Skip to main content

AI Capabilities

SpendOne is an AI-based platform for indirect corporate spend. AI is used to remove manual data entry, suggest matches between documents and existing records, and surface anomalies for human review. It does not approve spend, post journal entries, or send payments on its own: humans remain in the loop on every financial decision.

This page describes what is automated, what is suggested, and where human review touches the workflow.

Document Extraction​

The extraction service reads PDFs and returns structured data. It is used for four document types:

  • Invoices: supplier, invoice number, dates, line items, tax breakdown, totals, payment terms, IBAN, and any PO numbers found in the body.
  • Order confirmations: order number, date, seller, buyer.
  • Delivery notes / slips: delivery number, date, seller, buyer.
  • Quotes: quote number, validity, totals, line items, plus subscription hints (billing frequency, term length) and product hints (URL, category).

Flow:

  1. Input: a PDF arrives via dashboard upload, API push, or the inbox ingestion worker, which watches the tenant inbox and pulls down attachments automatically.
  2. AI step: the document is sent to the extraction service, which returns a structured object.
  3. Output: the platform creates or updates the corresponding entity (invoice, delivery note, quote-backed purchase request) with the extracted fields pre-filled.
  4. Human review: for invoices, the user reviews the extracted fields before approval. Confidence is reflected in substep state, and any field can be edited.

Automated: field extraction itself, persistence of extracted values, creation of the invoice record from an inbox email.

Suggested, not committed: the mapping of extracted seller details to a specific supplier in your master data, and the assignment to a specific purchase order. Both go through the matching step below.

Supplier and Recipient Matching​

When an invoice arrives, the extracted seller and buyer information is matched against existing supplier and entity master data.

Flow:

  1. Input: extracted seller block (name, address, tax ID, IBAN) and buyer block.
  2. AI step: the recipient and semantic matchers parse the billing address and compare it against existing suppliers and tenant entities.
  3. Output: a ranked list of candidate suppliers with confidence indicators.
  4. Human review: the top candidate is pre-selected; the user can confirm or pick a different one. New suppliers are not created automatically.

Purchase Order Matching​

Invoices are reconciled against open purchase orders so that two-way and three-way matching can happen without manual lookup.

Flow:

  1. Input: the invoice text and any explicit PO numbers found by extraction.
  2. AI step: PO references are extracted from the document and matched, then reconciled against open purchase orders on amount, supplier and line items.
  3. Output: one or more candidate POs with a match status (exact, partial, none).
  4. Human review: if the match is exact and within tolerances, the link is pre-applied; the user confirms before approval. Mismatches are surfaced as accountant tasks.

Subscription Detection and Follow-up​

Recurring spend has its own loop. The platform detects when an invoice belongs to an existing subscription and flags amount drift.

Flow:

  1. Input: invoice extracted fields plus any subscription metadata on the matched supplier.
  2. AI step: the new invoice is compared against the contract and the prior invoices in the series.
  3. Output: the invoice is linked to the subscription; if the amount changed unexpectedly, a flag is raised.
  4. Human review: accountant or admin reviews the flagged delta before approval.

Inbox Ingestion​

Inbox ingestion connects to a per-tenant mailbox, polls it on a schedule (every two minutes by default), downloads PDF attachments, and feeds them through the extraction pipeline. Successfully processed messages are moved to the trash mailbox; an outright delete is only the fallback when the move fails.

This is automated end-to-end up to invoice creation. The resulting invoice still goes through matching and approval before anything posts to accounting.

What AI Does Not Do​

To be explicit about the boundary:

  • AI does not approve invoices or purchase requests.
  • AI does not post journal entries to your accounting system.
  • AI does not release payments.
  • AI does not create new suppliers from extracted data without human confirmation.
  • AI does not modify approved records.

Every financial state transition is gated by either an explicit user action or a deterministic rule configured by an admin (the approval rules engine).

Security and Data Handling​

For the platform's overall AI security posture (model providers, data residency, retention, and customer controls), see https://spendone.com/en/ai-security.