Guides
What on-premises AI changes for clinics, law firms and schools, and where it stops
A plain guide for law firms, clinics and schools on what on-premises AI changes, what stays your job, and where Offline Base stands in October 2026.
Our page for professionals (opens in a new tab) opens with a blunt line: "You legally can't paste that into ChatGPT." The fuller version is narrower. The legal trouble starts when a patient's notes or a client's file go into a consumer chatbot with no business associate agreement (BAA) and no client consent. Some enterprise AI plans do come with a BAA, so the case for local AI is fewer vendors and fewer agreements.
In a law firm, a clinic or a school, the confidential document is usually the work itself. A client memo, a patient's intake notes and a student's records are the files people most want help with, and the ones they should least hand to a public chatbot.
On-premises AI means the model runs on a machine in your building. That changes which companies see your files. Your legal duties stay where they were.
What keeping AI in the building changes
When a model runs on hardware you own, your prompts and files are processed on that hardware. Our terms (opens in a new tab) say we never receive them, because there is no account, no telemetry and nothing to send.
The professionals page is written for three kinds of office:
- Law firms (client confidentiality). The page pitches drafting, summarizing and searching privileged material on a box in your office, so client files would stay inside the firm.
- Clinics (patient trust). Notes and intake can be handled on hardware that stays in the building, with fewer outside vendors touching what patients tell you.
- Schools and districts (FERPA). Student records and student work stay on site, which gives parents and boards a clear answer about where they are.
The common thread is a shorter list of third parties. Every outside service that sees your data is another company whose security, retention rules and own vendors you are trusting. A model on your own hardware takes that company out of the path for the work it handles.
What on-prem changes under HIPAA
For a clinic, on-prem has one narrow, concrete benefit. In our professionals page's words, it "reduces the business-associate agreements and sub-processor risk that come with sending protected data to a third party." The same page also says, "HIPAA doesn't actually mandate data residency, and we won't pretend otherwise."
In practice, the benefit comes down to two things:
- Each vendor that receives protected data is another BAA to sign, track and keep current.
- Each of those vendors has its own sub-processors (hosting, logging, support tools), and each is another place the data can be exposed or kept longer than you expect.
Run the model on your own box and, for that task, the AI vendor and its sub-processors drop out of the chain. Your own HIPAA obligations stay with you.
Where on-premises AI stops
Compliance frameworks. Offline Base products are not certified, accredited or audited under HIPAA, GDPR, FERPA, SOC 2, PCI DSS, CJIS, ITAR or any other framework.
No BAA or data processing addendum. Our terms say we do not currently offer either. If your counsel requires one from every AI vendor, we cannot be that vendor today.
Compliance stays your office's work. Our terms say running a model on your own hardware "does not by itself make any workflow compliant." Access control, retention, breach notification, staff training and vendor diligence remain your job.
You secure the machine. Our terms say we cannot protect data on a machine that is already compromised, shared, unencrypted or physically accessible to someone else: "Local means the data is yours to secure, and securing it is your job." A box under the reception desk needs the same locks, accounts and disk encryption you would give a file server.
Optional features narrow the promise. Outside help, web search and remote access send data off the machine when they are available and you turn them on. Our privacy page (opens in a new tab) says the moment you allow outside help or remote access, "data never leaves your machine" becomes conditional. As designed, outside help asks first, names what would be sent, and goes only to US-hosted providers under zero-data-retention terms. What zero data retention means, and where it stops explains how that term works.
Answers can be wrong. The terms warn that every model "can be confidently, fluently wrong," and that small models on your own hardware are more prone to it than the largest cloud systems. A summary of a patient note or a contract still needs a person who checks the source.
Where Offline Base stands today
As of October 2026, nothing we make can be ordered. The product meant for this kind of office is Base Router, an always-on box planned at a one-time $400 with no subscription. It is a preview, and our terms call every specification, price and date we publish "a current target." The hardware is covered in Base Router, the always-on AI box we are designing.
The professionals page lists the features below and tags three of them "In design." Our terms call activity records and local-only mode "design intent" that "may change or be dropped as we build."
Feature | Status |
|---|---|
Activity records for every question and approval, stored locally | Planned for Base Router, untagged |
Local-only mode, planned as the default, for offices that need no outside connection | Planned for Base Router, untagged |
AI prepared for your field: intake, summarization and search | In design |
Team accounts with role-based access | In design |
Optional secure remote access by user | In design |
A professional tier with bigger hardware, a support contract and compliance documentation | In design with pilot customers |
On compliance, the page says: "Talk to your own counsel, and we'll give you clear documentation to bring to them." Pilot customers are helping design the team features, and the page invites more offices to help shape a pilot.
Questions to put to any AI vendor
These apply to us as much as to anyone else. Ask them in writing and keep the answers.
- Where does our data go? Which machines, which companies and which countries handle a prompt or a file, from the moment someone types it.
- What gets logged, where, and for how long? Prompts, outputs, metadata and support tickets, on your side and on theirs.
- Who are the sub-processors? Ask for the list, and ask how you will hear when it changes.
- What is actually certified? Ask which framework, which product, who audited it and when, and ask to see the report.
- Will you sign a BAA or a data processing addendum? Get a yes or a no.
- What changes when optional features are on? Web search, cloud fallback, remote access and integrations each move data somewhere.
Our own answers, as of October 2026: Base Router is designed to keep prompts on the box, so we would receive nothing. Activity records would be stored locally. Outside help would go only to US-hosted zero-data-retention providers, after you approve it. Nothing is certified, and we offer no BAA or data processing addendum. If those answers do not fit your obligations, you should learn that before you buy from anyone.