The confidence score is only as good as the data under it

The most interesting model this year can't write a sentence

Last month TypeSafe released Jev, and unlike every other model, Jev’s biggest brag is all the things it does not do. It doesn't write. It does not summarize, and most notably, it does not chat. You give it a situation and a set of probable answers, and it responds with a probability and a confidence score.

Some people have called it a smart if-statement. But I think that kind of undersells it. For three years, we have asked generative models to make decisions, then dressed the answer up as prose. Jev has no costume.

A generative model gives you an answer and an essay. A decision model gives you an answer and a number, and only one of those can run a workflow.

That number can turn every decision into a routing decision. Above the threshold, the system acts. Below it, a person looks. That threshold is now one of the most consequential settings across enterprises.

Confident and correct are not the same thing

Let's turn this into a real practical example. A renewal lands in RevOps. The decision model is asked one question: route this account to expansion or to retention? It answers expansion, 93% confident. The workflow moves it to expansion.

The CRM classifies that customer as enterprise tier. The billing system shows they downgraded in March, but the downgrade was processed as a credit rather than a plan change, so the tier never updated. Finance counts that revenue one way. Sales counts it another. The model was given the CRM's version and was very sure about it.

So an already somewhat disenfranchised customer gets an upsell call or email. This signals one thing, and we all know it's not good. This company is out of whack with our client needs.

That is why we have to remember that confidence without conviction is meaningless.

Where Preql comes in

That certainty is what we build at Preql.

We connect the systems and the spreadsheets, find where they disagree, and fix it at the source. Every number gets one agreed definition, and every figure has a trail back to where it came from, so there is a closed, confined audit loop. Because if nobody is willing to put their name to what a number means, then why would you let software be confident about it?

My bias, and my bet

The companies that get the most from this shift will be the ones that settled what their numbers mean before asking a machine to be sure about them. Thresholds are easy to tune. Definitions are not.

It's easy to rent the model. But it will only work if you own the meaning.

If you're putting decision models into finance or revenue workflows and want to know whether your numbers are ready for them, send me a message.

Or if you want to debate where this new class of models is heading, I'm always up for that conversation.

What we've been up to and where we’ll be

Jerusalem. Last month I visited the CRB team at their Jerusalem office. It's one thing to talk about data on a call. It's another to sit in the room with the people who live with it every day.

A quick visit to Jerusalem.

Columbia. This semester, Columbia Business students are working on a capstone project for Preql AI. Few things humble your thinking like a room of students who aren't obliged to be polite.

Speaking at Columbia last week.

San Francisco. Next stop is SF Tech Week. If you're there, come find me!

Back in two weeks, Gabi!