Insurance fraud detection AI system. $840,000 recovered in 4 months
Zeacle AI deployed an autonomous fraud intelligence platform and claims triage system in 23 days. The platform recovered $840,000 in confirmed fraud in four months, increased claims processing capacity by 43% with no additional headcount, and reached full investment payback in 74 days.
Claims volume was outpacing the team's ability to triage, investigate, and settle accurately. Rules-based fraud detection was catching obvious cases and missing coordinated schemes hidden across providers, claimants, and time. Reserve accuracy depended on individual adjuster experience rather than systematic access to six years of historical settlement data. Capacity was capped and fraud leakage was rising.
We built an autonomous fraud intelligence platform with a claims triage layer trained on six years of historical settlement and fraud data. Every inbound claim is automatically classified, prioritised, and routed. A pattern detection layer surfaces complex multi-party fraud that no rules engine can catch. A reserve accuracy tool gives every adjuster systematic access to comparable historical settlements. Human investigators retain full decision authority on every flag.
- →Claims triage layer with automatic classification, prioritisation, and routing
- →Pattern detection over six years of historical claims and settlements
- →Provider, claimant, and event graph for coordinated fraud detection
- →Reserve accuracy tool with comparable-claim retrieval
- →Investigator console with full evidence trail per flag
- →Continuous evaluation against new confirmed fraud outcomes
Claims data corpus consolidated from six source systems.
Where the carrier was at start of engagement
The carrier was managing a steadily rising claims volume against a fixed team. Capacity was capped, cycle times were drifting, and the special investigations unit was working from a backlog that grew faster every quarter. Rules-based fraud detection had been in place for over a decade. It caught the obvious cases. the lone claimant submitting an obviously inflated repair invoice. and it missed almost everything coordinated. Multi-party schemes routed through colluding providers, repeat claimants using different addresses, and staged events distributed across months were invisible to a system designed to score one claim at a time.
Reserve accuracy was the second pressure point. Adjusters were setting initial reserves based on personal experience and the few comparable claims they happened to remember. The carrier had six years of paid claim data that, properly indexed, would have made every reserve decision dramatically more accurate. None of that data was systematically available at the point of decision.
The platform we built
We deployed an autonomous fraud intelligence platform with three layers working in concert. The claims triage layer reads every inbound first notice of loss together with attached documents, claimant history, policy data, and third-party feeds, and produces a structured triage output in under thirty seconds: severity score, complexity score, fraud risk score, recommended handler, recommended reserve range, and the specific evidence the next handler will need. The pattern detection layer is a graph engine. Every claim, claimant, provider, payee, address, phone number, vehicle, and incident is a node and every relationship between them is an edge. New claims are scored against the full graph, surfacing densely connected subgraphs that no rules engine can catch.
The reserve accuracy tool runs alongside triage. For every new claim it retrieves the closest historical matches. same loss type, same severity band, same jurisdiction, same coverage shape. and recommends an initial reserve range with the comparables attached. The investigator console pulls all three layers together into a single workflow. Every flag carries its evidence trail and every action retrains the underlying models.
Why human authority stays untouched
Every fraud flag, every reserve recommendation, and every triage routing is reversible by a human. The platform's job is to surface the right work to the right person with the right context. The investigator decides whether a flag becomes a referral. The adjuster decides the final reserve. The handler decides whether the routing is right. The system never closes a claim, never denies a payment, never makes a regulatory decision. That separation is the only basis on which a fraud AI system can be defended to a regulator and trusted by an operations team.
What the first four months looked like
By the end of month four the platform had surfaced confirmed fraud totalling $840,000. claims that would have been paid out without the system, recovered or denied before settlement. Processing capacity per handler per week climbed by 43 percent against the same headcount, almost entirely through the triage layer eliminating the constant context-switching that had been eating handler time. Reserve accuracy improved against final paid amounts and the capital efficiency benefit alone justified the build.
The payback period landed at 74 days. The carrier has continued to expand the system into adjacent lines of business and is on track to extend the platform into the subrogation workflow next quarter, where the same retrieval and pattern-detection infrastructure unlocks an entire second category of recoveries.
What we learned doing it
The biggest single lesson from this deployment was the importance of week one. The data lived in a claims system specified in 2003, a document repository nobody had audited in a decade, a payment system, an email mailbox where loss notices still arrived in PDFs, and a stack of spreadsheets where the special investigations unit kept its working memory. Consolidating that into a usable training corpus while respecting every regulatory and privacy constraint the carrier operated under was the largest single block of work in the project. Once the corpus existed, the model layer came together in days. Carriers chronically underestimate this and chronically overestimate the model work.