Automating invoice processing – how to stop rekeying data by hand

OCR (Optical Character Recognition) is a technology that turns an image or PDF into text. It has been around since the 1980s, it is a mature commodity today and on its own it delivers only one thing: raw text. It does not know where on the invoice the supplier tax ID, net amount or due date sits – it sees only a stream of characters scattered across an image.

Intelligent Document Processing (IDP) is the layer that turns raw text and layout into structured data. It classifies the document type (cost invoice, credit note, accounting note, transport document, contract), locates fields (supplier, buyer, document number, line items, totals), extracts values and validates them against business rules (matching to PO, budget limits, discount policy).

The difference between OCR and IDP is fundamental from a business perspective. OCR lets you search inside a document. IDP lets you post a document without a click. Only the second effect changes the economics of the AP function.

Classic manual invoice processing looks like this: the invoice lands in a shared mailbox, an assistant downloads the file, opens the ERP, manually types the supplier, amounts and cost category, picks the PO, routes for approval, chases the flow over email, fixes mistakes. An operation that requires 30 seconds of decision making takes 8–15 minutes with all the administrative steps.

At 500 invoices a month that is ~110 hours of team time – effectively an FTE on retyping data. At 5,000 invoices a month it is several FTEs. These are the visible costs. The hidden costs are larger: late payments (lost early-payment discounts), late-payment penalties, wrong accounting categories, missing audit trail during regulator inspections, and burnout in a finance team that spends days on tasks that require no expert decision. These costs are familiar from our approval bottlenecks analysis – AP is one of the processes where they hit hardest.

IDP does not eliminate all of this work – it eliminates the repetitive part. The human still approves the invoice, but approves a ready object with correctly filled fields and a matched PO, not a raw PDF requiring 12 manual steps.

A modern IDP pipeline has five steps that run automatically in seconds. The first is ingestion – the invoice enters the system (mailbox, OCR-ready scan, supplier portal integration, EDI API). Every source is normalised into the same entry point.

The second step is document type classification. The model decides whether we are dealing with a cost invoice, a credit note, an accounting note, a transport document, a delivery advice or a non-invoice attachment. Classification decides which sub-pipeline the document follows.

The third step is field extraction. For an invoice this means: document number, dates (issued, sale, due), parties (tax IDs, names, addresses), line items (description, unit, quantity, price, VAT), totals (net, VAT, gross, payable), currency, bank account, optional PO number. IDP models recognise these fields regardless of a specific supplier's layout.

The fourth step is validation. Does the tax ID exist in the official registry (GUS/VIES)? Does the line sum match the invoice total? Is the bank account on the VAT whitelist? Does the invoice match an open PO? Is the amount within the category manager's budget limit? All these rules run automatically and decide whether the invoice goes through straight-through processing (STP) or to human validation.

The fifth step is integration. The invoice with validated data lands in a Power Automate approval workflow and, after approval, in the ERP as a posted document. Audit trail of every step is captured automatically.

Classification is the most underrated element of IDP. Most organisations start with extraction and do classification manually – an employee picks the document type in a dropdown. Yet AI classification is one of the most mature applications today and reaches above 95% accuracy in stable domains.

The value of classification is that it automatically splits the document stream into sub-pipelines. A cost invoice goes to AP. A credit note triggers a different scenario (modify an existing entry). An accounting note needs different validation. A non-invoice attachment never enters AP – it is pinned to the original invoice as supporting evidence.

Classification is also the first place where Microsoft Copilot delivers real value. Instead of training a custom model, you can use AI Builder in Power Platform or Document Intelligence in Azure. For rare branch-specific document types, fine-tuning on your own corpus brings accuracy to production levels in 4–6 weeks.

Three years ago, invoice data extraction required per-supplier templates. A rollout meant hundreds of templates, and every layout change at a supplier broke things. Vision-based models (Document AI, Form Recognizer, Donut, LayoutLMv3) changed this fundamentally – they understand invoice structure regardless of layout, language and origin.

The second breakthrough is multimodal LLMs. Models like GPT-5, Claude Opus and Gemini Ultra can extract data from a PDF and at the same time answer contextual questions ('Does this invoice contain an unusual clause?', 'Is the VAT rate correct for this service category?'). This expands IDP from purely extractive to extraction-plus-decision – but it needs good AI governance to avoid introducing an uncontrolled decision layer.

The third breakthrough is availability. What used to need a data science team is now an API in Azure, Google Cloud and AWS. An average organisation can launch a production IDP pipeline in 8–12 weeks, not 12 months.

The value of IDP in the Microsoft ecosystem is that it is not an isolated tool. Power Platform (Power Automate + AI Builder) is the native orchestration layer, SharePoint is the document layer and Microsoft Copilot is the decision-support layer for the approver.

In a typical architecture the invoice lands in a dedicated SharePoint library via a mail connector. Power Automate runs an AI Builder model or Azure Document Intelligence, extracts fields and saves them as document metadata. The workflow matches the invoice against the PO in the ERP (Dynamics 365, SAP, IFS), validates the bank account against the VAT whitelist, checks budget limits and routes the case to the right approver through an adaptive card in Teams.

Microsoft Copilot, sitting on the approval layer, generates a 3–5 sentence invoice summary, flags anomalies (e.g. 'amount 38% higher than the 12-month average for this supplier'), highlights line items that need attention. The approver receives a ready decision to confirm, not a raw PDF to analyse. The decision cycle shrinks from 8–15 minutes per invoice to 30–60 seconds.

IDP needs governance alongside the technology rollout. The first pillar is retention and data classification – invoices are accounting documents subject to retention requirements (5 years in PL, varies elsewhere). SharePoint with retention policies handles this natively, but it requires architectural decisions before launch.

The second pillar is VAT whitelists (PL: bank accounts on the Ministry of Finance whitelist), KSeF procedures for structured invoices and GDPR alignment when invoices contain personal data (B2C invoices, named consultant invoices). Each area must be addressed in the workflow.

The third pillar is the approval and accountability model. Who approves an invoice in which value bracket? Who escalates? How do overrides work? How do we record deviations? IDP does not solve these questions – they are answered by company policy, which is worth designing in our advisory and strategy work with finance, legal and compliance.

The highest return comes in high-volume areas with repeatable structure. The first is classic AP – cost invoices from suppliers. The second is procurement – orders, advice notes, transport documents. The third is B2B customer service – customer orders, complaints, intake documents.

The fourth is HR – CVs, employee documents (tax forms, social security, declarations), requests. The fifth is legal – analysis of mass contracts (NDAs, service agreements), where IDP can pull out key clauses for human review.

The least cost-effective scenarios are very low-volume processes (a few documents per month) or highly varied document structure (each document different, no repeatable fields). For those, lightweight Copilot tools work better than a full IDP pipeline.

The first mistake is rolling out extraction without validation. Extraction itself has 90–97% accuracy – if the organisation does not build a validation layer (PO match, VAT whitelist, budget limits), 3–10% of bad extractions reach the ERP. Validation is often more important than the quality of the extraction model itself.

The second mistake is no exception path. IDP handles 70–85% of invoices without human input. The remaining 15–30% are cases where the model is uncertain (low confidence) or validation returns an error. Without a dedicated exception path (separate queue, dedicated UI, clear SLA), these invoices block the whole process.

The third mistake is deploying the AI layer without ERP and approval workflow integration. You then have very well extracted data that someone has to retype into the financial system. ROI drops to zero.

The fourth mistake is missing governance. Retention policy, permission model, SharePoint governance and audit trail must be designed before production – not after the first compliance audit. We run these stages with clients as part of solution design and implementation and growth.

Does IDP replace AP staff? No. It eliminates retyping and clicking, but people still make approval decisions, resolve exceptions and maintain validation rules. The work changes – from administrative to substantive.

Is Microsoft Copilot enough on its own, or do I still need a dedicated IDP pipeline? Copilot works well as a decision-support layer for the approver. For production extraction at volume, a dedicated pipeline with AI Builder / Azure Document Intelligence and rule-based validation is required.

How long does it take to roll out an IDP pipeline in a mid-size company? A pilot on one invoice stream – 4–6 weeks. A full rollout with ERP integrations, validations and workflow – 3–6 months. The cleaner the AP process before launch, the faster.

What about KSeF and structured invoices? KSeF changes the source for outgoing invoices, but incoming cost invoices will still come in many formats (XML, PDF, paper). IDP remains essential for the heterogeneous AP stream.

Does IDP require private AI? For most AP scenarios, Azure OpenAI / Azure Document Intelligence with EU residency is sufficient. For regulated industries (medtech, defence) it is worth considering private AI – covered in our AI on-premise vs cloud analysis.

OCR is a commodity today. Business value emerges only in a full IDP pipeline that combines classification, extraction, validation and ERP integration into one structured process. A well-built pipeline handles 70–85% of invoices without human intervention and changes the economics of the AP function at a scale no local improvement can match.

The most sensible first step is not picking a model but auditing the current AP process, mapping volumes and document classes and deciding which stream to pilot first. From there the rest of the architecture – Power Platform, SharePoint, Copilot, ERP – falls into place. At AlgorComp we support clients through this stage and run rollouts from pilot to full deployment.