LF AI & Data: DocLang working group develops open standard for AI-native documents

LF AI & Data Foundation (Linux Foundation for AI and Data) announced on June 9, 2026, the launch of the DocLang Specification Working Group — a working group tasked with developing an open, vendor-neutral standard for AI-native documents. Founding members are IBM, Red Hat, and NVIDIA, with additional contributors ABBYY and HumanSignal contributing to the specification’s development.

The core problem: business documents were not designed for AI

The overwhelming majority of business documents — PDF reports, Word memos, Excel spreadsheets, PowerPoint presentations — were created for human consumption. A person effortlessly recognizes a heading, the difference between a table and a chart, the context of a parenthetical note, or the logical hierarchy of a chapter. AI systems have structural difficulties with exactly this.

When such documents are fed into AI pipelines, multiple conversion steps are required, context loss is common, and processing reliability varies depending on the conversion tool. At enterprise scale — where companies process millions of documents annually — that problem becomes critical and expensive to address through ad hoc integrations.

DocLang aims to bridge precisely that gap between how documents exist and how AI needs to receive them.

What does DocLang preserve and how is it structured?

DocLang is a format specification that preserves both semantic meaning and geometric layout of a document in a single representation. This means the format describes not only what is written in the document, but also where, in what structure, and with what relationships between elements.

Key technical characteristics of the DocLang specification:

• Structural elements: headings, paragraphs, and tables described together with their position on the page — enabling AI systems to understand the spatial context of content
• Governance controls: embedded information about privacy, extraction permissions, and model training rights — directly in the document, not as external metadata
• Tokenization optimization: the format is designed with modern LLM tokenizers and modeling approaches in mind, reducing information loss on entry into AI systems

Relationship to Docling: complementary tools

It is important to understand what DocLang is not: it is not a replacement for Docling. Docling is an open-source tool already in use that converts documents — accepting PDF, DOCX, PPTX, XLSX, HTML, and images and converting them to structured output. Docling solves the problem of ingestion and conversion.

DocLang operates at the next level: it standardizes how that structured output is exchanged between systems. Without a standardized exchange format, every AI system receiving a document from another system must know the specific output format of that system — resulting in a multitude of custom integrations and growing complexity at scale.

Docling and DocLang therefore form a vertical stack: Docling for ingestion and conversion, DocLang for standardized exchange of structured output.

Vendor-neutral governance and open membership

Industry standards that succeed share one key characteristic: no single vendor controls the specification unilaterally. DocLang is housed under the Joint Development Foundation (JDF) — an organization that ensures open governance through community consensus rather than unilateral decisions by individual participants.

In practice, this means that IBM’s implementation of DocLang must be compatible with NVIDIA’s or Red Hat’s, and vice versa. ABBYY, a long-standing player in document processing, and HumanSignal, a data annotation platform, contribute experience from domains where AI pipelines for documents most often encounter practical challenges: information extraction, quality management, and training data preparation.

The project reference page is available at doclang.ai.

Primary applications: RAG, exchange, and governance

DocLang is primarily focused on three categories of enterprise use:

1. Document preparation for AI: standardized input into RAG (Retrieval-Augmented Generation) systems, fine-tuning pipelines, and extraction of structured data from documents
2. Exchange between systems: a document can pass through multiple AI systems from different vendors without losing context or requiring re-conversion and format adaptation
3. Document governance: embedded controls define what may be done with a document — which becomes especially important under regulatory requirements such as GDPR or sector-specific regulations in finance, healthcare, and public administration

The initiative arrives at a moment when the industry is increasingly moving from simple chatbot applications to agentic workflows — multi-step automated processes in which AI autonomously processes documents and forwards them between systems. In that context, the absence of a common format becomes a bottleneck that the DocLang ecosystem aims to resolve at the industry level.