Building a Financial Knowledge Graph From Unstructured Invoice Data
The Intelligence That Was Always There
A contract for office equipment supplies auto-renewed for a third year last October. The renewal clause was on page 3 of the original PDF, buried in a paragraph that began "unless the client provides written notice 90 days prior." Nobody reread page 3. Nobody set a calendar alert. The vendor's standard renewal terms included a 12% price increase — also on page 3 — that now runs for another 12 months. The accounting software recorded the new invoice amounts as normal entries. Nothing flagged. Nothing changed.
This is not a process failure. Your team is not negligent. It is the consequence of accounting software doing exactly what it was built to do: capture the regulatory minimum from each document and record the transaction. The contract clause was always there. The renewal date was always extractable. The price escalation was in black text on a PDF your system processed and filed.
Nobody built the layer that read those fields, linked them to your vendor entities, and surfaced them when they became time-sensitive. That layer is a financial Knowledge Graph.
---
Your Invoice Data Is Not Unstructured
The phrase "unstructured financial data" gets repeated so often it starts to sound like fact. It is the wrong diagnosis. Invoices are among the most structured documents in business. Every one has a vendor, a date, amounts, line items, payment terms, delivery references, serial numbers, and warranty clauses. The structure is inherent to the document type.
Extraction is the problem, not structure. Accounting software was designed to capture 3 to 5 fields from an invoice — date, amount, VAT, vendor name, and description — because those are the fields regulators require. The other 20 to 35 fields per document — serial numbers, warranty terms, price conditions, delivery references, payment conditions — are read by the software and discarded. They go into the PDF archive and stay there.
Your invoices are not unstructured data. They are structured intelligence that nobody extracted.
A company processing 200 invoices per month accumulates 7,200 invoices per year. At 25 extractable fields per invoice, that is 180,000 data points annually — of which accounting software captures roughly 18,000. The other 162,000 data points per year exist in your document history, technically readable, practically inaccessible.
Four years of invoice history represents approximately 648,000 unextracted data fields. Every vendor price trend, every warranty term, every contract clause, every serial number — all sitting in a filing system organized for compliance, not intelligence.
---
How the Knowledge Graph Is Built
Building a financial Knowledge Graph from invoice data happens in four steps. None of them require a separate IT project. All of them run automatically while the finance team works on other things.
Step one — SightCapture extracts every field. Stralevo's SightCapture reads each invoice the way a trained accountant would: vendor name checked against known aliases, every line item captured including fine print, payment terms compared to the contract on file, serial numbers and warranty dates noted. This covers 50-plus document formats and captures all 15 to 40 fields per invoice — not just the 5 that will appear in the accounting ledger.
Step two — SmartHub resolves entity identities. A mid-market French company with 8 years of invoice history across Sage for the parent entity, Cegid for the subsidiary, and QuickBooks from a recent acquisition has the same vendor listed as "ACME SARL," "Acme Corporation," and "ACME" across three systems. SmartHub recognizes these as the same supplier and creates a single vendor node with a complete 8-year transaction history spanning all three platforms. When the CFO asks "what is our full history with this supplier?" the answer crosses system boundaries automatically.
Step three — the Orchestrator connects related documents. An invoice does not exist in isolation. It connects to a contract (what terms were agreed), a payment record (when and how much was paid), prior invoices from the same vendor (price trend), and potentially a project or cost center. Stralevo's Orchestrator — the routing engine that coordinates between document types — establishes these links as each document is processed, so querying an invoice also returns the contract it was issued under, the payment status, and the trend context.
Step four — Needle Finder queries the connected graph. Once the graph is built, questions that previously required analyst projects become single queries. "Which suppliers have delivered late more than twice and also raised prices in the same period?" crosses vendor entity data, invoice date data, delivery reference data, and price trend data simultaneously. Needle Finder returns a sourced answer — each vendor named, each incident documented, the specific invoices cited — in seconds for typical query volumes.
---
What Becomes Possible
Questions that once required a multi-day analyst project become routine once the Knowledge Graph exists.
Somewhere in your invoice history from the last three years, a vendor charged above the contracted rate. Your accounting software recorded the amount as a normal invoice — because it processed the transaction correctly. What it did not do was compare the unit price on that invoice to the unit price in the contract PDF. The contract and the invoice were in separate systems, never connected.
Stralevo's Knowledge Graph links invoices to contracts automatically. Overcharges surface the first time a cross-entity query runs against the data. Companies running this query during their first Stralevo demo regularly discover duplicate payments or contract rate discrepancies that accounting software had been recording faithfully — and silently — for months.
Quarterly board preparation that currently takes two days of controller time — exporting from three systems, building an Excel model, cross-referencing supplier cost changes — becomes a 30-second query. "What changed in our supplier costs this quarter?" The Knowledge Graph compares current invoice data against the same period last year, flags the outliers, names the vendors, and cites the specific invoices. The two-day rebuild is now an afternoon review of exceptions already surfaced.
For French companies specifically, three regulatory formats that tax authorities can request on short notice become straightforward: FEC — the standardized accounting export that DGFiP can demand with 15 days' notice, carrying a €5,000 penalty for non-compliance — requires complete, sourced document references for every entry. A company running Stralevo is perpetually audit-ready because complete extraction is what the system does by default, not as a special preparation mode triggered by a letter from the tax authority.
---
The Three Questions That Used to Require Spreadsheets
Consider the questions finance teams currently answer most slowly.
"Is this vendor raising our prices gradually?" — answering it requires comparing unit price data across every invoice from that supplier over 18 months. From an accounting system: export to Excel, isolate the vendor, find the unit price column if it was captured at all, build the trend by hand. From the Knowledge Graph: one query, answered in seconds, with each price change cited to its specific invoice.
"Which contracts are within 60 days of auto-renewal?" — without a dedicated contract management system, this requires finding every contract PDF and rereading the renewal clause. From the Knowledge Graph: renewal clauses were extracted from each contract document when it was processed and linked to the vendor entity. The query surfaces every upcoming renewal with the clause cited and the deadline named.
"Which vendors share the same banking details?" — a fraud detection question accounting software cannot answer because banking detail data appears in the PDF body, not in the ledger fields. From the Knowledge Graph: extracted banking details are linked to vendor entities and flagged when the same account appears across multiple suppliers.
These are not exotic queries. They represent standard intelligence requests that finance teams currently handle inadequately — not because the data does not exist, but because the extraction layer was never built.
---
Financial Intelligence That Has Been Waiting
Once a CFO recognizes their invoice history is not a pile of PDFs — but a connected map of every supplier relationship, contract term, price movement, and payment behavior going back years — it changes which questions feel worth asking.
Everything already happened. The data already exists. The Knowledge Graph makes it navigable for the first time.
A company that starts extracting invoice data today will have 90 days of Knowledge Graph depth in 90 days — and a growing advantage over organizations that wait. Every document that arrives expands the graph. Every new invoice adds to the price trend. Every contract linked to its invoices extends the audit trail. The intelligence accumulates whether or not anyone asks a question about it.
At 162,000 unextracted data points per year for a typical mid-market company, the question is not whether your invoice history contains intelligence worth recovering — it does, more than accounting software has been capturing for every year you have been in business. The question is when you want to start querying it.