Here's an uncomfortable truth the AI industry doesn't like to say out loud: some policy language genuinely reads two ways, and both readings are defensible. An endorsement limits payment to the damaged portion of a roof, and reasonable professionals can scope "damaged portion" differently, with photographs supporting each of them. This isn't a failure of drafting or a gap in the AI's training. It's the nature of language meeting the nature of roofs.
What a system does at that moment tells you almost everything about how it was designed. There are three options, and two of them are wrong.
Two wrong answers
Silent resolution is the dangerous one. The system picks a reading (by training-data average, by whatever pattern happened to dominate) and reports the answer with unearned confidence. The adjuster never learns there was a fork in the road. Multiply that across a portfolio and you have thousands of judgment calls made by no one, invisible until an auditor or an attorney makes them visible for you. The shrug is the useless one: "ambiguity detected, please review." No analysis, no framing, no consequences. Thanks. The adjuster now does the full reading themselves, which is precisely the work the system was bought to absorb.
The third option: make the fork visible and price it
The right behavior is what a good senior colleague would do: lay out both readings honestly, say which one they'd lean toward and how strongly, and (this is the part software almost never does) put the dollar consequence on each option. Repair basis: net stays at one figure. Replacement basis: adds a specific amount. Exclude the line: zero. In ClaimVision, that's a first-class object we call the judgment call: the operative policy language on one side, the claim evidence on the other, ClaimVision's reading with its confidence stated, and each choice carrying its price. The adjuster rules. An override can carry the adjuster's written reason. Everything is preserved.
Attaching dollars to each option isn't decoration. It's what makes the decision real. "Interpretation A vs. interpretation B" is a seminar question. "This reading versus $1,708.08 more" (on an example claim) is a claims decision, made with full knowledge of what's at stake, by the person accountable for it. And because the ruling and its reasoning enter the permanent record, the file can defend the decision years later and teach the next adjuster who hits the same fork. You can see the whole flow on a worked example in the claim walkthrough.
"It depends" is the most expensive sentence in insurance when it's resolved silently or shrugged at. Resolved in the open, with evidence, confidence, prices, and a human ruling on the record, it becomes something else: the moment your organization's judgment compounds instead of evaporating. That's not AI replacing the adjuster. That's the adjuster, finally armed.
