AI-supported invoice checking (CSS)
by CSS Versicherung · Primary insurer
Market unit: CH
CSS uses AI algorithms and AI-based review logic in its benefit-entitlement checking of invoices from basic and supplementary health insurance. The AI prevents duplicate payments for invoices submitted twice, assesses accident-related claims and flags unusually high billing items or quantities; separate data analytics retrospectively identify providers with conspicuously high costs. CSS checked 23.4 million invoices in 2023 and states that over 85% are processed automatically. A data-driven process control introduced in early 2024 removed about 190,000 manual checks within a few months. CSS attributes avoided benefit costs of around CHF 800 million per year partly to AI; in 2025 it checked 25.4 million invoices and prevented about CHF 872 million of unjustified spending (all checking, not AI alone).
1 free view left this month
AI capability and maturity
AI capability level
Capability Maturity Score (CMS)
5 / 6
Business Adoption Score (BAS)
4 / 6
AI Development Index (ADI)
20 / 36
Business context
- Purpose of use
Automate benefit-entitlement checks of health invoices and flag duplicates and anomalies.
- Value chain steps
- Claims managementFraud detection and prevention
- Insurance branches
- Supplementary health insurance
- Application scope
- Internal only
- Business criticality
- Mission critical
- Country or market unit
- CH
Value and impact
- Added value category
- Cost reductionEfficiency gainQuality improvement
- Added value estimation
About CHF 800 million benefit costs avoided per year, partly through AI (CSS)
- Time to value (months)
- 10
- Competitive advantage gain
Domain-tuned extraction for Swiss medical tariff systems (TARMED, TARDOC).
- Quality gain
About 190,000 manual checks eliminated within a few months (early 2024)
Technical details
- AI category
- Machine learning and predictive analyticsAnomaly detection
- Interfaces used
Claims intake queue
- Architectural insights
Event-driven extraction with confidence-based human review thresholds.
- Automation level
- 4/5
- Estimated IT difficulty
- 4/5
Compliance and governance
- EU AI Act risk category
- Not available
- EU AI Act role
- Not applicable
- Relevant policies
- Swiss Federal Act on Data ProtectionFINMA guidelines
- Responsibilities
Health claims operations owns exception handling and audit sampling.
- Data used
Medical invoices, tariff catalogues, policy coverage rules.
Metadata
- Development type
- Scaled application
Publications over time
When this solution was reported on. Rumors are unverified mentions.
Community
Average rating
Comments
Log in to join the discussion.
No comments yet. Be the first to comment.
Related solutions
ZFinder is a fraud-detection model developed in-house by Zurich Seguros (Spain) that combines big data and machine learning, comparing each claim with thousands of similar claims and improving with every analysis. Zurich applied it first to home insurance (announced June 2018), with plans to extend it to other lines under the supervision of its internal investigation team. Zurich reported that automatic fraud detection doubled after launch and that in 76% of claims investigated via ZFinder a loss adjuster's work was not essential. The most common fraud types it found were simulated claims (47%), claims for excluded cover (22%) and concealed pre-existing damage (17%). No newer figures were found.
Since September 2021 Baloise has used a virtual voice assistant, developed with Spitch and NTT DATA, in its Swiss non-life customer service. Callers state their concern in standard German or Swiss German dialect instead of using the phone menu; the assistant recognises the intent, mainly claims notifications, and transfers the call with an on-screen text summary for the agent. At launch about 40% of German-language non-life enquiries passed through the assistant and 85% of callers reached the right expert at the first attempt; French was planned for a later stage. Spitch's case study (updated April 2025) describes a hybrid on-premise and cloud setup. Status after the December 2025 Helvetia Baloise merger is not documented.
by Zurich Insurance Group (Zurich Switzerland, Zurich Brazil)
CH, BR
Zurich's in-house claims fraud-detection system was developed by ZCAM (Zurich Customer Active Management), its wholly owned data-analytics company with about 50 specialists. In Switzerland, most of the more than 600,000 claims received each year pass through a proprietary fraud-scoring tool fed by the Allegro claims system and receive a 0-100 risk score, with high scores prompting closer review by the Special Investigation Unit. Alongside the rule-based system, a machine-learning model has been piloted in German since February 2024 for property and commercial lines. Zurich Brazil adopted ZCAM's system in its affinity business (extended warranties, mobile-phone theft) and extended it to motor claims; the affinity fraud detection rate trebled (Zurich, September 2024).