Subrogation recovery, the money an insurer is owed back, is often hiding in the details of claim files no one has time to read closely. We looked at using large language models to find it.
Subrogation is an insurance term worth explaining, because the whole project sits inside it. When an insurer pays a customer's claim but someone else actually caused the damage, the insurer has the right to recover that payout from the responsible party's insurer. Pursued properly, it's real money coming back in the door.
The catch is that finding these cases means reading claim files closely, and no team has the hours to read all of them. So recoverable money slips through, not because anyone decided to let it go, but because nobody had time to spot it.
As always, we worked through the six CRISP-DM steps.
Most of what we had to learn was how subrogation works day to day. The clues that someone else was at fault live in the details of a file: adjuster notes, police reports, descriptions of what happened. A specialist can spot them, but only in the files they have time to open.
We also learned that the two possible mistakes cost very different amounts. Missing a recoverable claim loses the full payout. Wrongly flagging one costs a specialist a few minutes to dismiss it. That difference shaped everything downstream.
Claim files are long, inconsistent, and written by different people for different purposes. The preparation work was getting them into a shape where a model could read each file whole, notes and reports together, the way a specialist would.
Reading dense text and flagging what matters is exactly what large language models are good at. We looked at using an LLM to review each closed file and flag the ones with likely recovery potential. Every flagged file goes to a specialist who makes the actual call.
Because a miss costs the full payout and a false flag costs minutes, a system like this should lean toward flagging. The measure of success is whether the short list catches the recoverable claims, and whether specialists find the list worth their time. If they stop opening it, the system has failed no matter what its accuracy says.
The deliverable is a short list with the model's reasoning attached, landing in front of a person who knows subrogation. Nobody's job disappears. The money just stops slipping through.
This was an applied study. We're glad to discuss how a pilot like this would be run responsibly in a regulated setting.
That's a language-model problem. Let's talk about it.