Everyone's Talking About Jev. Here's What Actually Worked for Us.
Decision models promised instant, cheap document classification. We tested them against real claims documents. Here's what held up, and what actually worked.
Read the post
What "Explainable" Actually Means in Claims
Explainability gets used in two different ways in claims, and most conversations never bother to separate them...
What "Explainable" Actually Means in Claims
Explainability gets used in two different ways in claims, and most conversations never bother to separate them.
One is technical: a model that shows its reasoning instead of behaving like a black box. You can point to the weights, the prompt, and the confidence score and say, "Here's how it got there."
The other is what actually matters when a regulator, a reinsurer, or your own compliance team asks you to defend a decision made six months ago. Nobody in that room wants to see the model's reasoning. They want to know what was seen, when it was seen, and why the claim was routed the way it was.
Those are not the same problem, and treating them as the same problem is how carriers end up with AI that's technically explainable and practically indefensible.
The room usually wants the second one.
A model that shows its work is a nice feature. It's not what holds up under audit. What holds up is a record: point of receipt forward, every decision logged, every version of the workflow that touched the claim identified, every "why" reconstructable without tracking down the one person who remembers.
That's not a model property. It's an infrastructure property. It lives underneath the workflow, not inside it.
Vendors demo the model. Reasoning traces are visible, legible, easy to put in a slide. Audit trails are boring by comparison, logs, timestamps, version history, the stuff nobody looks at until they have to.
But "until they have to" is exactly the moment explainability is supposed to be ready for. A model that explains itself in the demo and a workflow that can reconstruct a decision a regulator is questioning six months later are two different bets. Only one of them is the one your compliance team actually needs you to have made.
The question worth asking
Next time "explainable AI" comes up in a vendor conversation, ask which kind they mean. A model that shows its reasoning, or a decision that can be fully reconstructed after the fact, what was seen, when, and why.
If they can't tell you, that's the answer.
Get new posts by email.
New posts on claims risk and AI in claims. Unsubscribe any time.
