Chatbot Development in Helsinki, Finland

Insurance / Customer Service / Finnish NLP

A Finnish insurance company (€1.2B premium income — 1.8 million customers, offering home, vehicle, travel, and health insurance, 2,400 employees, headquarters in Helsinki) needed a Finnish-language chatbot for customer service. (1) Contact volume: 1.4 million customer contacts annually — 680,000 phone calls, 420,000 digital messages (email, chat, social), and 300,000 self-service attempts that failed and escalated. Agent count: 320 customer service agents. Cost: €18.5M annually in customer service operations. Peak: claims season (winter ice damage, summer storm damage) creating 50-60% volume spikes requiring temporary staff. (2) Finnish language barrier: existing chatbot (deployed 2021) being keyword-based — failing on Finnish morphology. Example failures: customer writing "autovakuutuksestani" (about my car insurance) — chatbot not recognising the word because it knew "autovakuutus" but not its elative case form. Customer writing "kotivakuutus" (home insurance) — chatbot understanding this but not "kodinvakuutus" (a common alternative compound). Abandonment: 67% of customers abandoning the existing chatbot within 2 messages — "I'd rather call, this bot is useless" being representative feedback. Reputation: the existing chatbot actively damaging brand perception — Finnish social media posts mocking the chatbot's inability to understand Finnish. (3) Routine queries: 55% of contacts being routine — "Am I covered for X?", "How do I file a claim?", "What's my deductible?", "When does my policy renew?", and "Can I add my new car?" Resolution: routine queries averaging 7 minutes per contact — agents looking up policy details and explaining coverage. Opportunity: 55% of 1.4M = 770,000 routine contacts × 7 minutes = 89,800 hours/year on queries AI could handle. (4) Claims: claims process being the most critical customer journey — and the most painful. Current: customer calls to report claim → agent takes details → claim form generated → customer receives form for signature → documents collected → claim assessed. Time: claims initiation averaging 15 minutes by phone — agent asking questions, customer describing situation, and data entry. Drop-off: 8% of customers not completing claims process — giving up due to friction. Claims volume: 280,000 claims annually — each requiring initiation, document collection, assessment, and resolution. (5) After-hours: 35% of digital contacts occurring outside business hours (8 AM-8 PM) — no response until next morning. Finnish: Finns being active digital users in the evening — insurance questions arising when reviewing finances at home. Claims: property damage often discovered in the evening — customer wanting to initiate claim process immediately.

We built a Finnish-language AI chatbot deployed across web, app, and Facebook Messenger over 7 weeks. (1) Finnish NLP engine: genuine Finnish language understanding. LLM: GPT-4 with Finnish-specific system prompting — understanding Finnish morphology, compounds, and informal language. Finnish RAG: 45,000 Finnish-language knowledge documents — policy wordings, coverage tables, claims procedures, FAQ content, and regulatory information. Morphological handling: Finnish lemmatization preprocessing — reducing inflected forms to base forms before RAG retrieval, then understanding full forms in conversation context. Example: "autovakuutuksestani" → understands as "auto+vakuutus" in elative case with possessive → retrieves car insurance coverage information → responds using appropriate Finnish grammar. Compounds: Finnish compound word understanding — "kotivakuutus," "kodinvakuutus," "asuntovakuutus" all recognised as home insurance concepts. Informal: handling informal Finnish — "mul on sellane tilanne et" (I've got this situation that) → understanding this as a claim or coverage question preamble. Swedish: Swedish language support — detecting Swedish input and responding in Swedish using the same knowledge base. English: English fallback for international customers — seamless language switching. (2) Coverage assistant: answering "Am I covered?" in Finnish. Personalised: chatbot accessing customer's specific policy (after authentication) — answering based on their actual coverage, not generic information. Example: "Kattaako mun vakuutus jos putki jäätyy?" (Does my insurance cover if a pipe freezes?) → checking customer's home insurance → finding water damage coverage details → responding with specific coverage, deductible, and what to do. Comparison: "Mitä eroa on laajalla ja superlajalla kotivakuutuksella?" (What's the difference between comprehensive and super-comprehensive home insurance?) → detailed comparison of their specific product options. Finnish clarity: responses in clear, natural Finnish — not legalese, not translated English, but genuine Finnish explanation. (3) Claims initiation: starting claims in chat. Guided: chatbot guiding customer through claims initiation — "Kerro mitä tapahtui" (Tell me what happened) → natural conversation collecting incident details. FNOL: First Notice of Loss captured through conversation — date, time, location, description, damage, and third-party involvement. Photos: customer uploading damage photos through chat — AI initial assessment (vehicle damage severity, property damage type). Smart: chatbot asking relevant follow-up questions based on claim type — vehicle claims asking about other driver details and police report, property claims asking about temporary repairs and living arrangements. Submission: claim submitted directly from chat — customer receiving claim number and next steps. Time: claims initiation from 15 minutes (phone) to 8 minutes (chatbot) — and available 24/7. (4) Policy management: self-service through chat. Changes: "Ostin uuden auton, pitäis vaihtaa vakuutukseen" (Bought a new car, need to change insurance) → chatbot guiding through vehicle change process — registration number, vehicle details, coverage selection, and premium calculation. Renewals: "Milloin mun vakuutus uusiutuu ja paljon se maksaa?" (When does my insurance renew and how much?) → checking renewal date and premium → offering renewal confirmation or change options. Documents: "Tarviin vakuutuskirjan" (I need my insurance certificate) → generating and sending policy documents through chat. Payment: "Onks mun maksu menny?" (Has my payment gone through?) → checking payment status and providing information. (5) Agent handoff: seamless human escalation. Recognition: chatbot recognising when to escalate — complaints, complex claims, emotional customers, and queries outside its scope. Context: complete conversation history transferred to human agent — agent seeing what was discussed, what the customer needs, and what the chatbot has already checked. Finnish: agent receiving context in Finnish — consistent language throughout the experience. Queue: when escalating, chatbot providing estimated wait time and offering callback option — respecting customer's time.

Results over 10 months. Finnish accuracy: Finnish language understanding at 91% (from 34% with previous keyword chatbot — measuring correctly understood customer intent). Morphology: 94% of Finnish morphological forms correctly handled — compound words, all 15 cases, and possessive suffixes. Informal: 88% accuracy on informal Finnish — slang, abbreviations, and Finnish-English code-switching. Swedish: 86% accuracy on Swedish interactions — acceptable for Finland's Swedish-speaking minority. Resolution: chatbot resolving 58% of customer enquiries without human escalation (target: 55%). Interactions: 812,000 AI-resolved interactions annually (from 1.4M total contacts). Response time: instant (from 7 minutes average phone, 45 minutes average email). After-hours: 35% of chatbot interactions outside business hours — 284,000 interactions previously unserved. Coverage questions: 72% of "Am I covered?" questions resolved by chatbot — personalised, policy-specific answers. Customer satisfaction: chatbot CSAT from 1.8/5 (previous keyword chatbot) to 4.2/5 (AI chatbot). Finnish feedback: "Vihdoinkin botti joka ymmärtää suomea" (Finally a bot that understands Finnish) — most common positive feedback. NPS: company NPS improved from 28 to 41 — chatbot being one of several contributing factors but specifically cited in customer feedback. Claims: 42% of claims initiated through chatbot (from 0%). Claims initiation: from 15 minutes (phone) to 8 minutes (chat) — 47% faster. 24/7 claims: 28% of chatbot-initiated claims filed outside business hours — previously impossible. Claims completion: chatbot-initiated claims having 97% completion rate versus 92% for phone-initiated — guided process reducing drop-off. Staff: customer service agents reduced from 320 to 240 (80 redeployed to complex claims, underwriting support, and customer retention — higher-value roles). Agent quality: agents handling complex, interesting cases instead of routine queries — agent satisfaction improved from 3.1/5 to 4.0/5. Seasonal: winter/summer claims peaks handled without temporary staff — chatbot absorbing volume increase. Temporary staff cost eliminated: €480,000/year. Financial: annual savings — agent redeployment €4.8M (80 agents), temporary staff €480K, after-hours service €320K (estimated value), and claims efficiency €580K (faster initiation, higher completion). Total annual value: €6.18M. Development: €320,000 (7-week project). Annual operation: €95,000 (LLM API, hosting, Finnish knowledge base maintenance). ROI: 19.3x first-year return.