AI Voice Agent Development in London

Insurance / UK Composite Insurer

A UK composite insurer (GBP 8.5B GWP, 15 million policyholders across motor, home, pet, travel, and commercial lines, 4,200 employees including 1,800 in claims operations) faced claims notification and customer service challenges. (1) Claims call volume: the insurer received 3.2 million claims-related calls annually — first notification of loss (FNOL), claims status enquiries, documentation submission questions, and settlement discussions. 1,200 claims handlers and customer service agents processed these calls. Average handle time for FNOL: 14 minutes. Peak periods (winter storms, major weather events) overwhelmed capacity — during Storm Ciarán (2023), inbound calls tripled to 45,000/day from a normal 12,000/day, creating 90-minute hold times and significant customer dissatisfaction. (2) After-hours claims reporting: the claims line operated 7 AM to 10 PM — but motor accidents and home emergencies occurred 24/7. An outsourced after-hours service handled overnight calls — but outsourced agents lacked system access, could only take basic details, and could not arrange emergency services (emergency glaziers, plumbers, alternative accommodation). 28% of motor claims and 35% of home claims were reported after hours — and the quality of after-hours FNOL data was significantly lower than daytime (missing fields, incomplete descriptions, wrong policy numbers). (3) FNOL data quality: human agents collecting FNOL data varied in thoroughness. Some agents captured comprehensive claim details; others rushed through data collection during busy periods. Incomplete FNOL data caused downstream inefficiency: claims handlers spent an average of 45 minutes per claim re-contacting customers to collect missing information. With 1.1 million new claims per year, even a small improvement in FNOL completeness saved substantial claims handler time. (4) Emergency service coordination: for home claims involving emergencies (burst pipes, roof damage, break-ins), the insurer needed to arrange emergency services (approved contractors) within 2 hours of notification. During normal hours, agents arranged this during the call. After hours, emergency coordination was delayed until morning — causing additional damage (a burst pipe running overnight caused far more damage than one stopped within 2 hours). (5) Vulnerable customer identification: the FCA required insurers to identify and appropriately support vulnerable customers (the Consumer Duty). Human agents received training on vulnerability indicators, but consistent identification was difficult — agents under time pressure during peak periods were less likely to detect vulnerability signals (speech patterns, confusion, distress).

We built a claims notification voice agent over 12 weeks, integrated with the insurer's claims management and emergency service coordination systems. (1) Conversational FNOL collection: the voice agent conducted natural FNOL conversations — adapting its questions based on the claim type (motor: accident circumstances, third party details, injuries, vehicle damage; home: incident description, damage extent, emergency needs; travel: location, medical situation, documentation). For motor claims: "Can you tell me what happened?" followed by contextual follow-up questions that collected all required FNOL fields through natural conversation rather than rigid question scripts. Data completeness: 96% of required FNOL fields collected (from 78% with human agents during peak periods) — because the agent systematically covered all fields without being affected by time pressure or conversation fatigue. (2) 24/7 operation with full capability: the voice agent operated continuously with complete system access — policy lookup, claims creation, emergency service dispatch, and payment processing. After-hours capability matched daytime: customers reporting claims at 3 AM received the same quality of service as those calling at 10 AM. (3) Emergency service coordination: for emergency home claims (water damage, roof damage, security breach), the voice agent immediately: assessed the emergency severity based on the customer's description, dispatched the appropriate emergency contractor from the insurer's approved network (matching contractor availability, location, and speciality to the emergency type), provided the customer with contractor details and estimated arrival time, and followed up with the contractor to confirm dispatch. Emergency response time: from next-morning (after-hours claims) to average 85 minutes 24/7. (4) Vulnerability detection: the voice agent incorporated continuous vulnerability assessment throughout the conversation — monitoring for indicators of vulnerability (distress, confusion, hearing difficulty, cognitive impairment, financial distress indicators) using both speech analysis (tone, pace, hesitation patterns) and content analysis (statements indicating vulnerability). When vulnerability was detected, the agent adapted its communication style (slower pace, simpler language, more empathetic tone) and flagged the claim for priority human review by the insurer's vulnerable customer team. (5) Surge capacity: during catastrophic weather events, the voice agent scaled automatically — handling 45,000+ daily calls without degradation (from the human capacity of 12,000 calls/day). During Storm events: the agent handled initial FNOL, arranged emergency services, and queued non-urgent claims for daytime follow-up — ensuring that emergency needs were addressed immediately while managing the volume that would have overwhelmed human capacity.

Call handling: the voice agent handled 65% of claims calls without human intervention (2.08 million of 3.2 million annual calls). Remaining 35% transferred to human agents with full context. Hold times: eliminated for voice-agent-handled calls. Human agent queue: peak wait from 90 minutes (storm events) to under 8 minutes. FNOL completeness: from 78% (peak period human average) to 96% — eliminating 18 percentage points of missing data. Claims handler re-contact time: reduced from 45 minutes average to 12 minutes (only truly complex cases required follow-up). Across 1.1 million annual claims: 605,000 person-hours saved. After-hours service quality: full-capability claims service 24/7 — after-hours FNOL data quality now matched daytime levels (previously significantly lower). After-hours claims: resolved 28% faster than previously (emergency services dispatched immediately rather than waiting for morning). Emergency response: average 85 minutes response time 24/7 (from next-morning for after-hours claims). Reduced secondary damage: estimated GBP 18M in reduced claim costs from faster emergency response (particularly water damage — early intervention reduced average water damage claim from GBP 8,200 to GBP 5,400). Vulnerable customer identification: detection rate improved from 34% (human agents) to 72% (voice agent continuous monitoring) — ensuring appropriate support per FCA Consumer Duty requirements. Surge handling: weathered 3 storm events in the first year with zero capacity overflow — 45,000+ daily calls handled without service degradation. Previously, each storm event required emergency staffing costing GBP 180,000. Customer satisfaction: claims NPS improved from +12 to +34 — driven by elimination of hold times, 24/7 access, and faster emergency response. Annual value: GBP 18M (reduced claim costs from faster emergency response), GBP 14.5M (claims handler productivity — 605,000 hours freed from re-contact), GBP 8.2M (agent cost optimisation — voice agent handling 65% of volume), GBP 2.8M (surge event cost avoidance — 3 events x GBP 180K + avoided customer churn), GBP 1.5M (after-hours quality improvement — reduced claim leakage from better FNOL data) = GBP 45.0M. Development cost: GBP 340,000. Payback: under 3 days.