Skip to content
Case studies

Systems, documented end to end.

Legal AI workflows with their architecture, operating decisions, available evidence and failure modes exposed.

  1. The reply-triage workflow canvas, including routing and failure paths.

    Reply triage and client routing

    Every inbound reply read, classified by intent, routed to the right end client and logged — with a failure path that keeps a lost reply from being lost silently.

    FeaturedClient engagement
  2. The scheduling-agent workflow canvas from reply to confirmed meeting.

    An agent that negotiates meeting times and books them

    Reads the reply, infers the prospect's timezone, checks four weeks of real availability, writes in the account owner's voice, and books the meeting end to end.

    FeaturedClient engagement
  3. The two-pass extraction workflow canvas from document intake to verified rows.

    Two-pass extraction, where the second pass checks the first

    A purchase order dropped in a folder comes out as validated rows — extracted by one model call, then checked by a second before anything is written.

    FeaturedOpen build
  4. One reporting layer across two platforms and every client account

    Daily collection from two sending platforms with incompatible vocabularies, normalised at the edge into a single schema — which is what made any of the other numbers knowable.

    Client engagement

The failure sections make the work useful.

They show where controls need to sit, what changed after testing and what still needs evidence.