Finance Invoice Extract

finance/invoice-extract — skill payload

This file is rendered by an MCP agent when it asks for skill://finance/invoice-extract/SKILL.md. The manifest at ../invoice-extract.manifest.json is the machine-readable contract; this file is the narrative a host LLM reads before deciding to invoke the skill.

What this skill does

Pulls structured fields (supplier, invoice number, line items, totals, tax) out of a single invoice document (PDF, image, or scanned page). Returns an InvoiceFields record that downstream skills can consume — notably finance/po-match for three-way reconciliation.

When to use it

• The user has an invoice and wants to do anything programmatic with its contents.
• A composed workflow needs InvoiceFields as input.
• You already have raw OCR text but need it structured into named fields.

When not to use it

• The document is not an invoice (use a general document-parse skill).
• You need multi-document batch processing (this skill is single-doc; call it in a loop or look for an invoice.batch-extract capability).
• The data classification is restricted or higher — this skill is certified up to confidential only.

Inputs and outputs

Determinism and risk

Determinism score: 0.7 — same input, same output ≥ 70% of the time on the validation set. Risk: low — read-only, no side effects on external systems. Safe to call speculatively.

How it composes

document.parse → invoice.extract → (downstream) invoice.match → po.reconcile. See the ontology graph at prototype/out/ontology.mmd.