AI features users actually need

Problem

AI feature requests often arrive as vague ambition: “add AI to the product.” Without a concrete desk failure, teams ship chat widgets and summarizers that practitioners ignore. In legal workflows, ignored AI is harmless; trusted-but-wrong AI is dangerous. The design problem is identifying repetition experts tolerate but should not have to perform.

Context

IP firms already make high-stakes judgments — conflict analysis, renewal strategy, client advice. They do not need a model to replace that. They need relief from translation work: bulletin PDFs to register rows, register rows to matter context, scattered documents to searchable records. AI earns adoption when it removes steps practitioners already resent.

Constraints

Wrong-but-confident output creates liability. Firms will abandon features that cannot show sources. AI must attach to matters, auth, and audit — not a floating chat widget.

Architecture

Retrieval-first. Document intelligence as a gated pipeline into D1. Generation only scaffolds drafts or field suggestions with confirmation. No auto-write into client matters.

Implementation

R2 → OCR → field mapping → review queue → merge. Confidence gates every field. Corrections logged to improve parsers. Search and reminders consume verified data only.

Result

Bulletin intake saves hours without invisible magic. AI disappeared into workflow — acceleration, not a sidebar “AI mode.”

What I Learned

If you cannot name the manual step being eliminated — and show the source when the machine guesses — you are building a demo, not a feature.