In most of the businesses we work with, the question is no longer whether staff should use ChatGPT, Claude, Copilot or Gemini. They already are. Someone is drafting emails with it, someone is summarising a long contract, someone has pasted in a spreadsheet of customer complaints to find the pattern. Usually on a personal account, usually without anyone in management knowing.
Most of that is useful work, and banning it outright tends to push it further out of sight. But some of it is a POPIA problem, and as the employer, it’s your problem rather than theirs.
This is a practical guide to where the line sits, written for owners and managers rather than lawyers.
The short answer
Yes, your staff can use AI tools at work. What POPIA cares about is personal information: anything that identifies a living person or an existing company, from a name and email address to an ID number, a CV or a customer’s complaint history. The moment someone pastes that into an AI tool, your business is sharing it with a third party, and the Act’s conditions apply.
South Africa has no AI-specific law yet. That doesn’t leave a gap: POPIA covers what goes into these tools just as it covers what goes into your CRM or your payroll provider.
Where POPIA bites
You stay responsible for what staff share
Under POPIA your business is the “responsible party” for the personal information it holds. Section 19 requires you to keep it secure, with reasonable technical and organisational measures. A staff member pasting a client list into a free AI account is exactly the kind of disclosure those measures are meant to prevent, and it doesn’t matter that they meant well.
The provider is your operator, and that needs a contract
When an outside service processes personal information on your behalf, POPIA calls it an “operator”, and Section 21 requires a written contract that obliges it to keep that information secure. Business plans from the major AI providers come with terms you can point to: a data processing agreement, a commitment not to train their models on your data, admin controls. A personal account signed up with a Gmail address gives you none of that.
Most of these services sit outside South Africa
Section 72 restricts sending personal information to another country unless the recipient is bound by laws or agreements offering comparable protection, the person has consented, or one of a few other grounds applies. The large providers address this in their business terms. It is one more reason the account type matters.
Some information needs much more care
Health, religion, race, union membership, biometrics, criminal records and information about children are “special personal information” under the Act, with stricter rules. As a rule of thumb, none of it should go into a general AI tool at all, on any plan, unless a system has been set up deliberately to handle it.
Decisions about people can’t be left to the machine
Section 71 says a person can’t be subjected to a decision with legal or substantial effect based solely on automated processing that profiles them. If AI is ranking CVs, scoring customers or flagging staff performance, a person has to make the actual decision and be able to explain it.
If something leaks, you must say so
Section 22 requires you to notify the Information Regulator and the people affected, as soon as reasonably possible, when personal information has been accessed by someone who shouldn’t have it. An AI account compromised with months of pasted client data in its history is a notifiable incident.
What to do this week
You don’t need a project to get this under control. You need four things.
- Pick the tools you allow, and pay for business accounts. One or two tools, on a business or team plan, signed up with company email addresses. Check the provider’s terms say your data isn’t used to train their models, and keep a copy of their data processing terms with your POPIA records.
- Write a one-page policy. There’s an outline below. Short enough that people actually read it.
- Spend twenty minutes with each team. Show them what’s fine, what isn’t, and how to take the names out of something before they paste it. Real examples from their own work land better than any slide.
- Ask what’s already being used. Without blame. You’ll find out where AI is genuinely saving time, which is where to invest next, and which accounts to close.
A one-page AI policy you can adapt
Copy this, change what doesn’t fit, and put your information officer’s name on it.
Using AI tools at [Company]
Approved tools: [Tool] on our business account, signed in with your work email. Don’t use personal AI accounts for work.
Fine to use them for: drafting and editing, summarising documents that contain no personal information, research, brainstorming, writing code, and working with information that’s already public.
Never put into any AI tool: ID numbers, bank or card details, passwords, health or medical information, anything about children, staff records, or client files, unless a system has been approved for that specific purpose.
Take the names out first. If you need help with something that mentions people, replace names, contact details and account numbers before you paste it in.
Check what comes back. AI tools are confidently wrong from time to time. You are responsible for anything you send, sign or publish.
People decide about people. AI can help organise information about candidates, customers or staff, but a person makes the decision and can explain it.
If something goes wrong, for example personal information pasted into the wrong tool, tell [information officer] the same day. We’d rather know.
Questions: [information officer], [email].
Beyond the policy
A policy keeps you on the right side of the Act. It doesn’t make the tools useful.
In our experience the gap between “everyone has a subscription” and “something has actually changed in how the work gets done” is where most businesses are stuck right now.
The step that usually closes it is to take the knowledge people keep asking each other for, such as policies, procedures, product information and pricing rules, and build an assistant that answers from those documents alone, inside a system where you control who sees what and where the data lives. That’s a design question as much as a technical one, and POPIA is part of the brief from the first day rather than something checked at the end.
If you’re not sure where your business stands, bring one process or one document your people struggle with to a 30-minute call. We’ll tell you plainly where AI would help, and where it wouldn’t.
Sources: the Protection of Personal Information Act 4 of 2013 (sections 19, 21, 22, 26, 71 and 72), and the Information Regulator. This article is general guidance, not legal advice.
Grant SlabbertDesign and development lead, JuneDay · About us
AI consulting at JuneDay