What text recognition used to be for us
Whenever you upload a scan or a photo of a document — a scanned contract, a photographed letter from a law firm, a handwritten note on a form — the platform first has to turn the image content into real, searchable text before any AI analysis can even begin. Until now, this was handled by a general-purpose AI chat model with the simple instruction to "transcribe this image faithfully." That works surprisingly well in practice — but it is technically not the same as a product actually trained for text recognition.
The difference: a purpose-built tool instead of a workaround
We are now rolling out, step by step, direct access to a dedicated OCR model — a product whose only job is text recognition, not holding a conversation. That sounds like a small technical nuance, but it has two concrete consequences.
First: no translation risk, structurally. A chat model interprets an instruction like "transcribe this" within the context of a conversation — a language model that sees a foreign-language document could, in rare cases, misread the instruction as a request to translate rather than transcribe. A dedicated OCR model has no concept of such an instruction in the first place — there is no "conversation" for it to misinterpret, only text being extracted from an image.
Second: reliability we verified ourselves before rolling it out. We ran the new path through a test document full of French and German special characters (é, à, ç, ê, œ, as well as ä, ö, ü, ß) before switching it on. Every single character came back exactly right — no encoding errors, no silent substitutions.
Handwriting: better, but not "solved"
This is where it gets honestly interesting, because a common use case for us is exactly this: a handwritten note on an otherwise printed form, a handwritten will, a handwritten addition to a contract. The model provider itself describes this new model as its best result on handwriting to date. We don't want to withhold that from you — but we also don't want to oversell it.
Independent industry benchmarks show a consistent picture: handwriting recognition is meaningfully harder than printed text across virtually every provider, and there are specialised offerings on the market that score even better on handwriting alone than a general-purpose OCR product. What this means for you in practice: a handwritten note is now recognised more reliably than before — but a document where a handwritten passage is legally decisive (the handwritten line under a will, a handwritten addition to a contract) should still be checked against the original before you rely solely on the AI transcript.
Where this applies
The new path applies everywhere text recognition already ran: uploads in the client portal, scanned case files in the lawyer portal, and documents you upload directly as a consumer — including the anonymous quick check on the homepage. Only images and scans without their own text layer are affected; a normal, digitally created PDF is read directly anyway, without any text recognition needed.
Who pays for it
For law firms, the OCR cost flows into the same AI quota that applies to any other analysis — a regular debit, no separate billing. For private and business users, nothing changes: text recognition is part of the relevant analysis action and its already displayed fixed price, not an extra cost item.
Gradual, not all at once
We're deliberately not switching this path on for everyone at the same time — we're observing it in live operation first. And because document content is additionally transmitted to a further external provider for this, that provider is listed in our subprocessor directory — transparency about who sees your data, and why, is not optional for us.
From the same piece of work: the relationship graph, now for you as a consumer too
In the same round of work, we rolled out a second, independent improvement: the relationship graph — which shows how the facts the AI recognised in a case relate to each other, including detected contradictions between two documents — is no longer limited to the lawyer portal. It's now available to you as a consumer, too. Open a case file and you'll find the same connections view there from now on, either as a searchable list or as a visual network.
