A finance team processing 50 invoices a day spends most of their time on three things: keying data from invoices into their system, checking that the invoice matches the purchase order, and chasing approvals. The actual judgment — “should we pay this?” — takes about 10% of the time. The other 90% is data entry and verification.
That ratio is backwards. The AI can handle the data entry and verification. The human should focus on the judgment.
The Invoice Processing workflow
The Finance starter pack includes a workflow called Invoice Processing. It takes invoice text (pasted or extracted from a document) and produces a validated, approval-ready package.
-
Invoice Data Extraction — The AI reads the invoice and extracts structured data: vendor name, invoice number, date, line items, quantities, unit prices, subtotals, tax, total, payment terms, and bank details. The output is structured fields, not a summary — every data point is labeled and extractable.
-
Discrepancy Check — The extracted data feeds into a recipe that compares the invoice against the purchase order or expected values. It flags mismatches: wrong quantities, unexpected line items, price differences, duplicate invoice numbers, math errors. Each discrepancy includes the expected value, actual value, and severity.
-
Condition: If Discrepancy Found — If any high-severity discrepancies exist, the workflow takes a different path. Minor discrepancies (rounding differences, formatting issues) are noted but don’t block processing.
-
Manager Approval — For invoices with discrepancies, the workflow pauses and sends an approval request with the full context: extracted data, flagged issues, and the original invoice. The approver can approve, reject, or request changes. For clean invoices, this step can be configured to auto-approve or still require sign-off depending on your policy.
Why AI extraction works for invoices
Invoices are semi-structured. They all have the same basic information — vendor, amount, items — but every vendor formats them differently. Some are PDFs, some are emails, some are scanned images converted to text.
AI models handle this variability well because they understand the semantics, not just the format. They know that “Total Due” and “Amount Payable” and “Balance” mean the same thing. They can handle multi-page invoices, line items that span columns in different ways, and tax calculations that vary by jurisdiction.
The extraction recipe uses a detailed output schema that defines every field the AI should look for. This structured approach means the output is consistent regardless of how the input invoice is formatted.
What the team still does
The workflow handles extraction and validation. The finance team handles decisions:
- Approval authority stays human. Every invoice goes through a human approval step. The AI provides the data and flags concerns. The human decides whether to pay.
- Vendor relationships are managed by people. When a discrepancy comes up, someone calls the vendor. The AI identifies the issue; the person resolves it.
- Policy exceptions are human calls. A late invoice with a discount that’s technically expired — the AI flags it as a discrepancy, but the finance manager might approve it to maintain the vendor relationship.
Running at scale
Most finance teams process invoices daily. The typical setup:
- Schedule the workflow to run every weekday at 8 AM
- Pull invoices from an email inbox or shared folder via webhook
- Process the batch, with results waiting for the team when they arrive
At 50 invoices per day, the workflow saves roughly 2 hours of data entry and 1 hour of manual verification. The cost in AI tokens is about $2 per day — the math isn’t close.
The full Finance pack
Invoice Processing is one of four workflows in the Finance starter pack:
- Budget Review — Variance analysis comparing planned vs. actual spend, with narrative explanations for stakeholders
- Revenue Reporting — Monthly revenue summary with forecast, trend analysis, and written narrative
- Audit Preparation — Documentation gathering, compliance checks, and gap identification
The pack also includes standalone recipes for expense report review, KPI dashboard narratives, department budget planning, and financial summary writing — each usable individually or as building blocks in custom workflows.