AI for professional services.

Billable hours lost to admin.

The work that pays is client work, and too much of the day goes to intake forms, file notes, meeting summaries and chasing documents. AI could take most of it, except client files are confidential and you can't hand them to a service that trains on your data.

We build assistants that answer from your own files, draft the routine correspondence and turn a recorded meeting into a clean summary. Where privacy matters they run on AI models installed on your own hardware, so nothing leaves the building. Intake, document requests and the paperwork trail get automated inside the systems you already use.

We run our own practice on exactly these systems. The cases page shows them with the logs to match.

See our own setup

Productivity · our own operation

AI bills, under control

AI bills grow when expensive models do cheap work. Our own operation runs on a router that sends every mechanical task, summarising, classifying, extracting, drafting, to free AI models on our own machine first, and saves the paid models for judgement calls. Every call is logged: 519 of them across the first nine days, including a 72,000-character job handled entirely on local models. At one measured checkpoint, roughly 320,000 tokens of work had been kept off the paid meter. A budget script reads the log and flags anything that should have been free but ran on a paid model.

See more work