GermanyPublic sector2026 – ongoing

European Patent Office

AI transcription of patent documents into ST36 XML

The European Patent Office's tailored OCR system processed 400,000 pages per day, Mistral reported

WorkingVendor case studyOfficial ↗Vendor ↗

Problem

Standard OCR struggled with formulas and tables in multilingual patent applications. The office needed structured output compatible with ST36 XML.

Approach

EPO and Mistral jointly fine-tuned an OCR model on patent documents. They built a conversion pipeline that produced ST36 XML and integrated the solution into EPO workflows.

Results

400,000pages transcribed per day, according to MistralVendor case study[1]

<1%character recognition error rate reported by MistralVendor case study[1]

Takeaway

The project combined recognition of technical content with conversion into the office's required document format.

Sources

From PDFs to structured output: automated patent transcription at scale.
Mistral AI2026Vendor publication
EPO adopts cutting-edge AI solution developed in Europe
European Patent Office11 Mar 2026Company publication

Source summaries

[1]Mistral AI2026

Mistral described a tailored OCR model and a pipeline that converted patent documents into ST36 XML. It reported throughput of 400,000 pages per day and a character recognition error rate below 1%.

[2]European Patent Office11 Mar 2026

EPO confirmed that it jointly developed the OCR solution with Mistral and integrated it into its workflows. It planned further refinement using operational feedback.

Full details are in the original sources.

Added 1 Nov 2026 · Updated 1 Nov 2026