Document Processing & Intelligence is the job of turning PDFs, forms, invoices and contracts into structured data — classification, field extraction, validation, and routing into downstream workflow and systems of record.
16 providers on the network offer it, publishing 256 APIs.
The cohort
| Provider | APIs | Band | Score |
|---|---|---|---|
| Webflow | 23 | exemplar | 71.9 |
| Adobe | 32 | exemplar | 69.4 |
| Reducto AI | 10 | exemplar | 67.1 |
| Nanonets | 10 | strong | 63.5 |
Also in the cohort: S&P Global, and eleven more.
Adobe belongs here for obvious reasons — it invented the format. Reducto AI and Nanonets are the current generation, both publishing ten APIs and both scoring well on small surfaces.
The category changed underneath itself
Document processing used to be OCR plus a template. You drew boxes on a sample invoice, told the system that the box in the upper right was the invoice number, and it worked until a vendor changed their layout.
That approach is gone. Modern extraction is model-based: hand it a document it has never seen, in a layout nobody templated, and get structured fields back. Reducto AI and Nanonets exist because that shift created room for companies that were not in the imaging business.
The consequence for anyone building on this is that the hard problem moved from extraction to confidence. Getting fields out of an invoice is now largely solved. Knowing whether the number you got is right — and what to do when it might not be — is not.
A build path
Classify before you extract. An invoice, a contract and a shipping manifest need different downstream handling, and running the wrong extractor produces confidently wrong fields rather than an error.
Extract with confidence scores, not just values. This is the field to insist on when evaluating vendors. An extraction API that returns invoice_total: 4820.00 is less useful than one returning the same value with a confidence and a bounding box, because only the second one lets you build a review queue.
Validate against something you already know. The purchase order, the contract, the vendor master. Extraction accuracy is a claim; cross-referencing is verification.
Route the uncertain to a human. This is the human-in-the-loop pattern that the rest of the catalog is missing, and document processing is the one category that has always had it — because nobody was ever willing to post a misread invoice straight to the general ledger.
The agent overlap
Document processing solved, years ago, the problem the rest of the API landscape is now confronting: an automated system produces an output that is probably right, and something has to decide whether to act on it.
The answer this category settled on — confidence scores, thresholds, and a review queue for anything below the line — is a working model of the consent infrastructure that agent-facing APIs almost universally lack. It is worth studying for that reason alone.
Takeaway
256 APIs across 16 providers, in a category where extraction is largely solved and confidence is the remaining problem. Document processing has been doing human-in-the-loop properly for a decade because the failure mode was always financial — which makes it the best-developed model of a pattern the rest of the catalog now needs.
Browse the use case at apis.io/use-cases/document-processing/.