The worry is reasonable
When someone pastes a client contract into a free chatbot, a fair question follows: where did that go, and could it turn up in someone else's answer? For a law firm, an accounting practice or an engineering firm, that isn't a technical curiosity. It's client confidentiality, and sometimes it's the law.
The good news is that a business doesn't have to choose between useful AI and keeping its information private. The answer depends on three things: which version of an AI service you use, how the assistant is built, and where your data is stored. All three can be chosen deliberately.
Business AI services and model training
Consumer apps and business services from the same AI company often come with different terms. The business versions from the major providers are explicit about training:
- OpenAI says that, by default, it does not use data from its business products (such as ChatGPT Enterprise and ChatGPT Business) or its API platform, including inputs and outputs, to train or improve its models.
- Anthropic's commercial terms, which cover its API and business plans, state that it may not train models on customer content from those services.
Two cautions. Terms change, so read the current version rather than a summary, including this one. And these commitments apply to the business products: an employee using a personal account is under different terms. That's why we build on business AI services whose terms keep your data out of model training, and go through those terms with you before anything is connected.
How a private assistant actually works
A private assistant doesn't learn your business by being trained on it. It looks things up, the way a new colleague would search the shared drive. When someone asks a question:
- It searches your own sources, such as shared drives, email, your CRM, past quotes and contracts, for the passages most likely to hold the answer.
- It only searches what that person can open. If someone couldn't open a document themselves, the assistant won't answer from it.
- It answers from those passages, in plain language, and links every answer to the document, email or record behind it.
- It says when it can't find support, instead of guessing, and flags what a person should check before acting.
Because every answer cites its source, the assistant is also easier to trust: anyone can check it in one click. We explain this in more detail on our AI assistants page.
Where your data lives
"Private" should mean you know exactly where each piece of your data sits. An assistant has more parts than people expect: the original documents, a searchable index built from them, the questions people ask, the answers, and the logs. Each needs a home and a retention rule.
For clients that need it, we can deploy in your own cloud tenant or data center, so the documents, the index and the logs stay inside an environment your team already controls and audits. Wherever it runs, each client's project data is kept separate, and client data is never used to train a model. For export-controlled work, we design the setup with your compliance team, so the decisions about where data can go are made by the people accountable for them.
Permissions: the part most often skipped
The quickest way to leak information internally is an assistant that can see everything. If it was indexed with an administrator's access, a junior employee can ask about salaries or a partner's private client file and get an answer. A private assistant should follow the permissions you already have: the same groups, the same shared-folder rules, and the same removal of access when someone changes roles or leaves. Test it with real accounts before launch, not just with an administrator's.
What to ask any AI vendor
Whether you work with us or anyone else, these questions separate a private assistant from a chatbot with your files attached:
- Which AI service and plan does it use, and do those terms keep our data out of model training?
- Where are our documents, the search index, questions and logs stored, and for how long?
- Can it run in our own cloud tenant or data center?
- Is our data kept separate from other customers' data?
- Does it follow our existing permissions, and how is that tested?
- Does every answer show its source, and what happens when it can't find one?
- What is recorded about who asked what, and who can see those records?
- If we stop using it, what happens to the index and the logs?
A vendor who can't answer these clearly, in writing, isn't ready to hold your clients' information. For how we approach the rest, see our security work.
How we set one up
We start in discovery, which is free: we map where your answers live today, who should see what, and where the data must stay. Then we connect the sources, follow your existing permissions and try it on real questions from your team, with every answer showing its source. If you'd like to see what that would look like for your firm, book a free discovery call (opens in a new tab), or read where to start with AI first.