Invoice and document intake with human approval

Demonstration, not a client case study Professional services / any Finance Standard · 2 credits

Before

Supplier invoices arrive as PDF attachments in a shared mailbox. Someone opens each one, types the supplier, amounts and dates into the accounting tool, guesses the cost centre and forwards anything large to a manager for approval by email. Approvals get lost, duplicates get paid and month-end is a scramble.

Trigger

New email with a PDF attachment in the accounts-payable mailbox

Connected systems: Outlook mailbox, OCR/AI extraction, lexoffice or DATEV interface, Teams

Automated steps

  1. Detect invoice attachments and skip newsletters, statements and reminders
  2. Extract supplier, invoice number, dates, net, VAT, gross and line items (OCR and AI extraction, per-field confidence)
  3. Match the supplier against the master list and check the invoice number for duplicates
  4. Suggest cost centre and account based on supplier history (AI step)
  5. Route invoices above the approval threshold or with low-confidence fields to the approver in Teams
  6. On approval, create the booking-ready record in lexoffice or the DATEV interface with the PDF attached
  7. Log every step, every extracted value and every decision to the audit trail

Human checkpoint

The approver receives a Teams card with the extracted fields, the suggested coding, the original PDF and any low-confidence flags. They approve, correct a field or reject with a reason. Invoices below the threshold with all fields at high confidence are queued for a daily batch review rather than approved silently.

Output

Booking-ready records in the accounting system with the source document attached, an approval record for every invoice above the threshold and a searchable log of what was extracted and who approved it.

What is measured

  • Minutes per invoice from arrival to booking-ready
  • Duplicate payments
  • Corrections after booking
  • Approval turnaround time

Risks and limitations

  • Extraction accuracy varies with supplier layouts; per-field confidence and review catch most errors but not all
  • Financial data is sensitive; access, retention and processor arrangements must be defined before launch
  • Accounting interfaces have strict formats and change over time
  • The workflow does not make payment decisions; it prepares them

Demonstration using sample data — not a client case study

Walkthrough

A supplier sends an invoice for €4,180 to the accounts-payable mailbox on Tuesday afternoon. The workflow recognises the attachment as an invoice, runs extraction and gets every field back with high confidence except the due date, which the supplier printed in an unusual format. It matches the supplier to the master list, confirms the invoice number has not been seen before and suggests the same cost centre used for this supplier’s last six invoices.

Because the amount is above the €2,500 threshold and one field is flagged, the invoice goes to the finance manager in Teams. The card shows the extracted values, the flagged due date with the original text beside it, the suggested coding and a preview of the PDF. The manager corrects the due date, approves, and the record is created in lexoffice with the PDF attached, four minutes after the email arrived.

A second invoice from the same supplier arrives the next day with an identical invoice number. The duplicate check stops it before extraction and posts a note to the finance channel.

Invoices under the threshold with every field at high confidence are not approved automatically. They wait in a daily batch that the bookkeeper reviews in one sitting. The workflow prepares bookings; people still decide.

Next step

Start with one workflow worth fixing

Answer nine short questions and get an honest read on whether your process is a good automation candidate — and which package fits. No sales call required.