Docupath, the AI-native document intelligence platform, has published AIDP, a new benchmark for measuring accuracy in document extraction and processing. It grades systems by the accuracy across 25 capabilities in a document processing pipeline, rather than by field-level extraction accuracy alone. Docupath’s own platform is the first system measured against it, with an AIDP Score of 97.66.

The release comes as document processing moves through another major transition with the emergence of AI. Over the last decades, the technology has gone from recognising characters and extracting fields, to processing documents and data with context and intelligent reasoning. According to Docupath, the way the industry measures accuracy has not kept pace.

“Most IDP and OCR platforms claims a high ‘field-level accuracy’ today. Yet, when conducting a study of almost 1,500 accurately captured documents, we found that only one in three was ‘system-ready’ for automated processing without at least one human correction,” says Sheran Corera, CTO at Docupath. “The industry needs a way to measure and compare accuracy that links to actual customer value – process automation.”

An old metric built for fields, not business processes

Field accuracy counts successfully extracted values from a document, one at a time. It is often ambiguous which fields are counted, whether empty fields are included, or how line items are tallied. As the original document is used for this benchmark, at perfection, the output data entering downstream systems is an exact match of the original, and often fails.

Measuring a system’s ability to turn documents into system-ready data, means going beyond reading the page. In example: 

  • - Normalize data formatting (i.e. dates, currencies)

  • - Reasoning conclusions (i.e. understanding that ACME GmbH & Co. KG", "ACME" and "ACME Group" are related)

  • - Apply business rules (i.e. applying data checks and validations to fit business requirements)

  • - Data cleansing (i.e. translate, enrich or correct document data to simplify processing)

  • - Conditional logic (i.e. route, flag or escalate documents fulfilling certain criteria)

AI brings new capabilities and errors

The emergence of AI raises the bar of ‘what’s possible’ in document processing. The contextual understanding and deep level reasoning of AI models leaped the industry forward but also opens for new types of errors. Older systems failed visibly, with garbled text and empty cells that a person could spot at a glance. AI models fail fluently. Ask one for a purchase order number that isn’t on the page, and it may return a well-formed PO number that looks entirely plausible and is entirely wrong.

Thus, a benchmark should both measure a system’s ability to apply these new capabilities as well as mitigate its risks and challenges. 


The AIDP Benchmark: 5 layers, 25 capabilities

AIDP grades a system across 25 capabilities divided into five layers: four layers describe the work. The fifth, Trust, describes the system doing it. Each is one way the work can fail, and each asks a single yes-or-no question.

 
Keeping capabilities separate matters because each failure has a different fix. A lost page, a missing value and a value in the wrong field are three different problems, and a single blended number hides which one a system has.

AIDP comes with an open corpus of 2,000 generated business documents. For each one, the answer key is written before the page is drawn, which makes invention provable: a value that isn’t in the key was never on the page.

 
The first result: Docupath

Docupath is the first system measured against AIDP. Testing a fixed software release, it achieved an AIDP Score of 97.66 (95% range 97.35 to 97.95) across 26 measures covering 21 of the 25 capabilities. The evaluation used every document in both benchmark datasets: AIDP-2K, with 2,000 business documents, and AIDP-CAP, with 960 more documents built to test harder conditions. 

Capabilities

Four of the 25 capabilities do not yet carry a score: Decoding, Computation, Anomaly detection and Escalation. 

The future of the AIDP Benchmark

“Technology advancements in document processing is moving fast, so any accuracy benchmark must evolve to stay relevant. That is exactly how we have designed AIDP, each version will raise the bar for what accuracy means. In the next version, we will add measures to test all 25 capabilities and make the benchmark available to other document processing platform provider,” says Sheran Corera. 

“As for Docupath, we are proud of the result, but we are not standing still. We keep shipping improvements and new capabilities to push the platform further.”

The full white paper, AIDP Accuracy Benchmark, sets out the specification, the corpus and Docupath’s complete results.

About Docupath

Docupath is an AI-native document intelligence platform that transforms business documents, such as invoices, orders and contracts, into structured and system-ready data. Built for finance, customer service, and legal teams, Docupath reads, validates, and corrects documents in real time, applying context-aware reasoning. The result is clean, reliable data flowing into systems, without need for manual review or training. Document processing made intelligent.