Most AI coding tools fail one simple test. Can they show their evidence?

I built a risk adjustment gap engine for our practice. It finds conditions that haven't been billed this year, using plan claims, our own billing and the EHR problem list.

The first version matched diagnoses on approximate names. Most of its problem list suggestions were wrong. So it switched to exact codes, and now every suggestion shows where it came from.

Then our practice manager checked six flagged patients by hand. The review found false alarms. A more serious code we had already coded. A cancer that was history only. Each one became a rule.

A worklist full of false alarms gets ignored. Evidence is what makes a team trust it.

So before you buy an AI coding tool, ask the vendor one thing. Can every suggestion show its evidence?

The full story, with the numbers: Coding gaps, with the evidence attached.