POPIA and AI tools: what your staff can (and can't) paste into a chatbot

Somewhere in your business today, someone has probably pasted an email, a spreadsheet extract or a draft letter into an AI chatbot to tidy it up. Usually the intent is good: they want to work faster. The question is what was in that text, where it went, and whether your business can account for it.

The Protection of Personal Information Act (POPIA) does not mention AI chatbots by name. It doesn’t need to. Its principles apply to any way your organisation processes personal information, and pasting it into an online tool is processing.

This article is general guidance for South African business owners and managers. It is not legal advice. For decisions about your specific obligations, speak to a qualified legal adviser or your Information Officer.

What counts as personal information

POPIA defines personal information broadly. It covers information relating to an identifiable, living natural person and, where applicable, an identifiable existing juristic person such as a company. In practice that includes names, ID numbers, contact details, addresses, financial information, employment history, opinions about a person, and correspondence of a private or confidential nature.

The Act also sets out “special personal information”, which carries stricter rules. It includes things like health information, religious or philosophical beliefs, race or ethnic origin, biometric information and criminal behaviour.

Think about what a typical 30-person business handles in a normal week: client files at an accounting practice, patient notes at a medical practice, payroll records, CVs, HR disciplinary notes. Much of that is personal information, and some of it is special personal information.

Three parts of POPIA that matter most for AI tools

Security safeguards (section 19)

Section 19 requires a responsible party (your business, when you decide why and how information is processed) to secure the integrity and confidentiality of personal information in its possession or under its control. It does this by taking appropriate, reasonable technical and organisational measures to prevent loss, damage, unauthorised destruction and unlawful access or processing.

If staff paste client information into a personal chatbot account that your business does not control, has not assessed and cannot audit, it becomes hard to show that reasonable measures were in place.

Operators processing on your behalf (sections 20 and 21)

POPIA calls a third party that processes personal information on your behalf an “operator”. An AI service your business uses to process client data may well fall into this category. The Act expects operators to process only with your knowledge or authorisation and to treat the information as confidential. It also requires a written contract between you and the operator that ensures the operator establishes and maintains appropriate security measures.

A business agreement with a reputable provider, with data processing terms you have actually read, is very different from an individual clicking “I agree” on a free consumer app.

Cross-border transfers (section 72)

Most popular AI services process data outside South Africa. Section 72 restricts transferring personal information to a third party in a foreign country unless certain conditions are met. These include the recipient being subject to a law, binding corporate rules or a binding agreement that provides an adequate level of protection, the data subject consenting, or the transfer being necessary for a contract with the data subject, among other grounds.

This does not mean offshore AI tools are automatically unlawful. It means your business should know where the data goes and on what basis, and be able to explain that if asked.

Consumer accounts versus business accounts

This is where most of the practical risk sits. The same brand of AI tool can behave very differently depending on the type of account.

• Consumer or free accounts are aimed at individuals. Depending on the provider and the settings, conversations may be retained for a period and may be used to improve the provider’s models. The business has no central control, no visibility of what was shared and no way to delete it when someone leaves.
• Business or enterprise accounts typically come with contractual commitments, including that your data is not used to train the provider’s models, along with admin controls, retention settings, single sign-on and audit logs.

Do not assume. Settings and terms change, so check the current terms for each tool you approve, and check them again periodically.

Keeping data in South Africa

For many businesses the simplest way to reduce cross-border questions is to use AI services hosted in-country. Microsoft Azure offers AI services in its South Africa North region, which means prompts and documents can be processed in a South African data centre.

We have built governed AI assistants for professional firms, including a South African law firm, on Azure in South Africa North. Data stays in-country, access is controlled firm-wide through Microsoft 365 accounts, and every interaction is logged. That combination of location, access control and audit trail is what turns “we think it’s fine” into something you can demonstrate.

Write an acceptable-use policy

You do not need a 30-page document. A clear two-page policy that people actually read is far more useful. It should cover:

1. Approved tools. A short list of AI tools staff may use for work, and which account type (business accounts only).
2. What may never be entered. Client personal information, special personal information, passwords, bank details and confidential contracts, unless the tool is approved for that specific purpose.
3. Review of outputs. AI output must be checked by a person before it is sent to a client or used for a decision. Models can be confidently wrong.
4. Who to ask. A named person, usually your Information Officer or IT lead, for questions and new tool requests.
5. What to do if something goes wrong. How to report an accidental disclosure promptly, so the business can assess it. POPIA has requirements for notifying security compromises, and an early report gives you time to act.

Pair the policy with technical controls where you can: business accounts tied to Microsoft 365 sign-in with MFA, blocking unapproved AI sites on company devices if appropriate, and data loss prevention rules for highly sensitive information.

A practical do and don’t list for staff

Do:

• Use only the AI tools on the approved list, signed in with your work account.
• Remove names, ID numbers, account numbers and contact details before asking for help with a document, unless the tool is approved for client data.
• Use AI for drafting, summarising public information, rewording your own text and brainstorming.
• Check every fact, figure and reference in the output before you rely on it.
• Ask before trying a new tool, even a free one.
• Report it straight away if you think you pasted something you shouldn’t have.

Don’t:

• Paste client files, payroll data, medical details or HR records into a personal or free AI account.
• Enter passwords, API keys, banking details or one-time PINs anywhere in a chatbot.
• Upload whole mailboxes, client databases or contracts to see “what it can find”.
• Send AI-written content to clients without reading it properly.
• Install browser extensions or plug-ins that read your screen or email without IT approval.

Treat it as part of your security, not a separate project

AI acceptable use belongs alongside the security basics you should already have: MFA on every account, Conditional Access, endpoint protection, tested backups and documented access permissions. If you are considering Microsoft 365 Copilot, the same thinking applies to who can see what inside your own tenant; our Copilot readiness checklist explains why.

A sensible next step

If you are not sure which AI tools your team is using, or whether your setup would stand up to scrutiny, our free Automation & Risk Check is a good place to start. In 30 minutes we look at where automation would genuinely help, and at the basics: backups, security and documentation. You get a one-page report with what we found, and no obligation to do anything with us afterwards.