A financial services firm cut invoice processing time by 70% and eliminated manual errors

The accounts payable team manually opened email attachments, keyed invoice data into the ERP, matched line items against purchase orders, and routed approvals through email chains. Processing a single invoice took 15-20 minutes. Errors in data entry caused payment delays and strained vendor relationships. Month-end closes regularly ran two days late.

The firm processed between 400 and 600 vendor invoices per month, spanning over 80 active suppliers across legal services, technology, facilities management, data providers, and professional services. Invoice formats varied enormously, some vendors sent structured PDFs with clearly labelled fields, others sent scanned documents or partially filled templates, and a small number still sent paper invoices that were scanned by the receiving mailroom team. The AP team of three handled all of this manually.

The data entry step alone consumed the majority of AP capacity. A typical invoice required the team member to open the attachment, identify the vendor, invoice number, invoice date, due date, total amount, and individual line items, then key all of this into the firm's ERP system. For invoices with multiple line items, a common occurrence with IT service providers and legal firms billing by matter, this could take 20 minutes or more per invoice. Across 500 invoices per month, that represented over 150 hours of pure data entry work, before any matching, approval, or exception handling had begun.

Matching was the next bottleneck. Every invoice needed to be matched against a corresponding purchase order in the ERP to verify that goods or services had been ordered and received at the agreed price. Purchase orders were tracked in a system that did not integrate with the AP workflow, so matching required the team member to search for the relevant PO, compare line items manually, and flag any discrepancy for resolution. Discrepancies, wrong prices, missing line items, quantity differences, were common, and resolving them required email back-and-forth with the relevant department manager and sometimes the vendor.

Approval routing happened through email. Once an invoice was matched, the AP team member sent it to the appropriate approver based on a rules document that lived in a shared drive. Approvers were busy; invoices sat in inboxes. Month-end close was consistently delayed because the finance team was still chasing approvals on invoices that had been in the queue for two weeks.

We built an automation pipeline that extracts invoice data from email attachments and scanned PDFs using OCR and an AI classification model. Extracted fields are validated against purchase orders in the ERP. Discrepancies are flagged for human review; clean invoices route automatically through the approval chain based on amount thresholds and department rules. Approved invoices queue for payment with full audit trails.

The ingestion layer monitors two inboxes, the firm's AP email address and a dedicated invoices@ alias, for incoming messages with attachments. When an invoice arrives, the pipeline downloads the attachment and runs it through a multi-stage extraction process. Standard PDFs with embedded text are parsed directly. Scanned documents and images go through a two-step OCR pipeline that first cleans and deskews the image, then runs character recognition optimised for financial documents. An AI classification model then reads the extracted text and maps it to structured fields: vendor name, vendor ID (matched against the ERP's vendor master), invoice number, invoice date, payment due date, currency, total amount, tax amount, and line items with individual descriptions, quantities, unit prices, and amounts.

The extraction model was trained on a dataset of 3,000 historical invoices from the firm's 80 active vendors, which gave it strong accuracy on the formats it would encounter most often. For vendor formats it had not seen before, the model defaults to a higher-confidence extraction with manual review flagged. The extraction accuracy on clean PDFs is above 99%. On scanned documents, it is above 96%.

Validation runs immediately after extraction. The system looks up the vendor in the ERP, retrieves any open purchase orders linked to that vendor, and attempts to match the invoice line items against PO lines by description and amount. Three-way matching, invoice, PO, and goods receipt, is performed for invoices above a defined threshold. Matches that fall within a 1% tolerance are treated as clean. Discrepancies outside this tolerance are flagged with a specific exception type (price variance, quantity mismatch, no matching PO) and routed to the relevant AP team member with both the invoice and the PO displayed side-by-side for resolution.

Clean invoices route automatically through the approval chain based on two rules: the invoice amount (invoices below £500 auto-approve; between £500 and £5,000 require department head approval; above £5,000 require CFO sign-off) and the department code on the matched PO. Approvers receive a structured notification with the invoice summary and a single-click approve or query button. Approved invoices are posted to the ERP payment queue automatically with all supporting documentation attached, maintaining a complete audit trail.

Invoice processing time dropped by 70%. Manual data entry errors were eliminated. Month-end close moved from two days late to on schedule. The AP team was reassigned from data entry to vendor relationship management and exception handling.

The throughput improvement was transformational for the AP team. Processing an invoice went from 15-20 minutes of manual work to approximately 2 minutes of exception review, and for the 78% of invoices that processed cleanly with no discrepancies, no human time was required at all. The team of three was processing the same 400 to 600 invoices per month in a fraction of the time they had previously spent.

Manual data entry errors were eliminated entirely. Before automation, the AP team estimated they caught 2 to 3 keying errors per week, wrong invoice numbers, transposed amounts, incorrect vendor IDs, and suspected that a similar number went undetected until they caused payment problems. The extraction model does not make transposition errors or misread amounts under time pressure. The 99%+ extraction accuracy on standard PDFs, combined with the validation layer that cross-checks every extracted figure against the ERP, means errors are caught structurally rather than by human vigilance.

Month-end close moved from consistently running two days late to reliably completing on the last working day of the month. The improvement came from two sources: faster invoice processing meant there was no backlog accumulating through the month, and the automated approval routing with deadline reminders meant approvals were no longer stacking up in managers' inboxes. The finance team reported that the last week of the month felt qualitatively different, instead of a frantic chase to clear the AP queue, it was a normal week with a scheduled close process.

The AP team shifted their role entirely. The two team members who had been doing data entry and manual matching were redeployed to vendor relationship management, proactively managing payment terms, resolving disputes, negotiating early payment discounts, and building stronger relationships with key suppliers. The third team member focused on exception handling and process improvement, using the data now available from the automated pipeline to identify which vendors generated the most discrepancies and addressing the root causes.