First Notice of Loss is the moment that defines an insurer's relationship with a claimant. The caller is stressed — often distressed. They've just had an accident. Their car is damaged. Maybe they're injured. Maybe someone else is. They pick up the phone and reach a contact centre.

What happens in the next ten minutes shapes everything: the accuracy of the claim, the speed of assessment, the claimant's perception of their insurer, and the insurer's exposure to fraud. And in most Dutch non-life insurers today, those ten minutes look like this: an agent navigating four disconnected systems simultaneously while keeping the caller engaged and providing accurate information.

It is an extraordinarily difficult task to do well under volume. Most of the time, it isn't done well.

The real numbers

50–70
FNOL calls per agent per day — with 4 systems, manual data entry, and a live stressed caller on every one
3–5×
Cost of rework caused by data entry errors made under time pressure — compounding downstream
68%
Customers who say the FNOL experience shapes their overall satisfaction with their insurer

Every piece of data entered manually, under time pressure, in a live conversation with a stressed human being. The result: data entry errors that create rework costing three to five times the original call. Fraud checks skipped — not by negligence, but by arithmetic. When you handle sixty calls today, the tab-switch to the fraud system takes time that doesn't exist.

"An incorrect claim file delays assessment. A missed fraud flag creates exposure. A claimant who had to call back is a claimant who tells other people about their insurer."

The compounding cost of FNOL errors

Why this is the hardest test for AI

FNOL is not a chatbot problem. It is a multi-system, multi-step, compliance-critical, real-time execution problem. Here's why generic AI fails it:

🔀
IVR and voicebots break on deviation
Real FNOL calls are non-linear. Callers interrupt themselves, add information out of order, correct earlier statements. A rigid script creates friction at the worst possible moment.
🎲
Generic LLMs are probabilistic — FNOL is not
FNOL requires determinism: the fraud check must run, claim fields must be validated, step order is legally significant. An LLM cannot guarantee any of this without a constraint layer it doesn't have.
👤
Agent-assist tools don't remove the bottleneck
CCaaS platforms help agents work faster. They do not remove the agent from the loop. The bottleneck is not the agent's speed — it's that a human must be in every loop at all.
🤖
RPA and workflow automation breaks on live voice
A caller who changes their account mid-call, adds a third party, or corrects a policy number — all normal FNOL behaviour. Static automation cannot handle it.
SKYLA FNOL — 5 autonomous steps, one call
SKYLA FNOL 5-step autonomous workflow 🪪 Identity Policy verified real-time No hold required 🗂 Accident Details Natural conversation Handles non-linear input 🔍 Fraud Check Mandatory — unskippable OPA/Rego enforced CANNOT BE SKIPPED 📋 Claim File Atomic write · Schema valid Guidewire / Duck Creek 📅 Assessor Booked Confirmed in same call SMS + voice confirmation ✓ End-to-end: under 200ms per step · GER audit record on every action · Runs on existing infrastructure · No rip-and-replace

What SKYLA does instead

SKYLA handles the complete FNOL intake autonomously. Not agent-assist. Autonomous. Five steps, one call, before the caller hangs up.

1
Identity and policy verification in real time
The caller continues speaking — no hold required. SKYLA verifies identity via voice and policy lookup simultaneously. Result confirmed before moving to step two.
No hold · Real-time
2
Accident detail extraction from natural conversation
The caller speaks normally. SKYLA handles interruptions, corrections, and non-linear disclosures natively. Every field is schema-validated before it's written to any system.
Natural voice · Schema validation
3
Fraud check — mandatory and unskippable
The ARF policy layer enforces this architecturally. It cannot be bypassed regardless of call volume or time pressure. This is the step that proves the architecture.
⚠ Mandatory · Cannot be skipped
4
Claim file creation — atomic, zero re-entry
Every field from steps 1–3 carried forward. Atomic write. Claim ID confirmed to caller live. Runs on existing Guidewire or Duck Creek infrastructure.
Zero re-entry · Existing systems
5
Assessor appointment booked and confirmed
In the same call. No callback. No follow-up. The intake is complete before the caller hangs up. SMS confirmation dispatched automatically.
Complete in one call · SMS confirmation

What passing this test means

If an agentic AI system can handle FNOL — with its regulatory constraints, its emotional complexity, its multi-system integration requirements, its mandatory compliance steps — it can handle almost any enterprise workflow.

That's why we started here. Not because insurance is our only market. Because insurance proves the architecture works under real conditions. The same ARF layer that handles FNOL handles healthcare back-office workflows, telco CX, and any high-volume process where the workflow is rules-driven and the execution is currently manual.

"Two to four weeks from contract to live. No rip-and-replace. The same Guidewire or Duck Creek infrastructure you already run — SKYLA integrates over REST/OAuth2."

SKYLA deployment model

The live demo is available now. SKYLA will call your phone and walk you through a complete motor claims intake — the same five steps, the same infrastructure as production. Not a simulation.