What to Ask Before You Let AI Near a Client File

Every firm we speak to is being sold AI. The pitches arrive weekly, they demo beautifully, and they are mostly delivered by people who have never had a professional duty of confidentiality in their lives.

The useful question is not whether to adopt AI. It is which specific jobs it genuinely helps with, and which ones it quietly puts at risk. These are the questions we ask on a firm's behalf.

1. Where does our data go, and who else can see it?

Ask for this in writing, and ask specifically whether inputs are used to train models. A vendor who cannot answer plainly, or who answers by describing their encryption instead, has told you something.

Encryption in transit is not the question. The question is what happens to the content after it arrives, who at the vendor can read it, what a subpoena to the vendor would produce, and what happens to it if the company is acquired or folds.

2. Is this a capability, or a wrapper?

A great deal of what is sold to law firms as AI is a thin interface over a general model, priced as though it were the model. That is not automatically bad, and sometimes the interface is the value. But you should know which you are buying, because it determines what you are paying for and how easily it can be replaced.

A fair way to ask: what does your product do that we could not do by pasting this into a general assistant ourselves? A confident vendor has a good answer.

3. What happens when it is confidently wrong?

Every one of these systems produces output that is fluent, plausible, and occasionally false. That is not a bug being fixed next quarter; it is a property of the technology.

So the question is about the workflow, not the model. Who checks the output, at what point, and what is the cost if a wrong answer gets through? A summarisation tool whose errors are caught in a two-minute read is a different risk from one whose output goes into a filing.

4. Can it run somewhere we control?

For some jobs the answer needs to be yes. Models that run on your own hardware are meaningfully more capable than they were even a year ago, and for a firm whose main hesitation is confidentiality, local deployment turns an unacceptable tool into an acceptable one.

It is not always the right answer. It costs more, it is slower to set up, and it will lag the frontier models in raw ability. But for anything touching client material it deserves a genuine look rather than a reflexive dismissal.

5. What does it cost to stop?

If your data is in their format, on their servers, and your staff have built two years of habits around their interface, the renewal conversation is not really a negotiation.

Ask what export looks like before you sign, not after.

The Answer Is Sometimes Nothing

The most useful outcome of this exercise is often that a firm buys nothing. A consultant paid to recommend technology will always find technology to recommend, which is a good reason to be careful about who you ask.

We do not resell anyone's AI product, which means we have no reason to tell you that you need one. If it would help to have someone read the same pitch you are reading and tell you which of the five questions it fails, that is what our consulting work is for.