Every time you paste a client's information into ChatGPT, you're making a POPIA decision — usually without realising it. There's a simpler option most SA businesses haven't considered yet.
South African businesses have embraced ChatGPT faster than almost anyone. It writes emails, summarises meetings, drafts proposals. It also quietly does something else: every query sent to it leaves the country, gets processed on servers in the United States, and is subject to whatever the vendor's retention policy happens to be that week.
For a lot of day-to-day work, that's fine. For anything involving client information, staff records, financials, or anything covered by an NDA, it's a POPIA question — and most businesses haven't properly answered it yet.
What POPIA actually says about this
Section 72 of POPIA governs cross-border transfers of personal information. You're generally allowed to send personal data outside South Africa only if the recipient is subject to similar protection, the data subject has consented, or a short list of other exceptions apply.
OpenAI, Anthropic, and Google all publish privacy terms that go some way toward compliance, but none of them give you the simple answer that would make a compliance officer relax. Every time an employee pastes a client's name, phone number, or document into a chat window, you are — strictly — performing a cross-border transfer of personal information.
In practice the regulator isn't knocking on doors over it. In principle it's still exposure you don't need to carry.
The cleaner option
A local AI model — running on a laptop, an office server, or a cloud VM inside South Africa — sidesteps almost all of this. The data never leaves the machine. There is no third-party vendor processing it. POPIA still applies internally, but you've removed the cross-border layer entirely.
The 2022 argument against this was that local models weren't good enough. That argument is no longer true. A Llama 3.3 or Qwen 2.5 running on a modest server handles email drafting, meeting summarisation, document Q&A, translation, and most of the actual AI work a business uses day to day. The things it can't do as well as GPT-4 or Claude are increasingly niche.
A practical setup for a small business
The minimum viable version looks like this:
- One machine with 32 GB RAM (a Mac Mini or a small server).
- Ollama running a Llama or Qwen model in the 7–14B range.
- An Open WebUI interface exposed to the internal network only.
- Staff accounts with individual logins and audit logs.
- A written policy about what may and may not be entered, same as you'd have for any tool.
Total setup cost: a one-time hardware purchase under R30 000 for a small team. Running cost thereafter: electricity. Capacity: tens of queries per minute on a fairly small machine.
Where the cloud still wins
Be honest about the trade-offs. Cloud AI is still better for:
- Cutting-edge reasoning on the hardest prompts.
- Multi-modal tasks that include high-quality image generation or voice synthesis.
- Situations where absolute state-of-the-art quality matters more than privacy.
- Extremely bursty demand that would idle your local hardware 90% of the time.
For most businesses, those cases are the minority — and when they do come up, there's nothing stopping you from using cloud AI for specifically those tasks, with the data properly anonymised first.
The regulatory wind is shifting
The Information Regulator has been increasingly vocal about AI and cross-border transfers. The EU's AI Act came into full effect in 2025 and SA's regulatory thinking tends to track European direction. A local-first AI posture today is likely to become the default expected posture in two or three years.
Getting ahead of that curve — and being able to tell clients, partners, and auditors that no personal information ever leaves your premises when AI is involved — is a real differentiator.
The easiest compliance question is the one you don't have to answer — because the data never left the building.
What to do this month
Run an internal audit: which tools does your team paste client or staff information into? Count the number of distinct systems involved, and whether each has an appropriate data processing agreement. For any gap you find, a local AI deployment usually closes it faster and cheaper than a new vendor negotiation.
If you've read our earlier post on POPIA basics, this is the 2026 extension of it. Privacy used to mean choosing your SaaS vendors carefully. Increasingly it means choosing whether you need a vendor at all.
