Lyzr · Billing Automation for a Global Insurance Brokerage

At a glance

RegionNorth America

FunctionBilling Operations

Team~40 billing analysts

Volume~1,000 invoices / day in scope

SystemsEPIC, RMS, IRD, CABR, Industry Mapper

Insurance Brokerage Finance Ops Agentic AI Human-in-the-Loop

The Client

Global scale.
A workflow still running on people.

A leading global insurance brokerage and advisory firm. North America. B2C billing operations processing a continuous stream of activities — every day, without pause.

Financial and risk services — one of the largest in the world

North America billing team processing 1 million+ invoices annually

Mature enterprise stack: EPIC, SharePoint, Azure AD already in place

Gap: AI remained assistive, not operational — every workflow hit a ceiling within 12 months

The Challenge

Four problems.
One number that explains all of them.

Five systems. One biller.

EPIC, RMS, Industry Mapper, IRD, CABR

Every activity = five separate logins

Constant context switching — no single view

19 fields. All typed by hand.

Policy numbers, premiums, dates, commissions

Re-keyed from unstructured PDFs and emails

Every activity. Every time.

Every biller, slightly different.

No standardized process between operators

Errors surface weeks later — in compliance

Expensive to trace. Costly to fix.

No audit trail that holds up.

No field-level record of who changed what

Compliance questions had no clean answer

The answer lived in the biller's memory

~1,000

invoices / day

minutes / invoice

300 hrs

of analyst time. Every working day.

The Solution

9 agents. 22 steps.
One review screen.

An orchestrated pipeline of AI agents handles gathering, reading, extracting, and validating. The biller reviews a single confidence-scored dashboard and approves. Nothing posts without their sign-off.

AGENTagent/activity-parser

Activity Parser

Pulls open activities from EPIC every 5 minutes, no manual queue.

AGENTagent/rms-retrieval

RMS Agent

Retrieves binders, invoices, and quotes from SharePoint. Encrypted, logged.

AGENTagent/ocr

OCR Agent

Converts PDFs and emails to structured, machine-readable text.

AGENTagent/kv-extraction

Key Value Extraction

All 19 billing fields extracted, each with a confidence score.

parallel

AGENTagent/epic-validate

Validation Agent

Cross-checks extracted data against EPIC policy records.

AGENTagent/industry-mapper

Industry Mapper

Classifies by department, profit centre, and line of business.

parallel

AGENTagent/ird-validate

IRD Validation

Checks for special invoice routing instructions.

AGENTagent/cabr-validate

CABR Validation

Confirms premium payable entity and banking reference.

AGENTagent/dashboard-summary

Dashboard Summary

Assembles outcomes and recommended actions for the biller.

HUMAN GATEhuman/biller

Biller Approval

Reviews, corrects, or raises an RFI. Nothing posts without sign-off.

edit any field → partial rerun, only downstream steps approved corrections retrain future runs 9 agents · 22 steps · 1 human gate

scroll to run

How it works — click any step to expand

The pipeline above compresses into six phases a biller actually experiences — most of which they never have to think about.

Activity Intake

System auto-fetches open billing activities from EPIC every 5 minutes. No manual pulling.

Filters by North America owner codes automatically

Search by Activity ID, client code, policy number, or entity

Auto-run scheduler clears backlog — no human trigger needed overnight

Document Retrieval

RMS Agent reads activity notes and retrieves binders, invoices, quotes from SharePoint. Encrypted and logged.

Follows direct paths when available; recursive search otherwise

Every document linked back to original for side-by-side comparison

Transfer integrity-checked on every download

OCR + 19-Field Extraction

PDFs become structured data. 19 billing fields extracted, each with a confidence score. Large documents split and parallelized.

Policy numbers, premiums, dates, commissions, payable entities — all automatic

Conflicting documents: most recent wins, reasoning recorded

Score 90%+ = auto-process. Below 85% = biller reviews.

Multi-Layer Validation

Five check groups run automatically: EPIC, Industry Mapper, IRD routing, CABR banking, billing consistency.

Fuzzy client name matching (85% threshold), exact policy matching, semantic LOB classification

Commission cross-checks: commission = rate × premium within tolerance

Every mismatch surfaces with expected vs. actual — not just a red flag

Human Review and Approval

One dashboard. All checks. Document alongside extracted values. Biller corrects, approves. Nothing posts without sign-off.

Edit any field — partial rerun executes only what's downstream. Upstream compute reused.

Every edit: field-level diff with user, timestamp, and biller notes

Raise RFI when data is missing — no guessing

Self-Learning Loop

Biller corrections feed back into the agents. The system improves with every use. Auto-rate rises. Manual corrections fall.

Approved corrections become better prompts — injected into future runs

Billing expertise encodes into the pipeline — not locked in senior billers' heads

Cost per invoice decreases as accuracy increases

9 agents in the pipeline

Activity Parser

Reads notes, extracts policy IDs and document links

RMS Agent

Locates and downloads billing documents from SharePoint

OCR Agent

Converts PDFs and emails to machine-readable text

Key Value Extraction

Pulls all 19 billing fields with confidence scores

Validation Agent

Cross-checks extracted data against EPIC policy records

Industry Mapper

Classifies policy by department, profit centre, GLOB type

IRD Validation

Checks for special invoice routing instructions

CABR Validation

Confirms premium payable entity and banking reference

Dashboard Summary

Produces outcomes and recommended actions for the biller

Inside the product

B2C Billing Automation — Agentic Data Flow diagram

Activity detail with attachments and policy list

The Outcome

What changed.
By the numbers.

Design targets, built into the architecture from day one and instrumented per activity. Baseline vs. actual captured on every run.

~0%

Straight-through automation. Activities auto-processed without manual data entry.

~0% faster

Billing turnaround time. From activity open to biller-approved, ready to post.

~0%

Manual validation steps eliminated. Five check groups run automatically — exceptions surfaced.

~0% fewer

Billing errors and corrections. Per-policy validation and pipeline consistency remove operator variance.

What Actually Changed

The shift that took longest to explain — and mattered most.

Ownership model

From a tech-owned project to a co-owned workflow.

The architecture was never just about speed. Business teams now design and evolve the billing workflow directly — IT became the enabler, not the bottleneck for every change.

Not every change needs a sprint anymore. — the shift leadership felt first.

Before

×AI projects owned and driven by technology teams

×Business teams consulted at the requirements stage, then handed off

×Every change required a developer; iteration was slow

×AI as a capability: assistive, advisory, not operational

After

✓Tech and business teams co-own the workflow — actively

✓Business teams design and evolve workflows, not just approve them

✓IT as architect and enabler — not every change needs a sprint

✓AI as infrastructure: live in the bill-to-cash operation, daily

Controls and Governance

Every result is reviewable.
Every edit is recorded.

Built for a finance operations team that could not afford an opaque system — where compliance, auditors, and senior leadership all need a clear answer to "why was this billed this way?"

Two HITL checkpoints: Activity Parser stage and Key Value Extraction stage

Edit any extracted field — partial rerun executes only downstream steps

Raise an RFI when data is missing — the activity is flagged, never guessed

Nothing posts to EPIC without explicit biller approval

Feature-flag gating — write-back can enable gradually as trust builds

Same judgment, far less transcription — billers keep the call, lose the typing

Field-level edit history on every correction: run ID, timestamp, step, policy, field, old value, new value

Who made the change, with biller status and notes attached

Full report exportable to CSV / Excel for compliance teams

Analytics dashboard: where runs fail, how often HITL triggers, throughput

Every document transfer logged, encrypted in transit, integrity-checked

Full prompts are not logged in production — compliance-aware by design

Azure Active Directory authentication — users explicitly assigned

Role-based feature sets: Admin and Biller roles with separate permissions

JWT auth, strict CORS allow-list, rate limiting, request-size limits

Managed-identity database access — no hardcoded credentials

Resilience: retries with backoff, startup reconciliation of interrupted runs

Slotted into existing Azure infrastructure — no new auth layer required

Owner reassignment — billing activities can be transferred between team members

No architectural change required to the existing EPIC or SharePoint environment

Read It From Your Seat

This story reads differently
depending on where you sit.

Pick your role. We'll surface what matters to you.

Cost recovery is built in

300 analyst-hours recaptured daily — the floor, before accuracy gains compound and the auto-rate rises with use.

Errors caught at creation

~60% fewer billing errors, caught before they ever reach compliance. Corrections that used to cost weeks now surface in seconds.

Audit questions have answers

Every field, every change, every reason — logged and exportable. Compliance stops chasing billers for explanations.

Cheaper the more it runs

Biller corrections improve future runs. Cost per invoice decreases as the pipeline learns from every approval.

No architectural change

Slotted alongside EPIC, SharePoint, and Azure AD — evaluation to production without a platform migration or new auth layer.

Security-first by design

JWT auth, Azure AD, CORS allow-list, managed-identity database access, rate limiting. Compliance-aware logging in production.

Built for operational resilience

22 dependency-ordered steps, parallel OCR with bounded concurrency, retries with backoff, startup reconciliation of interrupted runs.

Partial rerun architecture

Edit one field, re-execute only downstream steps. Feature flags control the rollout of write-back to EPIC.

Billers gained clarity, not less control

One dashboard, confidence scores, source document beside extracted values. Same judgment, far less transcription work.

Operations visibility that didn't exist

An analytics dashboard shows where runs fail, how often human review triggers, and throughput across the team, in real time.

The backlog clears without a trigger

An auto-run scheduler picks up open activities and emails a batch report automatically. Night shifts stop being a bottleneck.

One pattern, repeatable across teams

The same agentic pipeline and governance model already scales across billing teams and geographies — the blueprint exists.

300 hours of billing work.
Every day. Automated.

Global scale.
A workflow still running on people.