$10B+a year
What fraud takes.
Health care fraud losses alone are estimated in the tens of billions of dollars a year2.
V3Qtor AG/Industries/Financial services & insurance
Claims in a day. Models that stay calibrated.
Insurers lose money in the queue: claims that take weeks, fraud found after payment, models that drift between calibrations. We settle claims in a day and recalibrate the model every week.
Before and after
Where an insurer’s combined ratio moves.
Claim cost and fraud carry most of the bridge. Calibration and reporting effort follow.
Combined-ratio bridge, insurer engagement
Index points on a baseline of 100. Illustrative.
Illustrative bridge1. Hover a bar for the lever behind it.
80%of AI projects
Why risk models stall.
More than 80% of AI projects fail, twice the rate of other IT projects; the second root cause is the data3.
Gartner names weak data quality and insufficient risk controls among the reasons GenAI projects are abandoned4.
95%of GenAI pilots
No measurable P&L impact.
MIT NANDA's 2025 finding across 300 public deployments; purchased or partnered deployments succeed about twice as often as internal builds5.
Replayableevery score, with its inputs
What a governed model returns.
Fraud detection across millions of daily transactions in a typical engagement1, with every score replayable.
The numbers behind the sector
Most AI projects fail. The data under the model is the usual reason.
RAND, Gartner and MIT NANDA, cited in full below.
2×failure rate
AI projects versus other IT projects
RAND found more than 80% of AI projects fail, twice the rate of IT projects that do not involve AI3.
RAND Corporation, 2024
30%abandoned
GenAI projects after proof of concept
Gartner's prediction for end-2025, citing data quality, risk controls, cost and unclear value4.
Gartner, August 2024
67%success
Purchased or partnered deployments
Against roughly a third as often for internal builds, in MIT NANDA's 2025 analysis5.
MIT NANDA, 2025
Use cases
What we run in insurance operations and risk.
Insurance operations
Priced per risk, settled in a day, fraud stopped at intake.
Actuarial Risk EngineIndividual risk prediction for pricing and interventionDecision & Risk SciencesPer policyrisk priced individuallyIntelligent Claims EngineClaims processing with fraud detection at intakeOperations IntelligenceSame daypayment for clean claimsWellness Rewards SystemPreventive incentives priced by outcomeDecision & Risk Sciences$3,200annual savings per member
Risk, markets and reporting
Anomalies caught, distributions fitted, reports built once.
Transaction Anomaly MonitorAnomaly detection across millions of daily transactionsDecision & Risk SciencesReplayableevery score, with its inputsModel Calibration DeskAutomated distribution fitting and VARDecision & Risk SciencesWeeklycalibration, from quarterlyRegulatory Reporting LineRegulatory reports from governed dataEnterprise Data FoundationsOncebuilt, then re-run
What we start from
The data an insurer already governs.
Every model on this page is trained and calibrated on systems that are already under control. We read them, document the lineage, and agree the baseline with the CFO and the chief actuary.
- Policy administrationPolicies, coverages, endorsements and premiums, per product and cohort.Pricing
- Claims systemFirst notice of loss, adjuster notes, reserves, payments and closure dates.Claims
- Transaction logsPayment and account movements at the volume the anomaly models need.Fraud
- Model inventoryVersions, calibration dates, validation reports and the data each model was fitted on.Calibration
- Regulatory reporting packsSolvency II QRTs, IFRS 17 disclosures or FINMA returns, with their source tables.Reporting
- Member and counterparty master dataOne record per policyholder, provider and counterparty across systems.Master data
95%of generative AI pilots show no measurable P&L impact5
A model that cannot be replayed is not a control.
Regulators and auditors ask the same question: can you show how this score was produced, and would it be produced again? Fixed decoding, versioned data and logged runs turn a model into evidence. That is what makes it usable at the claims desk and in the capital calculation.
MIT NANDA, The GenAI Divide, 2025
How it starts
What the first quarter looks like at an insurer.
Thirteen weeks, from measuring the drift to a run-rate on the operations board.
01Weeks 1–3
Measure the drift and the days
We take the claims log, the model's back-tests and the fraud cases and measure what is leaking: days in queue, calibration drift between reports, fraud found after payment.
You get: one page per queue and per model
02Weeks 4–8
Pick the queue and the lever
We start on one queue, usually intake validation. Eligibility and fraud scoring at intake, clean claims paid the same day, every score logged with its model version. Our engineers sit with claims and risk.
You get: the first week of same-day settlements
03Weeks 9–13
Run the desk from it
The model calibrates weekly, the regulator pack builds itself from governed data, the reviewer queue carries the evidence. The number goes on the operations board.
You get: a run-rate the CRO can read
Practices and systems in this sector
The practices and systems behind the work.
Decision & Risk Sciences
Risk, fraud and calibration with the reasoning replayable.
AI GovernanceModels registered, outputs deterministic where a decision depends on them, every score logged.
Master Data Management LayerPolicy, claim, customer and counterparty records kept identical across core systems.
ShieldOSDecision audit trails and evidence packs for the regulator.
Talk to us about one model or one claims queue.
Tell us which model you trust least or which queue takes longest. We will come back with the drift we see and the days we think can come out of the queue.
Notes and sources
- 1Illustrative. Anonymised engagement scenarios; figures indicate modeled or targeted impact and are not audited outcomes.
- 2National Health Care Anti-Fraud Association, testimony to the U.S. Senate HSGAC subcommittee, October 2011: financial losses from health care fraud estimated in the tens of billions of dollars a year. Source
- 3RAND Corporation, The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed, RRA2680-1, August 2024. More than 80% of AI projects fail, twice the rate of IT projects without AI; 65 experienced practitioners interviewed. Source
- 4Gartner, August 2024: at least 30% of generative AI projects will be abandoned after proof of concept by the end of 2025, citing data quality, risk controls, cost and unclear business value. Source
- 5MIT NANDA, The GenAI Divide: State of AI in Business 2025, August 2025. About 95% of generative AI pilots show no measurable P&L impact; purchased or partnered deployments succeed about 67% of the time, internal builds about a third as often. Source