Staple AI

Inside the document pipeline

Staple extracts structured information from documents. Its API describes each field with a value and its origin: extracted, inferred, mapped or set. Extracted values can carry confidence scores. Inferred fields deliberately do not receive those same scores.

That separation makes it possible to tell whether a value came from the page or from an additional judgment. Developers can pass the result into an existing workflow through Staple’s APIs and software tools.

The evidence travels with the data

Staple also describes signed data records containing source-file hashes, processing history and model versions. It uses SHA-256 hashes and Ed25519 signatures to support verification. This gives downstream users a way to check the record’s provenance and integrity.

What I would test

A signature can show that a record has not changed. It cannot prove that the system read the invoice correctly. I would test those two things separately.

For extraction, I would use difficult layouts, unclear scans and totals that do not reconcile. For the evidence trail, I would change a document or a field and check whether that change is detected. I would also follow a human correction through the record.

This is why the distinction between extracted and inferred fields catches my attention. A system should make uncertainty easy to see. Otherwise, a neat spreadsheet can hide a very messy mistake.

The business model and moat

I see the buyer paying to reduce manual handling while keeping work reviewable. The useful economic measure is cost per correctly completed document, including the documents that still need a person.

The possible moat is a combination of difficult document handling, workflow connections and evidence that customers can depend on. Basic text extraction is becoming easier to buy. Reliable exception handling and a useful audit trail can make a product much harder to replace.

I would look for more document volume without a matching increase in manual review effort. That would help me understand how the model scales.

My take

I like the direction from extraction towards defensible data. It makes the product relevant after the first read, when someone needs to check an answer or resolve a dispute. The opportunity is to make that evidence practical enough that people actually use it.