First Notice of Loss is the most process-heavy, highest-volume call type in any insurance operation. A motor carrier handling £500m of premium will field tens of thousands of FNOL calls a year — each one a 15 to 45 minute structured interview with a stressed customer, captured by a contact-centre agent into a claims system that has not changed materially since 2010. The Association of British Insurers reported a record £6.1bn in UK property claims paid in 2025, driven by weather and inflation. Volume is rising. Cost per claim is rising. Cycle time is the FCA's new battleground.
The commercial logic for AI voice in FNOL intake is now hard to argue with. Manual intake runs 15 to 45 minutes per claim with 15 to 25% data-entry error rates. Voice AI handles the equivalent capture in 5 to 6 minutes at roughly a tenth of the labour cost, and writes a clean, schema-conformant record straight into the claims system. This guide is the practical playbook for getting that working — at carrier scale, under FCA Consumer Duty scrutiny, without the hallucinations that get you on the front page of the FT.
This guide is shipped by the team behind Dilr Voice — enterprise voice AI deployed in regulated UK call centres, including FNOL intake. Or see our approach to enterprise AI deployment, the framework we use before any production deployment in regulated industries.
FNOL is the only insurance call type where the structure of the conversation is already mandated by the claims schema. That makes it the safest, highest-ROI workplace for voice AI in any carrier — provided the design respects FCA Consumer Duty, ICO data minimisation, and the limits of what an LLM should be allowed to decide.
The wider AI context matters here. Per McKinsey's State of AI 2025, 88% of enterprises now use AI but only 6% are AI-mature and only 14% report material EBIT impact. Insurance is over-represented in the laggard cohort. The carriers that pull ahead in 2026 will be the ones who pick a workflow with structured I/O, contained risk, and a measurable economic anchor — and ship it. FNOL is that workflow.
Why FNOL is the right beachhead
If you ask a claims COO where their cost goes, the answer is rarely "the adjuster". It is "the four to six handoffs between the customer ringing in and the adjuster getting a clean file". FNOL is the front of that chain. Every minute saved at intake compounds: faster file completion means faster triage, faster reserving, faster settlement, lower cycle time, lower complaint volume, better Consumer Duty outcomes data.
The structured-conversation advantage
Unlike a sales discovery call or a customer-success conversation, an FNOL call has a fixed information schema. Carrier loss-notification forms have between 40 and 120 mandatory fields depending on line of business. Motor: date, time, location, vehicles, drivers, third parties, injury, witnesses, police reference, photos, recovery preference. Property: peril, address, room-by-room damage, occupancy, prior claims, alternative accommodation. Liability: incident, claimant, alleged injury, witnesses, CCTV.
Voice AI excels precisely where the conversation has a known schema and a known set of branching paths. The same dynamic shows up in adjacent verticals — see AI voice for fintech collections and KYC for the parallel pattern in financial services, where structured data capture under regulatory scrutiny drives the same ROI math.
What human agents are actually bad at. Human FNOL agents are excellent at empathy and edge cases. They are unreliable at: capturing every mandatory field on the first pass, spelling postcodes correctly, time-stamping incidents accurately, asking the awkward fraud-screening questions consistently, and writing structured notes during a live call. The 15 to 25% data-entry error rate at manual FNOL is not a training problem — it is a human-attention problem. Voice AI does not get tired at 4pm on a Friday.
The FCA Consumer Duty angle
The FCA's Consumer Duty framework, sharpened further after the December 2025 response to the Which? super-complaint, requires firms to evidence consistent, well-reasoned claims outcomes from FNOL through to settlement. Consistency is exactly what AI voice delivers — every caller gets the same questions in the same order with the same screening, with a full transcript and audit trail attached. That is a Consumer Duty story you can tell a regulator. A team of 80 contact-centre agents on rotating shifts is not.
The same diagnostic logic underpins our AI operating model consulting for regulated insurers — used before any deployment commitment, designed specifically to identify which call types in your operation will yield the cleanest economics under regulator scrutiny.
The three deployment patterns — and which one to pick
Carriers ask for "an AI agent for FNOL" as if it is one thing. It is not. There are three distinct operating models, with very different economics and risk profiles. Most procurement disasters come from picking the wrong one for your line of business.
| Operating model | Cost per FNOL | Time to FNOL complete | Data-quality vs human | Cycle-time impact | Where it fits |
|---|---|---|---|---|---|
| Human-only (today) | £18–£28 | 15–45 min | Baseline (15–25% error rate) | Baseline | Complex commercial, BI, casualty |
| AI triage + human capture | £8–£14 | 8–18 min | 30–40% fewer errors | 10–15% faster | Mixed motor/home book, Consumer Duty hybrid |
| Full AI capture (STP eligible) | £1.50–£3 | 5–8 min | 50–70% fewer errors | 60–75% faster | Standard motor, simple home, glass, recovery |
The trap is to deploy "full AI" across the entire book. The right answer is segmentation: identify the 30 to 50% of FNOL calls that are simple, low-severity, and schema-pure, route those to full-AI capture, and leave the bodily-injury motor and complex commercial cases to human agents augmented by AI triage. That is how the carriers in production are actually getting 30 to 50% straight-through processing — by being honest about what AI should and should not own.
The economics also depend on what you measure. Carriers that price voice AI on cost-per-call alone tend to over-pay for capability they will never use. The work in voice AI total cost of ownership covers the seven hidden cost lines — telephony, integration build, governance, QA, change management, fail-over capacity, and post-go-live tuning — that turn an attractive headline price into a £400k programme.
The hallucination question
Every claims director eventually asks: "What stops the AI from telling a customer their roof is covered when it isn't?" The answer is architectural, not aspirational. Production-grade FNOL voice AI does not freestyle on coverage. The voice agent captures the loss; a deterministic policy-lookup service returns coverage; the agent reads back what the system says, never what it inferred. Hallucination risk is contained at the system boundary, not the prompt. We treat this as a procurement gate — see voice AI hallucination as a procurement gate for the full screening criteria we apply on regulated deployments.
The integration that actually matters. Voice AI in FNOL is only as good as its write into the claims system. The big four UK claims platforms — Guidewire ClaimCenter, Duck Creek, Sapiens IDIT, and the assorted PAS-of-record stacks — each have a different integration shape. The carriers winning here are the ones who treat the voice agent as a structured-data producer, not a transcript producer, and invest in the schema-mapping layer that writes 40+ fields cleanly the first time. Lift-and-shift transcript-to-CRM is the pattern that fails.
What 90 days of production looks like
The carriers who pull this off follow a similar deployment shape. Weeks one to four: schema-mapping, scope segmentation, shadow-mode running on a recorded sample of past FNOL calls. Weeks five to eight: live pilot on one line of business, single language, narrow loss types — typically motor own-damage and glass-only claims. Weeks nine to twelve: expansion to mixed motor + home, second language, integration into adjuster routing.
By week twelve a well-run programme is handling 25 to 40% of FNOL volume autonomously, cutting average intake from 22 minutes to 6, reducing cost per call by 60 to 70%, and flagging a meaningfully cleaner file to adjusters. The same carriers report Consumer Duty audit time falling sharply, because the AI calls produce structured, queryable, consistent records — not 80 hours of unstructured contact-centre notes per week.
If you want to see how this looks in production for a UK carrier, you can take three routes. Try Dilr Voice live for an FNOL workflow, book an AI placement diagnostic, see our DATS methodology for regulated deployments, or read about our approach to placing AI inside enterprise claims systems.
The two things that determine whether this works: discipline at scope (don't try to do everything on day one) and a procurement process that interrogates how the vendor handles coverage decisions, hallucination, and fail-over. Carriers that get this right are pulling 12 to 18 month payback. The ones who skip the diagnostic phase are still in pilot purgatory eighteen months later, which we cover separately in our work on AI voice deployment risk patterns.
Take FNOL from 22 minutes to 6.
30-min scoping call · No deck · Confidential. We'll tell you which lines of your book are ready for full-AI FNOL — and which should stay hybrid.
Written by the Dilr.ai engineering team — practitioners who ship enterprise AI in production. Follow us on LinkedIn for shipping notes, or subscribe via the RSS feed.