AI-driven supplier invoice automation

The pain we solve

A mid-market company with a typical accounts payable volume (between 50 and 300 supplier invoices a month) spends two to five person-days each month typing data that is already written in the documents it receives. The math is simple: each complex invoice takes between six and twelve minutes of manual capture (header, lines, analytical allocation, link to PO or receipt, search for the right GL account, tax and withholding validation). Multiplied by monthly volume, that is 16 to 60 hours a month from a finance professional who should be doing closing, analysis or controlling, not transcription.

The cost is not just the hours. Typical errors in the manual process cost more than the capture itself: wrong account or cost centre allocations that go unnoticed until audit, duplicate invoices posted with one digit difference in the number, early-payment discounts lost because the invoice reaches the workflow on day 30 and the discount expired on day 15, miscalculated due dates, forgotten withholdings. Depending on volume, a moderately manual AP process leaves between 0.3% and 1% of annual supplier spend on the table in lost discounts and allocation errors.

In Spain the regulatory calendar adds pressure. The Ley Crea y Crece (the Spanish law mandating B2B electronic invoicing with a phased rollout by company size) and Verifactu (the near real-time AEAT tax reporting regime) require full traceability with no margin for error. Teams that already struggled to keep up with typing now have to comply with strict trace requirements. The answer is not to hire more accounting assistants: it is to take the typing away and leave the validation.

The underlying pain is always the same: the CFO knows the supplier information already exists written in some PDF, some XML or some email. But getting it well structured inside Business Central, with its PO, cost centre and GL account, costs days of team time. AI applied well closes exactly that gap.

What the AI does here

The core of the case is a three-layer pipeline. The first layer is Azure Document Intelligence, which extracts the structured invoice fields from each PDF or XML: tax ID, dates, amounts, taxable base, taxes, lines, discounts, due dates. It does so with a pre-trained invoice model that covers the standard case from day one, and it specialises per supplier with every correction the team makes, without requiring manual training.

The second layer is Azure OpenAI Service. This is what brings the missing context: which GL account each line corresponds to based on the supplier history, which cost centre or project fits the company usage pattern, which analytical dimension it should carry, which purchase order is the most likely match when the invoice does not carry an explicit reference. This is where a language model adds value to raw extraction: it turns fields into accounting entries.

The third layer is the integration with Business Central through the Davisa dvinvoice-hub extension and the standard BC REST connectors. The invoice enters as a pre-loaded purchase document, with header, lines, analytical allocation, link to PO and proposed due date. Finance reviews it on the usual BC screen, validates or adjusts, and posts. The approval workflow is triggered automatically via Power Automate to the right approver based on the authority matrix.

The entire pipeline runs inside your Azure tenant. Documents do not leave to public models, they are not sent to OpenAI direct, they are not used to train third-party models. Compliance aligned with GDPR and compatible with ENS (the Spanish National Security Scheme) depending on tenant level. The trace of every AI decision is logged: which invoice, what the model proposed, what the human validated, what was left as an exception for review.

What the AI does not do: post blindly. When the model does not reach the configurable confidence threshold, the line is flagged for human review. The default rule is restrictive: high-confidence invoices are auto-prepared; doubtful ones are routed to a human. The finance team stops typing, but does not stop deciding.

Before and after

How we deliver

When this case is NOT a fit

Some scenarios will not pay back the pilot. If you recognise yourself here, skip the discovery.

• If you process fewer than 30 supplier invoices a month. Hours saved will not justify the pilot investment. Below that threshold, a good BC template and disciplined manual capture performs better.
• If your chart of accounts and dimensions change every quarter. AI learns patterns. If patterns are constantly rewritten, the system never consolidates accuracy. First stabilise analytical governance, then automate.
• If you have not deployed Business Central yet. The case is native to BC with dvinvoice-hub. With a different ERP we can talk, but the value proposition drops sharply.
• If your real problem is approvals, not capture. If the bottleneck is approvers who do not respond, AI speeds up the first stretch and leaves invoices waiting all the same. The authority matrix and workflow need to be fixed first.