Document OCR & Data Extraction Engine

The problem

Most Dynamics 365 Finance & Operations implementations solve exactly one document problem: the AP invoice. Everything else that arrives as PDF or scanned paper is still keyed in by hand.

A purchasing team receives supplier order confirmations that change promised dates and quantities against a purchase order — someone opens the PDF, finds the order, and updates confirmed delivery dates line by line. A warehouse receives packing lists ahead of the truck and cannot pre-stage a receipt because nothing is in the system until the goods physically arrive. A treasury analyst receives a remittance advice listing forty invoice references against one lump payment, and has to type each reference into the settlement screen before the customer payment journal will balance. Bank letters, freight documents, certificates of analysis, and customs paperwork all end up in a shared mailbox that nobody owns.

The cost is not just typing time. It is late confirmations that never reach MRP, receipts that post against the wrong load, and unapplied cash sitting on account because the remittance never got matched. The data exists; it is simply trapped in a file format F&O cannot read.

What ECOSIRE builds

We build a document ingestion and extraction engine as an X++ extension package for Dynamics 365 Finance and Supply Chain Management. No overlayering — everything is extension classes, chain-of-command, table and form extensions, and new tables in your own model.

Ingestion channels

Documents enter through the channels that match how your business actually receives them: a monitored mailbox, a network or SharePoint drop folder, direct upload from the F&O document handling (Attachments) pane, or a REST endpoint for a scanning appliance. Each channel writes a row into a staging table with the raw file preserved in Azure Blob storage through the standard document management framework, so the original is always retrievable next to the record it produced.

Extraction and classification

An OCR and extraction service call runs inside the F&O batch framework, so throughput is controlled by batch groups and threads rather than blocking a user session. The engine first classifies the document type, then applies the field map for that type. Every extracted field carries a confidence score and the page coordinates it came from, stored alongside the value so a reviewer can see exactly where a number was read.

Matching into F&O records

Extraction alone is not useful — the value is in the match. The engine resolves extracted identifiers against real F&O data: vendor account from name, tax registration number, or bank details; purchase order and line from the order number plus item and quantity; customer invoices from invoice number, amount, and open balance for remittance settlement. Matching is legal-entity aware — a document is resolved within the correct company, and cross-company lookups only occur where you configure them.

Review workspace and posting

Documents that clear your confidence thresholds post straight through. Anything below threshold, ambiguous, or unmatched lands in a review workspace built on standard F&O form patterns, with the document image side by side with the extracted fields. A reviewer corrects a value, and the correction is captured so the field map improves over time.

Confirmed documents write through supported APIs and data entities — purchase order confirmation dates and quantities via the purchase order update path, arrival journals for packing lists, customer payment journal lines with settlement marks for remittance advices, and attachments back onto the source record.

Integration surface

Data entities and OData are exposed for the staging, extraction, and audit tables so extracted data is available to Power BI and to any external system. Where your process needs a canvas app for mobile capture or a Power Automate flow to route exceptions, we wire it through the Power Platform integration rather than building a parallel UI. Dual-write is used only where a document must also land in the Dataverse side of your estate.

Governance

Every extraction, every threshold decision, and every human correction writes an audit row: who, when, source file, field, old value, new value, confidence. Ledger dimension defaulting on any financial document follows your existing account structure and defaulting rules rather than a bypass. Security is delivered as duties and privileges layered onto standard F&O role-based security.

Who this is for

Organisations running F&O with meaningful inbound document volume: distributors and manufacturers processing supplier confirmations and packing lists, shared service centres handling remittance advices across multiple legal entities, and finance teams whose bank correspondence never reaches the ERP. It suits companies that have already automated the AP invoice and hit the wall on everything else.

It is a poor fit if you have a handful of documents a week, or if your documents are unstructured free-form correspondence with no repeatable fields.

How delivery works

1. Scoping call. We go through your actual documents — send us real samples with sensitive values masked. We agree which document types are in scope, which F&O records each one must create or update, and what your confidence and exception thresholds should be. 2. Fixed quote. You receive a written scope with the document types, field maps, target records, integration points, and a fixed price. No hourly drift. 3. Build. We develop the X++ extension package against a sandbox on your F&O version, using your document samples as the test corpus. 4. Install in test. The deployable package goes through your LCS pipeline into your sandbox. We run extraction against your samples with you, tune the thresholds and field maps, and iterate until the match rates hold. 5. Production. Once you sign off in test, the same package is promoted through your LCS pipeline to production. We are on the call for the first production runs. 6. Support. A support window is included from go-live, covering defects, threshold tuning, and questions.

Lead time is typically one working week from signed scope, depending on how many document types are in scope and how quickly sample documents and sandbox access arrive.

What this is not

This is not a shrink-wrapped product you download and install. Every build is scoped to your document set, your field maps, and your F&O configuration. There is no free trial, because there is nothing pre-built to trial — what exists is the engineering capability and the delivery process described above.