Skip to content
All articles

25 September 2026 8 min read

AI and client confidentiality: what to ask before your firm uses any AI tool

A lawyer's duty of confidentiality does not pause when software does the typing. When a firm puts client information into an AI tool, it is still the firm's information to protect, and in South Africa the Protection of Personal Information Act (POPIA) says how.

This article sets out the parts of POPIA that matter most when a firm uses AI, and eight questions worth putting to any vendor before client information goes anywhere near their product. That includes us.

Who is responsible

Under POPIA, the firm is the responsible party: it decides why and how personal information about its clients, their counterparties and its own staff is processed. A software provider that processes that information on the firm's behalf is an operator.

That distinction carries real duties:

  • Section 19 requires the responsible party to secure the personal information in its possession or under its control with appropriate, reasonable technical and organisational measures. Handing information to a vendor does not hand over this duty.
  • Sections 20 and 21 require operators to process information only with the responsible party's knowledge or authorisation, to treat it as confidential, and to be bound by a written contract that obliges them to keep appropriate security measures in place. An operator must also tell the responsible party immediately if there are reasonable grounds to believe information has been accessed or acquired by an unauthorised person.
  • Section 72 restricts sending personal information to a third party in another country. It is allowed only on one of a set list of grounds, the most common being that the recipient is bound by law, binding corporate rules or an agreement that gives protection substantially similar to POPIA's.

Professional privilege and the rules of conduct sit on top of all of this. A tool that is lawful under POPIA can still be a poor home for privileged material.

Eight questions to ask any AI vendor

1. Where will our information be processed and stored?

Ask for countries, not just company names. Many AI tools run on infrastructure in the United States or Europe. That is not automatically unlawful, but it brings section 72 into play, and your firm should know which ground it is relying on. Some firms will also prefer local hosting for reasons of client expectation, even where the law does not strictly require it.

2. Will you sign an operator agreement?

Section 21 requires a written contract. If a vendor cannot produce one, or offers only website terms you click through, treat that as an answer.

3. Do you use our information to train your models?

This is the question that matters most with AI, and the answer has to be in the contract. Consumer and free tiers of many AI products reserve the right to use what you type to improve the service. A promise on a help page or in a sales email is not the same as a binding commitment in the agreement that covers your account.

4. Who else touches our information?

Most software is built on other software. Ask for the list of sub-processors (hosting, AI models, email, support tools), what each one does, and where each one is.

5. Who at your company can see our information?

Ask how access is limited and whether it is logged. "Our engineers can see everything" is a different risk from "access is restricted to named people, for support, and recorded".

6. How do we get our information out, and how is it deleted?

You should be able to export your data in a usable form and have it deleted when you leave, with confirmation. Ask how long backups are kept.

7. What happens if something goes wrong?

Ask how quickly the vendor will notify you of a security compromise and what they will tell you. Your firm may have its own duty to notify the Information Regulator and affected people, and it cannot do that without knowing.

8. Does the tool decide anything on its own?

This is not a POPIA question, but it is a professional one. South African courts have now sanctioned practitioners more than once for filing AI-generated material that was never checked. Prefer tools that propose and let a person decide, and make sure your own processes require that check.

Inside the firm

The vendor is only half of it. A few habits do a great deal of the work:

  • Write a short AI policy. Which tools are approved, for which tasks, and what must never be pasted into them.
  • Keep client files out of consumer chatbots. Free tools are convenient precisely because they are built for everyone, and their terms reflect that.
  • Check everything that leaves the firm. If AI helped produce it, a lawyer still signs it.
  • Tell clients where it matters. Engagement letters are a natural place to explain, in plain terms, how the firm uses technology.

How Atora answers these questions

We would rather be asked than assumed safe, so here is where we stand today.

  • Operator agreement: founding firms receive our operator agreement before any client information is entered.
  • Approval: Atora drafts time entries and a lawyer approves every one. It never records screens or keystrokes.
  • Where your data lives: we are planning the move of our hosting to a South African cloud region. Until that is done, we will tell any firm exactly where its information is processed, and by whom, before it signs.
  • Sub-processors: we will give you the full list on request.

If any answer here does not satisfy your firm, we would like to hear why. Write to sales@atora.co.za.

This article is general information about POPIA and professional practice. It is not legal advice. Take advice on your firm's own circumstances.

Sources