Why private AI?

Keep Copilot or Gemini for mail, slides, and work that does not include client facts. That is a sound way to run the office. It is not the same as putting a client file through a model the firm does not operate. Private AI is the other purchase: confidential client work, on premises, under the firm’s control.

1Where the file is read

When someone pastes a client email or a contract into Copilot or Gemini, that text is processed in the vendor’s AI stack, under the firm’s tenancy terms. Encryption, tenant isolation, and a promise not to train on prompts can all be true. None of that means the model read the file inside the firm.

For a small or mid-size practice this is sharper, not softer. There is usually no specialist team to argue about whose cloud it was. There is a handful of clients, and they hired the firm.

The settling question is this: at the moment the assistant reads this file, where is that data, and who operates the system that can be compelled to produce it? “It is included in our package” is a procurement answer. It is not an answer to that question.

2The work that would actually help

Bundled AI is good at generic productivity: rewrite this email, summarise this public paper, tidy these notes. The high-value work is the opposite: this client’s file, this tax pack, this matter, this engagement.

If the policy is “do not put that in Copilot,” then the included assistant does almost nothing for the work that would give the firm capacity. Staff either leave confidential work unaided, or they use the convenient tool anyway because it is already open and already paid for. A ban without an alternative is not a policy.

Private AI exists so there is a firm-approved way to use AI on that work: authorised files, under the firm’s control, with a person still responsible for what leaves the building.

3The firm stays responsible

If a file is mishandled, the client does not ring Microsoft or Google. They ring the partner, the principal, the accountant they already trust. Professional duty, confidentiality, and the relationship stay with the firm, whichever logo was on the chat box.

That is why this is worth buying even at small scale. The firm is not buying a bigger model. It is buying the ability to say, calmly, to a client, an insurer, or a regulator: confidential processing stays under our control. “We use Copilot, it is in the package” does not give you that sentence. For a small practice, that sentence is commercial. One nervous client is a material part of the book.

4Three more, said quickly

You keep the package

Mail, files, and general AI stay where they are. Private AI is for the confidential subset: one or two workflows, named reviewers, a controlled environment. That is a smaller decision than replacing the IT estate, which is what people hear when they think you are attacking Copilot.

Drafting is not the same as inspectable work

Copilot drafts. Private AI, done properly, starts from authorised material and approved sources, shows what it used, and expects a human to review before anything becomes advice or a client communication. Grounding does not eliminate errors. It does make the output challengeable. For professional work, that difference matters more than which model is smarter.

Small firms get the same button, with less room for error

A large network can fund seats, training, and a crisis team. A ten-person firm usually cannot. They get the same convenient button, fewer people to police it, and a book of business concentrated in a few relationships. Private AI is not an enterprise hobby to grow into later. It is how a small firm adopts AI without renting the part of the work that has to stay theirs.

The longer ownership argument is Sovereign AI, or a rented assistant?. The products that do this work are Secure Law and Secure Accounting.