Sensible is the extraction layer under your AP workflow: schema-validated and traceable to the source page, so your finance team can trust the data without checking it by hand.
Accounts payable invoice extraction reads supplier invoices and returns typed fields such as vendor, PO number, line items and totals. Sensible uses a layout or LLM method for each field, and validation rules you set can check that line items, tax and totals add up before the data reaches your ledger.
| Description | Qty | Amount |
|---|---|---|
| Strategy workshop facilitation | 2 | $3,700.00 |
| UX research interviews | 8 | $3,400.00 |
| Design system audit | 1 | $4,200.00 |
| Account management | 12 | $2,220.00 |
| Travel & expenses | 1 | $684.00 |
| Description | Qty | Amount |
|---|---|---|
| Strategy workshop facilitation | 2 | $3,700.00 |
| UX research interviews | 8 | $3,400.00 |
| Design system audit | 1 | $4,200.00 |
| Account management | 12 | $2,220.00 |
| Travel & expenses (NTE) | 1 | $700.00 |
| Description | Qty | Amount |
|---|---|---|
| Account management: hours credited | −2 | −$370.00 |
Hover a field to trace it to the source. Sample data; output is illustrative.
Per field
LLM or deterministic method
Per document
pricing, whatever the page count
SOC 2 + HIPAA
SOC 2 Type II · BAA on Enterprise
No credit card
14-day free trial
Trusted by teams turning documents into production data






Headers, totals and line-item tables land somewhere different on every vendor’s document. Hybrid extraction handles the variation, and validation re-checks the math before anything reaches your ledger.
QuickBooks exports, SAP invoices, handwritten bills, custom ERP outputs: each one places fields somewhere different. Hybrid extraction (LLM parsing plus SenseML rules) absorbs the variation instead of breaking on it.
Multi-page tables, merged cells, multi-line descriptions and varying tax codes. Quantity, unit price, tax and amount have to stay attached to the right row, including across a page break.
Do the line items sum to the subtotal? Does the tax match the rate? Does the grand total reconcile? This check catches a bad extraction before it reaches the ledger, not after.
An invoice’s PO number is only useful if it can be checked against the PO and goods receipt it references. Extraction has to keep that reference clean enough to match on, not just capture it as a string.
Grouped the way AP teams ask for them. Each field comes back typed, keeps the text as printed, and can point to where it sits on the page.
invoice_numberstringinvoice_datedatesource.due_datedatepo_numberstringvendor_namestringvendor_addressstringbill_to_namestringbill_to_addressstringpayment_termsstringdescriptionstringquantitynumberunit_pricecurrencysource.unit_of_measurestringdiscountcurrencytax_amountcurrencyextended_amountcurrencyskustringsubtotalcurrencydiscount_totalcurrencytax_totalcurrencyshipping_freightcurrencygrand_totalcurrencyamount_paidcurrencybalance_duecurrencycurrencystringFive stages, and values can carry the page and polygon they came from all the way through.
PDF, scan, email attachment or spreadsheet. Any source, no pre-sorting.
Document type identified (invoice, PO, credit memo, remittance) before extraction runs.
Each field uses the method you chose: layout rules where a vendor’s format holds, an LLM where it varies.
Checks run on the extracted numbers: line items to subtotal, tax to rate, PO reference format. Failures go to review.
Typed JSON, with page and polygon per value when verbosity is set, sent by webhook or pulled by API.
Webhook or API into your AP system, ERP or matching engine. Three-way matching, approval routing, posting and payment happen there. Sensible hands over data you can trust; it doesn’t run those steps.
Getting a number out of a PDF is the easy part. The question your finance team asks is whether it’s the right number, so Sensible checks each invoice the way an AP clerk would. When a check fails, the field is flagged and can be routed for review instead of flowing into your ledger.
We don’t publish a flat accuracy percentage. Accuracy depends on document quality and configuration, and a single number hides both. You can test accuracy on your own invoices in a 14-day free trial, no credit card.
Sensible
Solutions engineers write the SenseML configs from your samples, adjust them when formats change, and help with integration work. You get the same engine as self-serve, with the configuration work done for you.
Engineers review your samples and pick the right method.
SenseML configs written from your samples.
Same engine as self-serve, ready for production.
Configs updated when formats shift or new edge cases appear.
Help with custom integration into your downstream systems.
The extraction logic is written out in SenseML rather than buried in prompts, so a new format can be added and reviewed.
QuickBooks · Xero · NetSuite · SAP · Oracle · FreshBooks · Wave · Custom / manual invoices
Standard invoices · Credit memos · Debit notes · Proforma invoices · Recurring invoices · Construction progress billing
Extracted against the same schema as the invoices that reference them, so the two can be compared field for field.
Received quantities captured in the same shape, so invoice lines can be matched against what actually arrived.
Payment references pulled out so they can be tied back to open invoices.
PDF (native or scanned) · Word · XLSX / XLS / CSV · JPEG / PNG · TIFF · Email bodies with attachments
Matching, duplicates, pricing, and where Sensible stops and your AP system starts.
No. Sensible is the extraction layer. It reads the invoice, the purchase order and the goods receipt, and captures the reference fields that tie them together (PO number, line quantities, unit prices, totals) in a consistent, schema-validated shape. The matching logic itself, including tolerances and exception handling, lives downstream in your AP system, ERP or matching engine.
Not as a built-in feature. Sensible’s validation rules run against one document at a time, and spotting a duplicate needs a comparison across documents you have already processed.
What Sensible does give you is a reliable duplicate key on every invoice: vendor, invoice number, invoice date and grand total, extracted the same way each time. Checking that key against your history belongs in your AP system or ERP. If you want that wired in as part of the pipeline, our managed-services team can help with the integration.
Sensible prices per document, not per token, so cost is predictable at 500 invoices a month or 50,000. One extraction result counts as one document, whatever its page count. A packet that Sensible splits into several documents counts each one. Enterprise plans add volume tiers; see pricing or talk to our team for a volume quote.
Validation rules in the configuration check that line items sum to the subtotal, that tax matches the stated rate, and that the grand total reconciles. A failed check flags the field and can route the invoice to human review. Results are delivered by webhook or API.
Accuracy depends on document quality and configuration, so we don’t quote a single flat number. What we can tell you is how the output is checked: schema enforcement and calculation validation, so weak extractions are flagged instead of passed through silently. The best way to know is to run your own invoices, including the ugly ones, through it in the 14-day free trial, no credit card needed.
Sensible holds a SOC 2 Type II report from an independent audit, and data is encrypted in transit and at rest. HIPAA is available on Enterprise with a signed BAA. Custom data retention is also on Enterprise, so talk to our team about the retention you need.
The starting schema has 25 fields across the header, line items and totals, listed in the table above. You can add GL code, cost center, remit-to details and more.
Sensible reads invoices whatever system produced them, including custom and manual formats. It returns JSON, CSV or Excel by API or webhook, so the data can go into QuickBooks, NetSuite, SAP or any ledger that takes an import.
Compare Sensible with Docsumo, Rossum, Nanonets and Parseur.