AI agent · Insurance · Score & detect
Subrogation detection
Recovery opportunities are usually written somewhere in the claim file: an adjuster's note, an incident report, a repair invoice. The agent reads every file and surfaces those worth a recovery specialist's time, quoting the passage that suggests it.
Typical volumes for this process, not a client figure.
Nobody has time to read every file for recovery potential.
Every file read, the recoverable ones surfaced with the reason.
Where the time goes today
When you pay a claim caused by someone else, such as a contractor, a manufacturer, another driver or a neighbouring property, you may be able to recover part or all of the payment from that party or their insurer. Spotting those claims depends on someone noticing the signal in the file: a note that the fire started in a newly fitted appliance, a report that the other vehicle ignored a red light, an invoice showing the leak came from the flat upstairs.
Adjusters are focused on settling the claim, and flagging recovery potential is one more field on a busy screen. Recovery teams usually review files by rule, such as certain claim types above a value, or by sampling. Files below the value line, or of a type the rule does not cover, are rarely read for recovery at all, while the time limits for pursuing recovery keep running after the file is closed.
The result is not a visible error. It is money not recovered, which nobody sees because nobody looked.
How the agent works
- Read the fileThe agent reads the settled or nearly settled claim file: adjuster notes, expert and incident reports, invoices, correspondence and the description of the loss.
- Find liability signalsIt looks for a third party whose act, product or property may have caused the loss, and extracts who they are, what they did and the passage that says so.
- Score the potentialIt scores the file on the strength of the liability evidence, the amount paid, whether the third party can be identified, whether evidence was kept, and the time left under the limitation rules you configure.
- Apply your thresholdFiles above your recovery threshold are surfaced. Files below it are recorded with their score and reason, so the threshold can be tested later.
- Prepare the referralFor each surfaced file, it drafts a referral for the recovery team: the third party, the basis for liability, the amount paid, the evidence held and the deadline.
What stays with a person
The recovery specialist decides whether to pursue. That decision weighs things the file alone may not settle: the strength of the evidence, the cost of pursuit against the likely recovery, whether the third party is also your customer, and how a recovery action would affect the policyholder. The agent's score is an input to that decision, with its reasoning visible, not the decision itself.
Letters of claim, negotiation and any legal action stay with the recovery team or their counsel. The agent can draft a first letter from an approved template, but a person reviews and sends it.
What it reads, what it produces
| It reads | It produces |
|---|---|
| Claim notes and adjuster reports | A score for each settled file, with the factors behind it |
| Police, fire and incident reports | A ranked list of files worth pursuing, quoting the passage that suggests liability |
| Repair invoices, expert and engineer reports | A referral pack for each surfaced file |
| Correspondence with the insured and other parties | A view of recovery deadlines across open referrals |
| Payment records for the claim | A record of files scored below threshold, kept for later review |
| Your recovery rules and limitation periods by claim type |
Controls that come with it
- The threshold is set from the cost of a specialist's review against the typical recovery for that claim type, and revisited as results come in.
- Each month a specialist reads a random sample of files scored below threshold, to see what the agent is missing.
- Every score records the passages and factors that produced it.
- No letter of claim is sent without a person's approval.
- A referral can be withdrawn, and a file re-scored when new information arrives.
How you know it works
- Share of past files with known recoveries that the agent would have surfaced
- Referrals accepted by the recovery team against referrals made
- Amounts recovered from agent-surfaced files, tracked separately
- Recoveries lost to expired time limits, before and after go-live
Is your process ready?
- Written rules: the recovery team can state which claim types, amounts and liability patterns they would pursue.
- Systems: settled claim files, including notes and attachments, can be read in bulk under your claims system licence.
- Cheap check: a specialist can accept or reject a referral in minutes, because the relevant passage is quoted.
- Volume: thousands of settled files a month, most with nothing to recover, is exactly where reading every file by hand is impossible.
- Consistent process: specialists agree on what makes a file worth pursuing, or can agree when shown examples.
The five candidacy checks are explained, with an exam, in the free Module 01.
What goes wrong
- A threshold set too low, which floods the recovery team and teaches them to ignore referrals.
- Using past recoveries as the only test set, which only teaches the agent what the team already found.
- Files where the key evidence is in photos or handwritten reports the agent reads poorly.
- Surfacing files after the time limit has already passed.
Questions we get
Why read every file rather than only high-value ones?
Value rules are a sensible way to ration a person's time, but they skip everything below the line. Reading a file costs the agent little, so the threshold can move from 'worth reading' to 'worth pursuing'. Whether reading below the line repays the build depends on your claim mix, and testing on your past files is how you find out.
How do you test it when past recoveries are incomplete?
Past recoveries show what your team found, not what was there. So the test set combines known recoveries with a sample of closed files that specialists read afresh during the build. The agent is measured on both. The second set is the more useful, because it shows whether the agent finds recoveries the old process missed.
Can it run on files that are still open?
Yes. Scoring open files gives the recovery team time to ask the adjuster to preserve evidence, such as a failed part or a damaged appliance, before it is thrown away. The score is updated as the file develops. Early scores are less certain, and the referral says so plainly.
Does it know the limitation rules in every jurisdiction?
No, and it should not guess. Time limits are configured by your team per claim type and jurisdiction, from advice you already hold. The agent applies them to the dates of loss and payment in the file, and flags any file where it cannot tell which rule applies.
Want this agent on your process?
Tell us about your version of this process — volumes, systems, what goes wrong. A person answers with an approach and a price, usually within two working days, or tells you it is the wrong project.