything the connector cannot see: joining systems that were never introduced to each other, agreeing what a word means before two people use it differently, and knowing how much of the business any given answer actually covers.
That last one is the subject of the rest of this.
Coverage is the number nobody prints
In the same account as the advertising figures above, roughly 29 per cent of inbound calls could be tied to a completed job.
Every figure derived from that data inherits the ceiling. A cost-per-booked-job number calculated on it is not an estimate and not a guess. It's a correct calculation over 29 per cent of the business, presented in exactly the same visual language as a calculation over all of it.
Nothing in how a number is displayed tells you which of those two you are looking at. A dashboard does not print the share of reality it can see. Neither does an assistant answering in a sentence. The confidence of the presentation is constant regardless of the coverage underneath, which means the presentation carries no information about reliability at all.
Which gives the rule this whole piece exists to deliver, and it is worth stating flatly because it governs every decision downstream of it: a number that looks authoritative and rests on a third of the picture is worse than no number at all. No number leaves judgement intact. The owner knows they are guessing and behaves accordingly. A wrong number replaces judgement with false precision, and nobody behaves cautiously around two decimal places.
Anything built on that account now states its own coverage next to the answer, permanently. Not as a footnote. As part of the answer, in the same sentence, because a coverage figure that lives somewhere else is a coverage figure nobody reads.
What the first discovered error actually costs
The obvious risk is a bad decision made on a wrong number. That one is real and survivable. Businesses make bad decisions constantly and recover from most of them.
The second risk is worse and gets discussed almost never.
The first wrong answer that somebody catches does not discredit only itself. It discredits every correct answer that came before it and every one that comes after. Trust in a system like this isn't assembled number by number, and it doesn't erode gradually. It holds completely until one figure is caught being wrong, and then it collapses in a single piece.
What happens next is quiet. Nobody announces that they have stopped believing the system. The organisation drifts back to running on instinct while continuing to pay the subscription, and the reporting keeps generating, and meetings keep referring to it in a slightly looser way than they used to. Saying it out loud would mean admitting the whole thing was a mistake, and there's rarely a moment when that's convenient.
This is why the order of work matters more than the tooling. Get the measurements right, then let something answer questions on top of them. The other order produces confident advice resting on a foundation that is a third complete, and it produces it fast, which is the part that makes it expensive rather than merely useless.
Whether a question is answerable at all is a different question
None of the above is about whether the data can answer. That is a separate and prior problem, and it has its own shape: some questions are answerable from a system you already run, some need one specific thing that was never written down, and some cannot be answered by anyone because nothing in the business records them.
We went through that sorting with one operator recently across twenty numbered questions, and roughly a fifth of them turned out to be unanswerable by anybody, at any price, because the field that would carry the answer had never been filled in. That breakdown, and the three rows every business question falls into, is written up separately.
Sort first. Then read this one. The two failures are independent: a question can be perfectly answerable and still produce a number that sees a third of the business, and those need different fixes.
The check to run on the numbers you already act on
This one isn't about future questions. It's about the figures already circulating in your business, the ones that appear in a monthly review and get acted on without argument.
Take three of them. For each, answer two things without opening anything.
What share of the relevant volume does this number actually see? Not in principle. Of all the calls, leads, jobs or invoices that should feed it, what fraction can be traced end to end right now?
And which definition is it using? If two people in the business were asked to define the metric, would they say the same thing? Revenue at booking and revenue at invoicing are both correct and they are not the same number.
Most people cannot answer either question for a number they have been acting on for a year. That's the finding, and it's available for free, this afternoon, without hiring anyone or connecting anything.
If the answers come back thin, the fix is not a better model or a better dashboard. It is printing the coverage next to the number, every time, until nobody in the business can read a figure without also reading how much of the business it saw. The commercial version of that work lives on the team side. The check itself is yours to run.