Data you can't share, and data you'd rather not
Two reasons an office keeps its AI in-house, and why most firms have both.
Every office that uses the frontier AI models, the big cloud chatbots, is sending its text to a company abroad and paying for the privilege. Most of the time that's fine. A recipe, a cover letter, a question about Excel. The trouble starts when the text is someone else's.
The first kind is data you can't share. A lawyer's client files are covered by professional secrecy. A doctor's notes are covered by patient confidentiality. A bank's KYC files are covered by banking rules and by the bank's own compliance team. Pasting any of these into a cloud chatbot means handing them to a third party in another jurisdiction, under terms of service that can change. For many firms that's simply not allowed, even if nobody has written the policy down yet.
The second kind is data you'd rather not share. Nothing forbids it, but you'd still prefer it stayed in the building: tender prices, salaries, board minutes, a draft of the strategy you haven't announced. Staff paste this kind of thing into chatbots every day because it's convenient, and the firm usually finds out later.
Most offices have both kinds, and they're mixed together in the same folders. That's why a policy of "don't put sensitive things into the chatbot" rarely works. People can't tell where the line is, and the tool is right there.
A local setup removes the question. The model runs on a machine in your office. Staff get the same kind of assistant they're used to, and everything they type, upload or ask stays on that machine. It isn't sent to an AI company and it isn't sent to us.
That doesn't mean the cloud is useless. Some jobs need the biggest models available, and some data is harmless. A sensible setup keeps sensitive work local and lets the firm decide, deliberately, what else may go out.
If you're unsure which kind of data your office handles, start with a list of the ten documents your staff open most often. In our experience at least half of them belong in the first group.