Small insurance claims: AI sorting between direct settlement and human review
In brief. An insurer can settle its small claims faster by letting an AI model sort incoming claims into two queues: those that meet all the conditions set by the company for direct settlement, without manual intervention, and those that must go to a claims handler. The model does not assess coverage in place of the insurer; it checks a grid of criteria defined by the claims department, and anything that falls outside the framework is sent back to a person. No refusal is issued automatically.
This article describes an illustrative scenario. It is not a named client and we give no quantified results: they depend on your portfolio and your thresholds.
The problem: simple files treated like complex ones
Broken glass, a damaged phone, damaged luggage, a small vet bill, a low-value private liability claim: depending on the line of business, a share of claims involves modest amounts and clear facts; this share is to be measured on your portfolio. In this scenario, these files follow the same route as the others: opening, data entry, document checks, coverage verification, decision, payment, letter.
In this scenario, policyholders wait for an answer that could have been immediate, and experienced claims handlers spend time on files that do not need their expertise.
The process before and after
Before. Each claim, whatever its size, is taken on by a claims handler who reads, enters, checks and decides.
After. Sorting takes place as soon as the claim arrives.
- The claim and its documents arrive through the portal, by email or by scanned post.
- The model extracts the useful elements: type of claim, date, circumstances, amount, documents supplied.
- Deterministic rules are applied by the management system, not by the model: policy in force on the date of the claim, cover taken out, deductible, ceilings.
- The model checks the criteria that require reading: consistency between the account and the documents, presence of a third party, mention of an injury, invoice legible and complete.
- If everything is met, the file goes to direct settlement. Otherwise, it goes to human review with a summary and the reason for the referral.
The split of roles is important: what can be calculated with certainty stays in the management system. The model is used only to read what a classic program cannot read.
The rules and thresholds belong to the insurer
| Criterion | Direct settlement possible | Human review required |
|---|---|---|
| Amount | Below the threshold set for the line of business | Equal to or above the threshold |
| Coverage | Cover clearly applicable, policy in force | Possible exclusion, doubt about cover, unpaid premium |
| Nature | Simple material damage | Bodily injury, third party involved, possible recourse |
| Documents | Invoice, photo or quote matching the required list | Missing, illegible, altered or inconsistent document |
| History | Claims frequency within the defined norm | Repeated claims, late notification, very recent policy |
| Warning signals | None | Indications raised by the company's anti-fraud rules |
| Clarity | Consistent account | Contradictions, ambiguity, several claims in one notification |
Three principles are added to the grid. A single unmet criterion is enough for a referral to review. The model never refuses a benefit: only a claims handler can do so. And the thresholds are parameters that the claims department changes itself, by line of business and by product, without touching the model.
Human in the loop
- Sample checks of claims settled directly, at a frequency set by the company.
- Systematic review of any file flagged by the anti-fraud rules, whatever its amount.
- A business owner designated to validate every change to the grid.
- A simple channel for contesting, for the policyholder, which always ends with a person.
- Possible suspension of direct settlement at any time, line of business by line of business.
Audit trail and explainability
For each file, the system keeps the version of the rules, the model version, the extracted elements, the result of each criterion, the queue selected and the human interventions. The explanation takes the form of a list of criteria checked, readable by a claims handler, by internal audit or by the supervisory authority.
This level of documentation also meets an expectation of the regulator. In its Guidance 08/2024, FINMA expects supervised institutions to identify, assess, manage and monitor appropriately the risks linked to the use of AI, and among these risks it cites the robustness, accuracy, explainability and bias of models, as well as dependence on third parties.
Errors: which ones, and who bears them
- Claim settled in error. The cost is capped by the threshold and borne by the insurer. It is measured by sample checks. This is also where low-value fraud occurs: a known and stable threshold can be exploited, hence the value of monitoring frequency per policyholder and of varying the checks.
- Claim sent to review unnecessarily. The cost is claims handler time. The system is tuned to prefer this error.
- Undetected bodily injury or third party. This is the error to avoid. Strict rules govern it: certain words, certain documents and certain lines of business always trigger a human review.
The policyholder must not bear a system error. If they disagree, their file is taken over by a claims handler.
Data protection and supervisory framework
A few reference points, to be validated by your legal department and your compliance function:
- Automated decision. Art. 21 of the Federal Act on Data Protection (FADP) requires the person to be informed when a decision is taken exclusively by automated means and has legal effects on them or significantly affects them; the person may ask for a review by a natural person. The law provides an exception where the decision is directly related to the conclusion or performance of a contract and the person's request is granted. A direct settlement that accepts the policyholder's request fits this logic; an automated refusal does not. This is one more reason never to automate refusals.
- Sensitive data. Health data is sensitive personal data (art. 5, let. c, FADP). Claims involving bodily injury are therefore excluded from direct settlement for two reasons: business risk and the nature of the data.
- Impact assessment. Art. 22 FADP calls for a prior impact assessment where processing is likely to result in a high risk.
- Supervision. Insurance companies are supervised by the Confederation under the Insurance Supervision Act (ISA). FINMA's expectations on AI governance and outsourcing must be taken into account from the design stage.
Why an open-source model hosted in Switzerland can be suitable
The task is narrow and repetitive: read a short claim, extract a few fields, check a list. A medium-sized open-source model (Llama, Mistral, Qwen or Gemma families) can be enough, subject to a test on your files.
Three arguments weigh in favour of this option. Volumes are high, which makes a fixed cost more attractive than pay-per-use billing. The model does not change without your decision, which makes auditing easier. And dependence on a third-party supplier, which FINMA cites among the risks, is reduced. If your claims management system is already in a given cloud, a proprietary model in that same environment may also be justified. Our decision guide details the criteria.
How a pilot runs and is measured
- Test on historical data. The model sorts closed claims from one line of business, and its sorting is compared with the decisions taken.
- Parallel operation. New files are sorted by the model and handled normally by the claims handlers.
- Limited direct settlement. Opening on one line of business, with a low threshold and reinforced sample checks.
| Indicator | What it measures |
|---|---|
| Share of files in direct settlement | The real potential |
| Agreement rate between the model and the claims handlers | The reliability of the sorting |
| Settlements in error, found by sampling | The residual financial risk |
| Undetected bodily injuries or third parties | The risk to keep at zero |
| Time between claim and payment | The benefit for the policyholder |
| Change in the frequency of claims below the threshold | A windfall or fraud effect |
| Disputes and complaints | The perceived quality |
Limits: when not to automate
- Any bodily injury and any claim involving a third party.
- Cover whose application requires interpretation.
- Recent products, without enough history to test on.
- Lines of business where supporting documents are too varied to be read reliably.
- Situations where the claims department cannot manage to write its own rule.
Frequently asked questions
Does the model decide on coverage? No. The policy conditions are checked by the management system and, in case of doubt, by a claims handler. The model reads the documents and checks criteria.
Can refusals be automated too? We advise against it. A refusal has consequences for the policyholder and must remain a human, explained decision.
Does direct settlement encourage fraud? It can attract it if the threshold is known and the checks predictable. Hence the monitoring of frequency per policyholder, the sample checks and the retention of anti-fraud rules upstream.
Do we need to replace our claims management system? No. The sorting plugs into the existing system, which remains the reference for policies, payments and history.
Which line of business should we start with? The one where files are numerous, amounts low and documents standardised. The discovery workshop is used to choose it.
Going further
The same scheme applies to minor liability claims at a hospital and to sorting KYC files and alerts.
Our AI consulting starts with a discovery workshop to choose the line of business and the rules. Contact us to talk about it.