How AI extraction works
When an invoice arrives, the agent reads the entire document and extracts structured data automatically. This page covers what gets extracted, how confidence scoring works, and how the system gets smarter.
Invoice detail page showing extracted header fields including vendor, invoice number, dates, tax, total, payment terms, and type.
Supported documents
The agent handles:
- PDFs, including scans.
- Images (PNG, JPG, JPEG, TIFF).
- Word documents (DOCX).
- Spreadsheets (CSV, XLSX).
A clean digital PDF, a photo of a crumpled paper invoice, and a CSV export from a vendor portal all work. The agent also handles low-quality formats: blurry scans, photos taken at angles, and images with poor lighting still extract cleanly. Multi-language documents are supported as well.
Fields extracted
Header fields
| Field | Description |
|---|---|
| Vendor Name | The company or person that issued the invoice. |
| Invoice Number | The unique identifier on the invoice. |
| Invoice Date | The date the invoice was issued. |
| Due Date | The payment due date. |
| Payment Terms | Terms like Net 30, Net 60, Due on Receipt. |
| Currency | USD, EUR, GBP, etc. |
| Subtotal | Total before tax. |
| Tax | Tax amount. |
| Total Amount | The final amount due. |
| PO References | Any purchase order numbers on the invoice. |
Line items
For every line item:
- Description of what was purchased or provided.
- Quantity.
- Unit price.
- Line total.
- Suggested GL code.
- Suggested dimension values (department, cost center, project, and so on).
- Line-level PO reference, if present.
Confidence indicators
When the agent is unsure about a value, the field is highlighted. Hover over a highlighted field for a tooltip prompting you to verify the value against the original document.
Fields without a highlight are high-confidence. A quick glance is usually enough.
Start your review with highlighted fields. They are the ones most likely to need correction.
AI Summary
Every invoice gets a structured AI-generated summary: vendor context, what this invoice is for, and any issues the agent flagged. The summary sits in its own panel, always visible on the right side of the invoice detail page. See AI Summary.
Vendor matching
The agent identifies the vendor and matches it to your vendor directory using name, address, tax ID, and other identifiers. Unmatched vendors surface as a notice for you to resolve, either by selecting an existing vendor or creating a new one. See Vendor Matching.
GL coding
GL account code suggestions come from:
- Vendor history: how you have coded invoices from this vendor in the past.
- Line item descriptions.
- Your Agent instructions, which tell the agent how your company codes.
Accept the suggestions or change them. See GL Coding.
How accuracy improves over time
Mod AI learns from your corrections. When you fix a vendor name, correct a total, or change a GL code, the agent factors that into future extractions.
- Vendor-specific learning. Corrections to Vendor A invoices improve accuracy for future Vendor A invoices.
- Pattern recognition. The more invoices the system processes, the better it gets at recognizing your invoice formats.
- Coding memory. GL code corrections teach the agent your coding preferences for specific vendors and line item types.
- Agent customization. For company-wide behaviors, write them into your Agent instructions so they apply to every future invoice.
Reviewing and correcting extractions is not just fixing today's invoice. It is training the system to be more accurate on every future invoice. Over time you correct less.
Fix invoices with Copilot
For corrections that span multiple fields (for example, reassigning all line items to a new GL account), Copilot lets you describe the change in plain language, see a diff, and accept or reject the update.