$11.8M in Receivables, Collected Systematically

The challenge: collections lived in inboxes and spreadsheets

This US insurance services firm bills insurance carriers for claims investigation work — an unusual AR profile where the people who approve payment are ~2,500 individual claims adjusters spread across 1,200+ carrier customers. Open invoices stretched back years, and the ledger stood at $11.8M gross with over 1,500 invoices more than 180 days overdue.

The ERP (Dynamics 365 Business Central) held the invoices, but the actual collections work — who called whom, what was promised, which invoice was disputed — lived in personal inboxes, spreadsheets and memory. There was no shared system of record, no prioritisation logic, and no way for a manager to see whether promised payments actually arrived.

The source data was messy in the ways real AR data always is: inconsistent status flags, invoices missing salesperson codes, free-text payer names that couldn't be joined to anything, and historical rows from a legacy accounting system polluting the queue. Any automation had to survive that reality rather than assume clean inputs.

What we built: a collections workspace on top of the ERP — not inside it

We built a custom AR collections platform that reads Business Central data through Azure Synapse, scores every account's risk daily, generates and routes follow-up tasks automatically, and sends templated dunning emails and statements through Microsoft Graph — while the ERP stays untouched as the clean source of record.

The design principle was read-only on the ERP: a daily sync mirrors open AR from Business Central (via an Azure Synapse view) into the platform's own Azure PostgreSQL database, and everything the team creates — call logs, notes, promises to pay, disputes, tasks, emails — is written back to the platform, never the ERP.

On top of that foundation sits the workflow engine the team actually lives in:

• Work queue & dashboards — a prioritised queue of open invoices with aging, ownership and follow-up status, plus KPI cards for total open AR, 30+ day overdue and promises due this week.
• Balance-weighted risk scoring — a weighted-average-days-overdue score recalculated daily for every payer, so a $50,000 invoice at 60 days outranks a $500 one at 90 — feeding four escalation tiers up to executive level.
• Automated task generation — grouped outreach tasks on a twice-monthly cadence for routine accounts, plus weekly tasks for broken promises to pay and seriously aged invoices — auto-closed when paid or resolved.
• Promise-to-pay tracking — every promise gets a date and amount, auto-creates a follow-up task on the promise date, and fulfilment is measured — including milestone tracking for payment plans.
• Dunning emails & statements — escalating templates from friendly reminder to formal notice, sent through Microsoft Graph from the company's own mailboxes with the full trail logged against the account.
• Role-based visibility — Entra ID single sign-on with five roles, so sales reps see only their own accounts while managers and executives see their tier of the escalation pyramid.

The Microsoft environment: Synapse, Graph, Functions, Entra

The whole platform runs on the client's existing Microsoft estate: Azure Synapse serves consolidated ERP views, Azure Functions run the daily sync and task-generation schedules, Microsoft Graph handles email and SharePoint document access, and Entra ID provides single sign-on — no new passwords, no data leaving their tenant.

Invoice PDFs stay in SharePoint and are opened on demand through Graph rather than being copied into the platform. A separate verification job even checks that each PDF actually contains the invoice it's supposed to — because filename-based matching had burned the team before.

Authentication uses managed identities throughout, so there are no connection-string secrets to rotate. Deployment is fully automated: a push to the main branch builds a container and ships it to Azure Container Apps.

Intelligent automation on messy data — without pretending it's AI magic

This platform is a case study in getting automation right before layering AI on top. The intelligence is deterministic and auditable: a risk-scoring algorithm the finance team can explain to an auditor, escalation rules they can edit in a settings screen, and task cadences they control — thresholds, templates, routing and override rules are all configurable in-app without a code deploy.

That matters in collections, where every automated action is a customer touchpoint with money attached. The result is a system the team trusts — and a clean, structured activity dataset (every call, promise, dispute and email) that makes an AI layer for reply drafting and prioritisation a straightforward next step rather than a leap of faith.

The result: $11.8M worked systematically instead of reactively

Collections went from inbox archaeology to a managed pipeline. Every payer has an owner, a risk tier and a next action. The executive team sees its escalation tier — around $3M of the ledger — without asking anyone for a spreadsheet. Broken promises surface automatically after seven days instead of whenever someone remembers.

Management reporting that used to be assembled by hand — collector productivity, promise fulfilment, aging trends — now comes off the platform's own activity data, with a daily snapshot history that makes the trend line real rather than reconstructed.

Frequently Asked Questions