The prototype was scrappy. I built it with Base44 to validate one question: could an AI read estate documents accurately enough to skip the form altogether? I demoed it live with 10 of our clients.

It extracted up to 85% of the relevant data. That was enough to earn the following quote from a paying client.

That was the moment we had the conviction to build it.

Vale's AI extracts up to 90% of a matter from estate documents. The other 10% can be a misread date of death, an executor name spelt wrong, or a beneficiary missed. In estate law, mostly-right is not safe enough.

Three design decisions ensured the lawyer remained the author of the matter:

1. The matter isn't created until the lawyer confirms it. A “Create Matter” CTA at the end. No auto-save, no auto-submit.
2. Read-only by default; editing requires a click. We assumed editing would be the exception, not the rule. At 90% accuracy, most extractions are correct, so lawyers can scan without fear of accidentally changing a value.
3. Sticky sidebar to jump between sections. Six sections is a lot. Lawyers should be able to go straight to what they care about, not scroll through everything.

Trust over autonomy.

At launch in December, the team marketed AI extraction as one of Vale's main value propositions. Since then, 86% of new matters have used the feature.

The current focus is making the AI better. We're closing the remaining gaps in extraction accuracy and broadening what it can pull from each document. Each improvement means less for the lawyer to review.