Every popular AI tool is a website. You type into it, the text goes to a data center, an answer comes back. For most questions that's fine. For a draft of a client's contract, a patient's chart, or a client's tax return, it means the most sensitive document in the building just left the building, under terms you didn't read.
Most firms respond with a policy that says "don't use AI on client files." What that produces is staff using it on their phones. A private AI server is the other answer: give them a tool that's as easy, and keep it inside.
What it actually is
A server, usually tower-sized, with a lot of memory and an AI accelerator card, running open language models. It sits on your network next to the file server (or is the file server). Staff open a page on the office network and get a chat window, a document summarizer, and a search box that reads your own files and answers in plain language with the source attached.
The models are the open ones that can be downloaded and run anywhere. They're smaller than the biggest public models and a step behind them on hard reasoning. They are very good at the things an office actually does all day.
What it's good at
- Summarizing. A 60-page deposition, a discovery dump, a year of correspondence, in a paragraph with page references.
- Drafting from your own templates. An engagement letter, a treatment plan explanation, a client update, in your firm's voice, because it's reading your firm's documents.
- Search that understands the question. "Which of our leases let the tenant sublet?" across every file you have, not just the ones with the word "sublet" in them.
- Transcription. Meetings, dictation and calls, turned into text that never leaves the room.
- Forms and intake. Pulling the fields out of a scanned document into your system.
What it isn't
- The newest public model. If your work needs the most capable AI on earth today, that's a cloud product, and you'll pay in privacy.
- Magic. It reads what it's given. If your files are a mess, it gives you organized answers from a mess.
- Free of maintenance. Models improve every few months; someone has to test and swap them. That's what the care plan is for.
Who it's for
Law firms, where a leaked document is a malpractice claim. Medical, dental and veterinary practices, where it's a reportable breach. Accountants during the season. Property managers with tenant files. Anyone whose insurer's cyber questionnaire asks "do you use AI tools, and where does the data go?" and who would like to write a short, true answer.
Ours start as a tower with a 16 GB accelerator and 64 GB of memory, which handles a small team comfortably, and go up to 96 GB cards for heavy use. See the AI server, or build one.
Written by Nikita Huttenlocher, founder and lead engineer at Huttenlocher Technologies, a web design and software company in Myrtle Beach, South Carolina.