✦ Insurance Operations · System Integration · AI-Assisted Automation
Pilot delivered
From Manual Insurance Processing to an Integrated, Auditable Workflow
How we delivered an integrated insurance operations platform in five months, connecting document processing, configurable business rules and existing enterprise systems without forcing an immediate change in how people work.
5 monthspilot deliveredDelivered
8tracked workflow statesDelivered
9product lines in scope, 1 pilotedDelivered
Duration
5 months
Systems integrated
2 + reporting layer
Architecture
Hexagonal .NET
→ or swipe to begin
✦ The challenge
Where it started
An insurance services organisation was managing document-heavy workflows across nine Property & Casualty product lines. Proposals, policies, endorsements and invoices arrived through email, insurer portals and spreadsheets.
Operational teams had to extract information, compare documents, resolve inconsistencies and manually update multiple applications. The same information was often entered or checked more than once.
The objective was not simply to digitise documents. It was to establish a consistent process connecting people, business rules and existing enterprise applications, while preserving familiar working practices.
Delivered scopeDelivered
9Product lines in overall scope
1Product line in the pilot
8Workflow states
2Operational systems integrated
✦ Our approach
How we tackled it
A centralised operations platform, delivered as a five-month pilot that started with one high-volume product line and was built to expand without rebuilding the platform for every product.
01
AI ingestion foundation
Month 1
AI ingestion and document-processing foundation.
02
Platform, rules, portal, integrations
Months 2–4
Core platform, business rules, consultant portal and enterprise integrations.
03
Reporting layer & stabilisation
Month 5
Reporting data layer and stabilisation.
Three independently managed components: AI-assisted document intake, a central workflow with configurable business rules, and a consultant portal with dedicated system integrations.
✦ Architecture
Logical architecture of the pilot
Loading diagram…
Integration adapters isolate external applications; the AI intake service and rules engine stay separate from the core workflow.
Drag to pan · Ctrl/⌘ + scroll or pinch to zoom · click a block to trace its connections
✦ Technology
The stack, and why
ATAWS TextractOCR for incoming insurance documents.
Amazon BedrockClassification, structured extraction and cross-document comparison.
.NET / C#Core workflow with hexagonal architecture.
DroolsConfigurable, versioned business rules in decision tables.
JavaIsolated runtime for the rules engine.
ReactConsultant portal for review, correction and tracking.
HubSpotCRM synchronised through a dedicated adapter.
Qlik SenseReporting on a structured, read-only data layer.
Hover or tap a technology to see why it is there.
All key engineering decisions
Technology / pattern
Why it matters
AWS Textract + Bedrock
Converts unstructured insurance documents into information that can be validated and processed
.NET/C# + Hexagonal Architecture
Separates business logic from external services, allowing integrations to evolve independently
Drools, isolated Java runtime
Executes configurable business rules through an internal API, with versioned decisions and traceable outcomes
React
Supports exception handling and progressive adoption of structured submissions
Integration adapters
Connects existing enterprise applications without embedding their APIs into the core business logic
Qlik Sense data layer
Provides a consistent reporting source without giving analytics tools control over operational transactions
✦ Technical highlight
Designing AI out of the critical path
The most interesting decision was to treat AI-assisted intake as a transitional capability, not a permanent dependency. The organisation could not ask every consultant to abandon email-based workflows overnight.
The platform accepts existing submissions, extracts what it can and asks for human input only where needed. As consultants adopt structured forms, fewer documents pass through OCR and language models.
Because ingestion is separated from the core workflow, the AI service can be modified, replaced or reduced without redesigning business rules, integrations or submission management.
AcceptTake submissions the way consultants already send them.
ExtractPre-fill what confidence allows.
AskSend a link to highlighted fields only.
FadeStructured forms gradually replace OCR.
A practical adoption path that avoids long-term dependence on document-processing technology.
✦ Evidence · 1/2
The numbers
Pilot timelineDelivered
Five months in three phases. The remaining product lines and extra insurer-portal integrations are future expansion.
Eight tracked states, audited end to endDocumented design
Received
Reading
Extracted
Validated
Approved
Loading
Completed
Exception
Every decision and status change is recorded in an audit trail. Exceptions are a deviation from the main flow.
Confidence decides who does the workDocumented design
≥ 90%Pre-filled
70–90%Pre-filled and highlighted for review
< 70%Consultant enters the value
Required field missingBlocks the load
Reference policy for field-level extraction confidence.
✦ Evidence · 2/2
More numbers
Manual minutes per submissionModelled scenario
Simulation, not a production measurement. Assumes 1,000 submissions/month, 30 min of manual work before, and 75% of submissions needing no re-entry (75% × 5 min + 25% × 18 min).
Hours released per month, by volume and automation rateModelled scenario
Share of submissions needing no manual re-entry
50%
75%
90%
500 / month
154 h
181 h
198 h
1,000 / month
308 h
363 h
395 h
2,000 / month
617 h
725 h
790 h
Simulation under explicit assumptions. At the base case (1,000 submissions, 75%) the model releases 362.5 hours a month (−72.5% manual effort).
✦ Results & impact
What we can say, and what we can't yet
Delivered scope of the pilot, as reported by the delivery team.
Metric
Result
Evidence
Delivery duration
5 months
Delivered
Product lines in overall scope
9
Delivered
Product lines delivered in pilot
1
Delivered
Workflow states
8
Delivered
Operational systems integrated
2
Delivered
Reporting
Structured dataset for Qlik Sense
Delivered
Qlik Sense dashboard development was outside the core pilot scope. The operational-impact figures are simulations based on explicit assumptions, not measured project outcomes.
✦ What this project demonstrates
Principles that travel
More than an AI document-processing service: a way to connect established enterprise applications, isolate complex business rules, introduce automation with human oversight and deliver a platform designed for gradual operational change.
The same principles apply to financial services, logistics, compliance and enterprise back-office operations.
Automation is a path from today's workflow to a simpler one, not a flag day
Business rules belong in configuration, with traceable outcomes
Keep AI replaceable: isolate it from the core
Let analytics read, never drive, operational transactions
How to read the badgesDeliveredTeam-reportedDocumented designMeasured · acceptance testModelled scenarioIn delivery
✦ Your turn
Still typing the same data into three systems?
Bring us one painful workflow, the messier the better. We will show you where the first hour of every day comes back, and what a calm pilot looks like.
Amazon BedrockClassification, structured extraction and cross-document comparison.
DroolsConfigurable, versioned business rules in decision tables.