AI can finally read your handwriting — here’s why enterprises care

Somewhere between the language models learning to code and the video generators learning to fake reality, a much less glamorous problem has been quietly piling up: the millions of PDFs, PowerPoints, and scanned forms sitting in corporate file shares that no AI system can actually read. Valantor, an enterprise visual intelligence company, is betting real money that fixing this is worth more than another round of model fine-tuning. On Tuesday it announced it acquired EyeLevel, a specialist in retrieval for unstructured documents, and rolled the combined tech into something it calls the Enterprise Visual Intelligence platform.

Benjamin Fletcher, EyeLevel's CEO and co-founder, puts the scale of the problem in blunt terms: about 80% of corporate knowledge lives in visually dense files — invoice stacks, claims paperwork, engineering drawings, contracts — that are simply too big and too messy for any LLM context window to swallow. Humans still do most of that processing today, and Fletcher says the results aren't pretty. Hand-built 'golden datasets,' the ones companies treat as ground truth, carry error rates of 10 to 25 percent by his team's own measurements. Yet those same organizations often demand near-perfect accuracy from AI before they'll trust it. That double standard is the quiet joke sitting underneath this whole announcement.

The actual product, GroundX, is designed to run inside a company's own walls — private cloud, on-prem, even fully air-gapped setups — rather than shipping documents off to some third-party API. Fletcher describes it as a single tightly-tuned ingestion-and-retrieval system exposed through REST APIs, SDKs, and MCP, with a Helm chart that drops into existing deployment pipelines. To keep costs and latency in check, GroundX never feeds a whole page to a language model at once. Instead a vision model breaks each page into elements — tables, paragraphs, handwritten notes — and small specialized agents process those pieces in parallel, which means processing time doesn't balloon linearly as documents get longer or more complex.

The handwriting angle is the hook getting attention, and it's a real one. E-ink tablets like the reMarkable and ViWoods AiPaper already dabble in AI-assisted note capture, but enterprise-scale handwriting recognition inside messy business documents has mostly stayed unsolved. Fletcher says EyeLevel's vision model, trained on more than a million pages of real enterprise documents, treats handwritten annotations as page elements with their surrounding layout intact, rather than stripping them out or ignoring them. Air France-KLM used the underlying platform to build a customer service assistant that reportedly hits over 96% accuracy on complicated policy questions, and AskVet leaned on it to automate resolution of up to 85% of customer inquiries using a decade of veterinary records.

None of this makes document management thrilling, and nobody involved is pretending otherwise. But the acquisition is a pretty clear signal about where the next wave of enterprise AI money is heading — not toward bigger models, but toward the unglamorous plumbing needed to make existing models useful on the data companies actually have.