How we built AI document reading into Casemind
The design choices behind Casemind's Read document with AI feature: a quote for every value, visible confidence, a deadline check, and nothing saved without a lawyer.
Casemind is our practice-management product for law firms in Kosovo. Lawyers there receive judgments, rulings and summons as PDFs or photos, mostly in Albanian and sometimes in Serbian or English. Each one means typing the case number, court, decision level and dates into the case file, then working out the appeal deadline.
We built Read document with AI to do the reading. These are the choices that mattered.
Suggestions, not changes
The AI never writes to a case. It returns suggestions; the lawyer ticks the ones to keep, edits them if needed, and saves. Until then, nothing in the case changes.
A quote for every value
Each suggested value comes with the passage it was taken from. A lawyer can check "8 October 2026" against "Pejë, më 08.10.2026" in a second, without opening the PDF.
Confidence you can see
Every value is marked Sure, Check or Unsure. Dates and numbers that were hard to read get flagged, so attention goes where it's needed.
A second opinion on the deadline
Casemind calculates appeal deadlines from the case type and the decision date. The AI also reads the appeal period written in the document. If the two disagree, or the document counts the period from the day the decision is received, Casemind says so and asks the lawyer to check.
Controls for the firm
- A document is sent to Google's Gemini API only when someone asks for it.
- Partners can switch AI off for the whole firm.
- Interns can't use it.
- Each firm has a daily limit on documents read.
- Usage is logged as token counts only; document content never goes into logs.
What we'd do the same way again
Treat AI output as a draft. Show the source. Make uncertainty visible. Keep a person in charge of anything that matters.
Those rules apply well beyond legal work, and they're how we build AI features for clients too.