Skip to content
← Blog

Confidently Wrong

The most dangerous answer a system can give isn't "I don't know" — it's a confident answer about a world that has already changed. We spent a week teaching VERA the difference between what it remembers and what's true.

Every operator has been burned by a confidently wrong answer.

"Sarah owns that account." She doesn't — she moved to the other client in April, and the person saying it just hadn't caught up. "The Meridian retainer has forty hours left this month." It doesn't — that was before the scope change last Tuesday. "We've got a developer free in June." We don't — that slot got committed on Thursday and nobody updated the plan in their head.

None of these are lies. They're worse than lies, in a way, because the person saying them believes them. They're answers about a world that existed last week, delivered with the confidence of the present tense. And you act on them — you promise the client, you assign the work, you tell the founder the number — because the answer came back fast and sure, and fast and sure is what you've been trained to trust.

An agency's state is never still. People get reallocated. Roles change — the contractor you brought on for one project is now effectively running two. Retainer hours move every time someone logs against them. Projects open, close, get renamed, get cancelled. The truth at 9am on Monday is frequently fiction by Thursday afternoon. This is not dysfunction; it is simply what a busy shop looks like. But it means that anything — or anyone — answering from memory is answering about a place that no longer exists.

This week we found VERA doing exactly that, in a handful of small but telling ways, and we spent the better part of it stamping the behaviour out.

It looked harmless at first. Someone asked whether a particular colleague was an owner, and VERA said no, she's a regular user — except she was an owner, had been for weeks. VERA hadn't lied; it had simply assumed. It knew who the person was, recognised the name instantly, and then guessed at the part it hadn't bothered to check. The recognition was real. The fact attached to it was stale.

The same shape showed up elsewhere. When VERA went to confirm a change to someone's allocated hours — the moment where it reads the plan back to you and waits for a yes — it was sometimes showing the wrong starting number. "I'm about to change this from forty hours to thirty-five." Except the real figure was forty-seven, not forty. The change would have landed correctly; the machinery underneath always used the true value. But you, reading the confirmation, would have approved a different change than the one you thought you were approving. The whole point of reading a change back is so the human can catch it. A confirmation built on a guessed number quietly defeats its own purpose.

And one more: when you referred to a project by a slightly-off name — a typo, an abbreviation, the way you actually talk — VERA would sometimes charge ahead, collect every other detail from you, walk you all the way to the final confirmation, and only then discover the project didn't exist by that name. You'd done all the work for nothing. The thing it should have sorted out first — which project do you actually mean? — it left until last.

The fix, stated plainly, is a single discipline: never state a fact you haven't just checked. VERA's own memory is now treated as a hint about what you meant — never as a source of what's true. Before it tells you someone's role, it looks it up, every time. Before it reads a change back to you, it reads the real current value first, so the number you're approving is the number that's actually on the books. Before it asks you for the details of a request, it confirms the thing you're referring to actually exists — up front, at the first turn, not after you've spent four messages filling in a form for a project that was never there.

What makes this worth a week of work, and worth writing down, is that it runs against the grain of what makes these systems feel good to use. A system that pauses to check looks very slightly slower than one that answers instantly from what it thinks it knows. The instant answer feels more capable. It is, in fact, the more dangerous one — because the cost of a confidently wrong answer isn't paid at the moment it's given. It's paid later, when you've already promised the client the hours, already told the founder who owns the account, already made the decision on intelligence that was true once and isn't anymore.

There is a version of "always current" that is pure theatre — a dashboard that refreshes every sixty seconds but reports figures it cached an hour ago. Currency without accuracy is worse than a stale report you know is stale, because it wears the costume of the truth. The entire promise of a tool that lives where the work happens and answers in the moment is that the answer is good in the moment. That promise only holds if the tool knows, at every turn, the difference between what it remembers and what is true right now — and refuses, every single time, to confuse the two.

VERA now refuses. It is a small-sounding discipline. It is most of what trust is made of.

VERA by talktalkmake