Insurance · AI engineering

AI applications built around how your insurance operation already works — not around a vendor's roadmap.

We design AI-assisted tools for claims, customer calls, policy documents, and reporting — based on your current systems, data sources, and compliance requirements. Custom engineering, delivered in phases your team can absorb.

Book a 30-min discovery call See what we build
GDPR-aware · EU data residency available · NDA on day one
Claims · call analytics · today live
1,284
calls · 24h
0.72
avg sentiment
18
flagged · review
Adjuster · cases resolved · last 30d
M. Reyes78%auto · water
K. Singh52%auto · fender
T. Okafor38%property · roof
L. Müller84%auto · glass
S. Park22%liability
› show me calls where claimant mentioned water damage in last 30 days NLQ · pgvector
n8n postgres pgvector gpt-4o rag
Outcomes we design for
Lower handling time

Routine claims handled with less adjuster touch and fewer hand-offs.

Searchable conversations

Call recordings turned into structured records you can query.

Faster reporting

Pipelines that produce regulator-ready data without weekend rework.

Phased rollout

Working prototype first, then production — not an 18-month bet.

What we hear in discovery

Most operational drag isn't about one system. It's the gaps between them.

The five patterns below show up in nearly every insurance discovery call we run. None of them require ripping and replacing a core system — most are handled with focused tooling on top of what you already have.

01

FNOL and claims intake are document-heavy

First-notice-of-loss arrives as forms, attachments, and voicemails. Staff spend hours triaging what should be routed automatically — and the queue still gets longer.

02

Customer calls hold information no one can use

Hours of recorded conversations contain complaints, churn signals, and recurring questions. Without transcription and summarisation, that record only matters when there's a dispute.

03

Fraud and leakage surface after the payout

Patterns adjusters notice in conversation often aren't reflected in case notes until it's too late to act. The signals exist; they're just not being captured or scored.

04

Policy documents stay locked in PDFs

Endorsements, schedules, and rider clauses are filed but rarely structured. When underwriters or adjusters need a clause, they read three pages of fine print.

05

Reporting depends on one person's weekend

Loss ratio, exposure, and IFRS 17 inputs run on bespoke SQL maintained by one analyst. New questions take a ticket, a week, and a meeting that doesn't get scheduled.

What we build

Custom AI applications for the parts of insurance that are still manual.

Each engagement starts with one high-value workflow, validated against your data. Production work is planned around integrations, security, and the team that will own it.

Claims triage

Read every claim as it arrives. Route the routine, hand the rest to a human.

An intake layer reads the FNOL, attached documents, and historical context, then scores the claim against your rules and underwriting position. Clean cases move to auto-acknowledge or auto-pay paths your team has approved. Borderline and high-value cases land in an adjuster queue with the model's reasoning attached.

  • Reads structured submissions, PDFs, and voicemail transcripts
  • Routing logic is visible — adjusters see why a claim landed where it did
  • Writes back to your existing claims system; no parallel system to maintain
claims · last hour
Received
28
Auto-routed
21
Queued for review
6
CLM-44291Auto · rear-endAuto-acknowledged
CLM-44292Property · waterAdjuster queue
CLM-44293Auto · windshieldAuto-acknowledged
CLM-44294Liability · slipAdjuster queue
CLM-44295Auto · total lossInvestigations
Call analytics

Customer calls become searchable records, not just compliance archives.

Every inbound call is transcribed, summarised, and tagged with topic, sentiment, and recurring complaint patterns. Team leads ask questions in plain English — "complaints about delayed payouts last week, by office" — and get a list of calls with quotes and one-click handoffs to a follow-up adjuster.

  • Transcripts mapped to existing call metadata (CRM, switch, IVR)
  • Sentiment and topic models tuned on your own data
  • Quote-level evidence on every search result
call analytics · search
complaints about delayed payouts, last 7 days
#CALL-22841 · 4m 12s

Caller escalated about a water-damage claim filed 18 days ago. Asked for a supervisor twice.

#CALL-22906 · 7m 03s

Customer mentioned ombudsman; cited two missed callbacks on a property claim.

#CALL-22987 · 2m 41s

Auto claim — promised payout date missed. Caller asked for written confirmation of next steps.

Showing 3 of 14 matching calls · all linked to original recordings and adjuster notes
Policy & document extraction

Turn endorsements, schedules, and rider clauses into queryable fields.

Policy documents are read by extraction pipelines that produce versioned, structured records — what's covered, what's excluded, what the limits are, what changed in the latest endorsement. Underwriters and adjusters query the structured layer; the original document is one click away for audit.

  • Document hashes and version history kept for audit
  • Human review on low-confidence fields before write-back
  • Outputs slot into rating, claims, and reporting systems
policy extraction · audit view
POL-2024-08412 · endorsement #3
Commercial property — extended cover
versioned
Sum insured
AUD 4,250,000
Excess
AUD 5,000 / event
Flood
Included (sub-limit AUD 250k)
Business interruption
12 months indemnity
Reviewed by
Underwriter — confidence 0.92
Source
endorsement-3.pdf · p.2
Reporting & ad-hoc queries

A natural-language layer on top of the warehouse, with guardrails.

Leadership asks questions of the data warehouse in English. A schema-aware query layer rewrites the question into safe SQL, runs it, and returns the answer with the columns it touched. No write access, no destructive operations, full audit log — designed for analyst review, not blind delegation.

  • Schema-aware generation, scoped to read-only views
  • Every answer cites tables and columns used
  • Sits alongside existing BI — replaces nothing your team trusts
reporting · natural-language
loss ratio by product line, NSW, last 90 days
generated SQL · read-only view
SELECT product_line,
  SUM(losses)/SUM(premium) AS loss_ratio
FROM v_claims_90d
WHERE state = 'NSW'
GROUP BY product_line
Auto
64.0%
Home
58.0%
Commercial
71.0%
Liability
48.0%
Not sure where to start?
Tell us where your workflow is stuck — we'll review it and suggest the most useful next step, whether that's a prototype, a review, or a referral.
Book a workflow review
How we work

From discovery to working prototype in weeks — production planned responsibly.

We start with one high-value workflow, validate feasibility, build a prototype, then plan production around integration, security, and the team that will own it.

01
Week 1

Discovery call

30 minutes. We learn the workflow, the systems already in play, and the compliance shape. The goal is to know whether we're a fit, not to pitch.

02
Weeks 2–4

Proof of concept

We design the application and build a working prototype against your real data. You see results, not slides — and we both find out where reality bites.

03
Weeks 5–12

Production build

Dedicated engineers ship the production version, integrated into your systems. Weekly demos, written status, and decisions captured as we go.

04
Ongoing

Controlled rollout & support

We deploy in phases, monitor usage, collect feedback, and improve the application before wider release. You own the code; we stay involved as much as you want.

Compliance

Built with security and audit requirements designed in, not bolted on.

We design against the controls auditors actually review — access, retention, lineage, and reversibility — and add a four-layer defense around AI outputs: input filtering, prompt constraints, schema validation, and least-privilege access to downstream systems.

Audit-aware controls

Role-based access, encryption at rest and in transit, audit logs, and deployment hygiene — designed against the controls auditors actually review.

EU data residency

Region-specific hosting, storage, access, and retention policies on AWS (eu-central) or bare-metal Hetzner in Frankfurt when contracts require it.

Auditable data pipelines

Claims and policy pipelines that produce reversible, reconcilable outputs — suitable inputs for regulator-facing reporting your finance team owns.

NDA & DPA on day one

Signed before discovery work begins. Redacted samples are fine while paperwork is in flight; nothing leaves a defined perimeter.

Discovery request

    NDA available on request · we never share your details

    Talk to us

    Claimants move faster when your internal workflows do.

    Tell us what's broken — we'll come back within one business day with a clear sense of what we'd build and what it would cost. No pitch deck, no marketing follow-up.

    • Identify one high-friction workflow worth automating
    • Validate data, security, and integration constraints
    • Build a focused AI-assisted prototype
    • Plan production rollout with human review and audit trails
    Prefer email?
    [email protected]
    Common questions

    The questions we get asked most.

    If yours isn't here, send us an email — we'd rather have a real conversation than a marketing one.

    Engagement

    For tightly scoped pieces — a specific integration, a migration, a focused prototype — we can quote fixed price after discovery. For broader work, dedicated-team engagements work better because insurance requirements shift as the build progresses.

    You do. Every line we write becomes your property at delivery. We can host and operate it, hand it off to your team, or both — your call.

    A working prototype against your real data takes two to four weeks. A production system typically ships in eight to twelve weeks, with rollout planned in phases after that.

    Data & security

    We sign NDAs and DPAs before discovery. Encryption at rest and in transit is default; access is role-based and audit-logged. We can work inside your VPC or run dedicated infrastructure for your tenant.

    On AWS (EU regions for European clients) or on bare-metal Hetzner infrastructure in EU regions (Frankfurt by default). EU data stays in the EU when contracts require it.

    Four layers: input filtering, prompt constraints, output validation against structured schemas, and least-privilege access to downstream systems. Every model output is auditable; nothing reaches a payout, claims record, or policy without a validation step.

    Delivery & integration

    We've shipped REST, SOAP, file-drop, and direct database integrations across legacy and modern systems. For your claims and policy admin, we map to whichever interface is most stable: vendor API, scheduled file transfer, or read-only database replica.

    Whichever fits the problem. Whisper for voice; GPT-4o and Claude for structured reasoning; open-source models when fully self-hosted inference is required. We architect for swap-ability.

    Either side, or both. We can operate the application on infrastructure you own, hand the runbook to your platform team, or stay on as a support contract — whichever matches your operating model.