Enterprise AI has a reading problem, and PDFs are a big part of it.
A PDF doesn't actually know what it says. It only knows where to put ink. IBM's lead researcher on the project explains that PDF tells a renderer where pixels go, DOCX tells a word processor how to lay out a page, and neither was designed for a machine that needs to understand the content (IBM Research). So when your AI reads a PDF, it's basically squinting at a photo of a document and guessing. Tables turn into word soup and captions wander off from their figures.
That's the problem DocLang wants to fix.
DocLang is a new file format, except it's written for AI instead of people. IBM, NVIDIA, and Red Hat founded the working group, with ABBYY and HumanSignal contributing, and it runs under a vendor-neutral open governance model (KMWorld). No single vendor controls where it goes, so you're not betting on one company's roadmap.
A photo of a recipe card looks fine to you. But type that recipe into a cooking app and suddenly the app knows "2 cups" is an amount, "flour" is an ingredient, and step three comes after step two. It can double the servings. The photo can't do any of that. The PDF is the photo. DocLang is the app version.
Supporters like to compare it to HTML, which gave browsers a common way to interpret web content (Unite.AI). The spec is at version 0.6 under an Apache 2.0 license, and governance metadata like privacy flags and model training restrictions live inside the document itself instead of in a separate file (It's FOSS).
We got tripped up here at first. You don't throw your PDFs away. The PDF stays as the record. DocLang is a copy made for the machine, like a translation you keep on hand for a colleague who doesn't read your language.
Every AI setup that grounds itself in company content already has a step where something reads your documents and preps them. It's invisible, and it's usually rough. DocLang is a better output for that step.
And it happens ahead of time. You chop the onions on Sunday, not while your guests are waiting at the table. Conversion runs in the background, not every time someone asks a question.
It also doesn't replace things like Microsoft Graph. Graph is the librarian who knows which shelf the book's on and whether you're allowed to read it. DocLang makes sure the book isn't written in smudged pencil. You want both.
At capture. When paper gets scanned, the OCR engine can produce DocLang right alongside the searchable PDF. ABBYY's FineReader Engine 12.8.0 already exports to DocLang (ABBYY). Structure gets captured once, at the moment the engine understands the page best, instead of being reverse-engineered later.
Inside the content repository. A background service in your document management system converts new and changed files as they land and keeps that version in an AI-facing index. Users never notice. Every AI tool pointed at that index reads clean content, and the governance flags ride along, so a "don't train on this" rule doesn't get lost in transit. For records-heavy public sector orgs, that's a big deal.
At the indexing step of a custom AI build. If you're building your own retrieval pipeline, you swap the crude PDF-to-text step for a DocLang converter. IBM's open source Docling toolkit can already parse DocLang files and serialize documents as DocLang for the LLM (PyPI). In one demo the DocLang version answered in 2.7 seconds versus 4.2 for the PDF, and the AI reading the PDF missed a subsection and botched a table (The Register).
Between systems. When AI agents hand documents to other agents, a shared format means nobody re-parses at every handoff. This one's the most speculative, and it's where a shared standard would help most.
It's early. IBM's own researcher admits adoption will make or break DocLang. The benchmarks so far are vendor-run, and Microsoft hasn't said a word about Copilot support. Analysts are supportive but point out that organizations will need controls in place to scale it responsibly (InfoWorld).
Still, we've spent years telling clients their AI is only as good as the content they feed it. This is the first serious attempt to fix the content side at the format level. If it catches on, a lot of AI accuracy work starts to look like content management. That's our home turf.