AI for SA business
How to anonymise data before using AI (POPIA-safe, step by step)
The cleanest way to use AI safely is to make sure the data is not personal information any more. POPIA governs personal information, so if you strip the identifying detail out of a dataset before you ask an AI tool anything, you have stepped outside the part of the law that creates most of the risk. The work is not complicated, but it is more than deleting the name. Here is the POPIA-safe way to do it.
Why this works (the POPIA shortcut)
POPIA (the Protection of Personal Information Act, Act 4 of 2013) only applies to personal information. Section 6 says the Act does not apply to information that has been de-identified to the extent that it cannot be re-identified. So if you genuinely de-identify a dataset, it falls outside POPIA’s scope and you can put it to work, including in an AI tool, without the same obligations. That is the shortcut: do not fight to use personal data safely when you can remove the personal element entirely.
Anonymise vs pseudonymise: the difference that matters
These two get used interchangeably, and the gap between them is where businesses get caught.
Truly anonymised data is irreversible. The identifying detail is gone and cannot be reconstructed, so under s6 it is outside POPIA. Pseudonymised data is reversible: you swap a customer’s name for “Customer 12” but keep a key somewhere that maps it back to the real person. That key means the data still identifies someone, so it remains personal information and POPIA still applies in full. POPIA’s own definition of de-identification (s1) is strict: you must delete the information that identifies the data subject, or that can be used or manipulated to identify them, or that links to other identifying information. Swapping in a label while holding the key does not meet that test.
The step-by-step
Work through it in order; each step removes a layer of identifiability.
List the identifiers. Go through the dataset and find every field that could identify a person, on its own or in combination: name, ID number, account or policy numbers, email, phone, physical address. Combinations count, not just single fields.
Remove direct identifiers. Delete or replace the obvious ones, name, ID number, contact details, account numbers, before anything goes near the AI tool. These name a person outright, so they go first.
Generalise quasi-identifiers. Blunt the indirect fields that can identify someone in combination: exact date of birth becomes an age band, full address becomes a suburb or region, a rare job title becomes a broad category. This is the step that separates real de-identification from a stripped name.
Check re-identification risk. Look at what is left and ask whether someone could still work out who it is, especially in a small town or a niche dataset where one record is unusual. If yes, generalise more.
Keep any key separate. If you have to retain a mapping back to the real person, store it separately and offline, and never paste it into the AI tool. The moment the key meets the data, it is personal information again.
Record what you did. Note what you removed and why. A short record lets you show your de-identification was reasonable if anyone asks.
Direct vs quasi-identifiers
The mistake is treating only the first column. POPIA’s de-identification standard means you have to handle both.
| Type | Examples | How to treat it |
|---|---|---|
| Direct identifier | Name, ID number, passport number, email, phone, account or policy number | Delete or replace before the data leaves your system |
| Quasi-identifier | Date of birth, suburb or postal code, gender, employer, job title, race | Generalise: age band, region, broad category, so no combination is unique |
| Linking field | A reference number that ties back to another identifying record or key | Remove it, or hold the key separately and offline, never with the data |
Checking re-identification risk
De-identification is judged on whether the data can realistically be re-identified, not on whether you deleted a checklist of fields. The risk rises sharply in small populations and unique combinations. A record reading “42-year-old female attorney in a town of 3,000” may name exactly one person even with no name attached. The same combination in Johannesburg points to thousands and is far safer. So the test is contextual: in a large, common dataset, modest generalisation may be enough; in a small or unusual one, you generalise harder, or you do not use that data in a public tool at all. “Reasonable” de-identification, in POPIA’s terms, means it cannot realistically be reversed.
When anonymising isn’t enough
Sometimes you genuinely need the real personal data, because the task is about a specific, named person: drafting a reply to a named customer, working a real claim, acting on one client’s account. You cannot de-identify that and still do the job. Anonymisation is not the answer here. The route is a private or Enterprise AI deployment with a data processing agreement (DPA) in place and model training switched off, a proper responsible-party and operator arrangement under POPIA, so the personal information is handled under contract rather than typed into a public tool. (Bear in mind that while data is still personal information, sending it to a US-hosted AI tool is also a transborder flow under s72, which is another reason removing the personal element matters when you can.)
Where Zaiq fits
Zaiq is an AI engineering studio in South Africa, and we build de-identification and safe-AI workflows in from the start: automatic stripping of personal detail before anything reaches a model, private deployments for the cases where you genuinely need the real data, and a clean processing record so you can show what happened and why. We do this in practice: a one-click publish pipeline we built for a client only ever sends the customer-safe version out, while the sensitive underlying data stays private and never leaves their own system. De-identification is a design choice, not an afterthought. If you want the speed of AI without putting customer data at risk, bring us the problem at zaiq.ai/work and we will tell you straight what safe looks like for your case.
This is general guidance, not legal advice. For a high-risk or consequential processing decision, confirm with a POPIA practitioner.
Related guides
Anonymise data before using an AI tool
Strip identifying detail so the data falls outside POPIA before you put it into ChatGPT or any AI tool.
List the identifiers
Find every field that could identify a person: name, ID number, account or policy numbers, email, phone, address, and combinations of these that together point to one individual.
Remove direct identifiers
Delete or replace the obvious ones, name, ID number, contact details and account numbers, before anything goes near the AI tool. These are the fields that name a person outright.
Generalise quasi-identifiers
Blunt the indirect ones: exact date of birth to an age band, full address to a suburb or region, rare job titles to broad categories. In combination these can identify someone, so widen them.
Check re-identification risk
Ask whether someone could still work out who this is from what is left, especially in a small town or a niche dataset where one record stands out. If the answer is yes, generalise further.
Keep any key separate
If you must keep a mapping back to the real person, store it separately and offline. Never paste that key into the AI tool, because it is what turns de-identified data back into personal information.
Record what you did
Note what you removed and why, so you can show your de-identification was reasonable. A short record is what evidences that the data was genuinely taken outside POPIA's scope.
Questions people ask
What's the difference between anonymising and pseudonymising data?
Anonymising means de-identifying data so it cannot be re-identified at all, which under POPIA s6 takes it outside the Act. Pseudonymising is reversible: you swap a name for "Customer 12" but keep a key that maps it back. That key means the data still identifies a person, so it remains personal information and POPIA still applies.
Does de-identified data fall under POPIA?
Genuinely de-identified data does not. POPIA s6 says the Act does not apply to personal information that has been de-identified to the extent that it cannot be re-identified. The catch is "to the extent that it cannot be re-identified": if someone could realistically reverse it, it is not truly de-identified and the Act still applies.
Is removing the name enough to anonymise data?
No. The name is only one direct identifier, and POPIA's definition of de-identification (s1) covers anything that can be used or manipulated to identify a person. Quasi-identifiers like date of birth, suburb and a rare job title can pinpoint someone in combination, so you have to address those too, not just delete the name.
Can anonymised data be used in ChatGPT freely?
If the data is truly de-identified so it cannot be re-identified, it is outside POPIA (s6) and far safer to use. The discipline is making sure it really is anonymised first: direct identifiers gone, quasi-identifiers generalised, re-identification risk checked. If any real personal detail remains, treat it as personal information and keep it out of the public tool.
What are quasi-identifiers?
Quasi-identifiers are fields that do not name a person on their own but can identify them in combination, such as date of birth plus suburb plus gender, or a rare job title in a small town. POPIA's de-identification standard means you have to blunt these, usually by generalising them, not just remove the obvious direct identifiers.
How do I anonymise financial data for AI?
Strip the direct identifiers first: name, ID number, account and policy numbers, contact details. Then generalise the quasi-identifiers, exact amounts to bands, full dates to months, branch to region, so a record cannot be tied back to one client. If you must keep a key to the real person, store it separately and offline, and never paste it into the AI tool.