Is legal AI secure enough for client files?
Your firm already sends privileged documents to an American technology company every day, through SharePoint. The reasons that was acceptable apply to legal AI too.
Cloud storage and legal AI
When a firm moved its document management to Microsoft 365 or its practice system to a hosted platform, it made a considered decision: client material would live on infrastructure owned by a foreign company, processed in data centres the firm would never visit, under terms the firm did not draft. Risk committees worked through it. Residency was checked, contracts were negotiated, and the profession concluded, correctly, that the controls were sufficient.
Legal AI, for the most part, runs on exactly that infrastructure. When a document reaches a well-built legal AI product, it is stored in the same class of cloud storage the firm already uses and processed by a model deployment running in the same class of cloud region. The storage activity is the activity SharePoint performs. The processing activity is the activity every server-side application performs.
Whether the cloud is safe for client files is a question your firm answered a decade ago.
So the useful question concerns the particular product: what is different about it, and is the difference fenced?
How AI processing differs
Three things distinguish AI processing from the storage and retrieval your firm already accepted. A vendor who dismisses them is telling you something.
- The content is read, not just held. SharePoint stores a deed; an AI product reads it end to end. That is the point of it, but it means the substance of a matter passes through a system that has to be trusted with substance, not just with bytes.
- It may be retained by a third party. Model providers have their own retention behaviour, often for abuse monitoring, and it is separate from the vendor’s own storage. This is the layer firms most often fail to ask about.
- It may improve the vendor’s product. Consumer AI services have trained on user content by default. Enterprise deployments generally do not. “Generally” is not a standard you can rely on; a clause is.
The bans most firms imposed on public chatbots were a correct response to all three at once. A free consumer service reads the content, retains it on terms that change without notice, and may train on it. Nothing about that is comparable to SharePoint, and the policies were right to say so.
Controls to ask for
AI processing carries a risk profile close to your existing systems when a set of ordinary controls is in place and written down.
- Enterprise inference under a data-protection agreement. Cloud providers offer AI deployments where prompts are not used to train foundation models and the provider’s data-protection terms apply. This is the single largest difference between a legal AI product and a chatbot, and it is invisible from the outside, so you have to ask.
- Isolation you can describe. Ask whose model endpoint your matters reach. A dedicated subscription per firm is stronger than a shared endpoint; a shared endpoint with enforced tenant scoping is stronger than a pooled consumer API. Any of these can be defensible. What is not defensible is a vendor who cannot tell you which one you are getting.
- Residency stated as a commitment. Where data is stored and where it is processed are different questions. Transient processing may cross a border even when storage does not. A serious vendor distinguishes them without being pushed.
- A published subprocessor list. Every vendor has a supply chain. The good ones publish it and give notice before changing it.
- Retention you set, and evidence it ran. A retention policy without a log proving deletion occurred is an intention.
Where privilege sits differently from security
Security and privilege are not the same test, and legal AI is one of the places the gap shows. A system can be secure in every ordinary sense, encrypted, access-controlled, audited, and still create a problem if client material has been disclosed to a third party in a way that is hard to characterise as confidential.
In practice, the analysis usually turns on the same things that make a cloud DMS acceptable: the provider is a service provider acting on the firm’s instructions, bound to confidentiality, without an independent right to use or disclose the material. That is a contractual posture, not a technical one, and it is why so much of this comes back to the agreement rather than the architecture.
It is also why the vendor’s own terms matter more than their security page. A page can say anything. A clause has to be honoured.
A practical standard
If a firm wants one test to apply to any legal AI vendor, we would offer this: ask for every security claim in the contract, and see what survives.
Vendors are rarely dishonest. But marketing pages are written to be reassuring and contracts are written to be true, and the distance between them is exactly the risk you are trying to measure. A vendor whose contract says materially less than its website has told you something useful at no cost to you.
We publish our Terms & Conditions, Data Processing Agreement and subprocessor register openly, before anyone signs anything, for that reason. We do it because a claim we would not put in the agreement is not a claim worth making.
Hold the category to the standard you already set for the cloud. Then hold each vendor to their own contract.
The controls in full are on the security page. Questions go to admin@solicia.ai. A person reads them.
Writing your firm’s AI policy? Get the AI use policy template for Australian law firms, free in Word.
Get the next briefing by email.
New briefings by email, a few times a year.
Check your inbox.
We have sent a confirmation to . The next briefing follows once you have clicked it.