NLP Development in Stockholm, Sweden

Insurance (Sakförsäkring)

A Nordic P&C insurer (headquartered in Stockholm, operating across Sweden, Norway, and Denmark, 3.8M policies, 280,000 claims annually, SEK 32B gross written premium) processed claims through 195 claims handlers across three countries. The claims narrative processing challenge: (1) Claims descriptions were unstructured. Customers described loss events in free text on the insurer's app and website — in Swedish, Norwegian, and Danish. Each description needed to be read, understood, classified, and converted into structured claims data (claim type, cause of loss, affected property, estimated severity). (2) Language variation created inconsistency. Swedish claims handlers processed Swedish text fluently but struggled with Norwegian and Danish descriptions (despite mutual intelligibility, insurance terminology differs: Swedish "självrisk" = Norwegian "egenandel" = Danish "selvrisiko" for deductible). Norwegian claims handlers had similar issues with Swedish and Danish text. (3) Claims classification was manual. Handlers classified each claim into one of 85 categories (brand — fire, vattenskada — water damage, stöld — theft, trafikskada — traffic damage, personskada — personal injury, and 80 sub-categories). Classification accuracy: 81% (measured by quality audit). Misclassification affected: reserve accuracy (wrong claim category = wrong reserve model), regulatory reporting (claim statistics reported to Finansinspektionen by category), and fraud detection (fraud patterns are category-specific). (4) Severity estimation was slow. Initial severity estimation (determining whether a claim is likely under SEK 25,000 — simple, SEK 25,000-250,000 — standard, or over SEK 250,000 — complex) determined the handling pathway. Manual severity estimation from narrative text took 8-12 minutes per claim. Incorrect severity routing caused: simple claims assigned to senior handlers (wasted expertise), complex claims assigned to junior handlers (inadequate investigation). (5) Fraud indicators were inconsistently detected. Narrative text contained fraud indicators — inconsistencies between the described event and physical evidence, unusually specific damage descriptions (suggesting pre-knowledge), and language patterns associated with fabricated narratives. Detection depended entirely on individual handler experience and attention.

We built a Nordic claims NLP platform over 11 weeks. (1) Cross-Nordic Claims Processing Engine: processing claims descriptions in Swedish, Norwegian, and Danish. The engine: detecting language automatically and processing each in its native language model (not translating to Swedish then processing — native processing preserving semantic nuance in each language), extracting structured claims data from narrative descriptions: claim type (mapping narrative descriptions to 85 claim categories — "vattnet sprutade ut från röret under diskbänken" → vattenskada/vattenledningsskada — water damage/pipe damage), cause of loss (rörbrott — pipe burst, inbrott — break-in, trafikolycka — traffic accident — extracted from narrative context), affected property (specific rooms, items, vehicle parts — extracted from description), severity indicators (damage extent, affected area, number of affected items, cost mentions), and timeline (when the damage occurred, when it was discovered — temporal extraction from Swedish, Norwegian, and Danish text). Processing time: 15 seconds per claim (from 8-12 minutes manual reading and classification). Classification accuracy: 93.8% across all 85 categories (from 81% manual). (2) Severity Intelligence Engine: estimating claim severity from narrative and extracted data. The engine: analysing damage descriptions for severity indicators ("hela golvet förstört" — entire floor destroyed = high severity, "liten repa" — small scratch = low severity), comparing against historical claims with similar descriptions (semantic similarity matching against 3 years of closed claims with known ultimate costs), incorporating property and policy data (home value, vehicle value, policy coverage limits), and producing severity classification (simple/standard/complex) with confidence scores. Severity routing accuracy: 91.2% (from 76% manual). Correctly routing complex claims to experienced handlers improved investigation quality and reduced ultimate claim costs. (3) Fraud Detection NLP Engine: identifying linguistic fraud indicators in claims narratives. The engine: detecting narrative inconsistencies (contradictions between damage description and timeline, physically implausible damage combinations), identifying coached language patterns (unusually formal or legalistic language in casual claims descriptions — suggesting the claimant received guidance on how to describe the event), comparing narrative similarity across claims (detecting multiple claims with suspiciously similar descriptions — suggesting coordinated fraud), analysing sentiment patterns (genuine claims tend to express frustration, confusion, or distress; fabricated claims tend to be emotionally flat and overly factual), and flagging claims for SIU (Special Investigation Unit) review with specific fraud indicators. Fraud detection: flagged 2.3% of claims for SIU review — of which 68% confirmed fraudulent or exaggerated (vs. previous manual detection rate of 1.1% flagged with 45% confirmation). (4) Cross-Nordic Analytics Engine: providing unified analytics across Swedish, Norwegian, and Danish claims. The engine: normalising claims data across languages into a unified taxonomy (mapping Swedish "vattenskada" = Norwegian "vannskade" = Danish "vandskade"), generating cross-Nordic comparison reports (claim frequency, severity, and trends across countries), identifying country-specific patterns (claim types more prevalent in specific countries — enabling targeted prevention programs), and producing Finansinspektionen/Finanstilsynet regulatory reports in country-appropriate formats.

Claims classification accuracy: 93.8% (from 81%). Misclassification-related issues (incorrect reserves, inaccurate regulatory reporting) reduced proportionally. Processing time: 15 seconds per claim intake (from 8-12 minutes). Claims handlers' time freed from reading and classifying descriptions redirected to investigation and customer interaction. Severity routing accuracy: 91.2% (from 76%). Complex claims correctly identified and routed to experienced handlers — resulting in improved investigation quality. Estimated reduction in ultimate claim costs from better routing: SEK 18M annually. Fraud detection: flagging rate increased from 1.1% to 2.3%. Confirmed fraud/exaggeration rate: 68% (from 45%). Net fraud savings: SEK 42M annually (identified fraudulent and exaggerated claims prevented or mitigated). Cross-Nordic consistency: unified claims taxonomy across Swedish, Norwegian, and Danish — enabling portfolio-level analytics previously impossible due to language-siloed data. Claims team: 195 handlers maintained (growing claims volume absorbed without additional hiring — effectively 35 additional FTE capacity from processing efficiency). Finansinspektionen compliance: improved regulatory reporting accuracy. Annual value: SEK 42M (fraud detection), SEK 18M (severity routing optimisation), SEK 28M (processing efficiency — capacity increase without hiring), SEK 8M (classification accuracy — improved regulatory reporting, better reserve setting) = SEK 96M (approximately €8.3M). Development cost: SEK 2.6M (approximately €224K). Payback: under 10 days.