Skip to main content
A.I. Commerce Studio

AI workflows and agents

AI-assisted approval of incoming invoices

Invoices from e-mail are extracted, checked against rules and reach accounting after a person approves.

Fictional scenario. An example situation that illustrates the approach. It is not a client project and describes no measured results.

Business problem

A fictional distributor receives about a thousand invoices a month by e-mail. Entry is manual, errors are found late and approval depends on one person.

Assumptions

  • Invoices are PDFs or photos in Latin and Cyrillic script.
  • The ERP has an API for draft documents.
  • Approval rules by amount and supplier are written down.

Constraints

  • No invoice is posted without a person.
  • Data stays in the EU.
  • No change to the accounting process in the first phase.

Reference architecture

Components by stage — an illustration, not a running system and not a certified design.

  1. 01Trigger

    • Invoice mailbox
    • Upload from a portal
  2. 02Input

    • Original kept unchanged
    • OCR and field extraction
  3. 03AI and rules

    • Extraction model
    • VAT number and amount checks
    • Match with the purchase order
  4. 04Review

    • Approval queue
    • Low confidence → a person
  5. 05Output

    • Draft in the ERP
    • Decision log
The diagram as a list
StageComponents
TriggerInvoice mailbox, Upload from a portal
InputOriginal kept unchanged, OCR and field extraction
AI and rulesExtraction model, VAT number and amount checks, Match with the purchase order
ReviewApproval queue, Low confidence → a person
OutputDraft in the ERP, Decision log

Interfaces

FromToVia
MailboxIntake serviceIMAP/Graph API, read-only
Intake serviceERPREST API, drafts
QueueAccountantWeb interface

Data flow

  1. The invoice arrives and the original is stored unchanged.
  2. Fields are extracted with a confidence score.
  3. Deterministic rules check amounts, VAT and the order.
  4. A person approves or corrects; the draft goes to the ERP.
  5. Every decision is logged.

Security boundaries

  • Read-only mailbox account.
  • The ERP API key is in a secrets store, not in code.
  • Personal data in notes is masked before the model.

Where a person approves

  • Every invoice above the threshold is approved by a person.
  • A new supplier always goes to review.

Delivery phases

  1. 01 · Discovery

    Sample invoices, rules and target fields.

  2. 02 · Pilot

    One supplier, extraction and review only, no ERP writes.

  3. 03 · Integration

    ERP drafts behind approval; more suppliers.

  4. 04 · Operation

    Evaluation set, error and cost monitoring.

Limitations

  • Handwritten invoices need manual handling.
  • Quality depends on the scan quality.

Trade-offs

  • More human review early on in exchange for fewer errors.
  • Drafts instead of direct writes — slower but reversible.

Facing something similar?

We start with an assessment of where you are and propose a realistic first phase.