There's a conversation we have with almost every healthcare practice, law firm, and public agency that calls us, and it usually starts with some version of the same sentence.
We'd love to use AI. We just can't send our data anywhere.
They're right to say it. A care provider's records are protected health information. A firm's files are privileged. An agency's data belongs to the public and comes with procurement rules that were written long before anyone was talking about large language models. The consumer version of AI — open a browser tab, paste in the document, ask a question — is off the table, and it should be.
What surprises people is how quickly the conversation changes once they realize that isn't the only way to run these systems.
The model can come to the data
Most people's picture of AI is a service somewhere else. You send it your text, it sends back an answer, and somewhere in between your text lives on someone else's servers under someone else's retention policy.
That picture is accurate for the consumer tools. It is not a law of nature.
Capable models can run on hardware you own, in a room you control, with no connection to an outside AI service at all. The documents never leave the building because the model is in the building. Nothing is sent, so nothing is retained by a third party, so the question of what a vendor does with your data has a very short answer.
We've built exactly this. For Potomac Mitigation, the agentic assistant their team calls Allie Bot runs locally on a dedicated Mac Mini, trained on the firm's own templates, doing contract and financial analysis on documents that were never going to be sent to a public cloud. The published result is in the case study: analysis that used to take hours now takes minutes, and the sensitive material stays where it started.
What private AI is good at
Private deployment isn't a lesser version of AI. For the work regulated organizations actually need done, it's often the better fit, because that work is mostly reading.
Summarizing and extracting. Pulling the relevant facts out of a long record, a contract, a case file, or an application — and putting them in the structure your team already uses.
Drafting from your own templates. First drafts of letters, reports, and forms that follow your language, not a generic model's idea of your language.
Answering questions against your own documents. "Where in this file does it say…" and "Which of these matters involve…" — questions that currently cost someone an afternoon of scrolling.
Routing and triage. Reading what comes in and sending it to the right person with a short summary attached.
None of that requires the internet. All of it benefits from the model having been set up around your documents and your rules.
What changes in the design
Keeping data in-house is the headline, but it's not the whole design. Regulated work needs a few other things built in from the start, not added after.
Access follows your existing rules. If a paralegal can't see a matter today, the AI shouldn't surface it to them tomorrow. If a caregiver only sees their own clients, so does the assistant. Privilege and minimum-necessary access are part of the build, not a policy memo.
Everything is logged. Who asked what, against which documents, and what came back. When someone asks how a draft was produced, you should be able to answer.
A human signs. The assistant drafts and extracts. A person reviews and decides. That boundary is the same one we write into every AI employee's job description, and in a regulated setting it isn't optional.
Retention matches yours. The system keeps what your retention policy says to keep and nothing else.
Our healthcare work is built on HIPAA practices from day one, and our government work follows SOC 2 practices and deploys on-premises when procurement requires it. Those aren't add-ons. They're the starting assumptions.
The honest trade-offs
Private AI costs something the consumer tools don't: hardware, setup, and someone responsible for keeping it running. The very largest models only run in large data centers, so a private deployment uses a model sized for the job rather than the biggest one available — which, for extraction, drafting, and triage, is usually plenty.
And it doesn't make every project a good idea. The same test applies here that applies everywhere: price the problem first. If the reading your team does each week isn't worth meaningfully more than the build, the math says wait, and we'll say so. That discipline is the point of the 2× guarantee — we only take on work we can defend on a spreadsheet.
Where to start
Pick the one document-heavy task your team does every week that they'd happily never do again — the intake packet, the record summary, the contract review checklist. Write down how many hours it takes. That's your candidate, and it's usually a reading problem, which is the kind of problem this handles best. We wrote more about that category in AI That Reads Your Email.
If you want to see how we build these systems, the Agentic AI Solutions page covers the approach. If you want to know whether your workflow is worth it before anyone builds anything, start with the AI Automation Scorecard.
"We can't send our data anywhere" isn't the end of the AI conversation. It's the first requirement.