AI business ideas in South Africa: 6 industries, 18 ideas, and a fact-check
I ran OppFinder across six South African industries, then checked 14 of the claims it made against real sources. Four were right, six were right with a wrong detail, three I couldn't confirm and one was wrong.
I build OppFinder from South Africa, so I wanted to see what it says about my own market. On 19 October I ran it on six industries, all scoped to South Africa:
- legal tech
- logistics
- healthcare
- agriculture
- retail
- property management
Each scan searched the web, kept the 13 to 17 sources it found, and returned three ideas. That's 18 ideas for under R10 in API costs.
Then I checked the results, which is what I'd tell anyone to do with AI research before spending money on it.
The fact-check
I took 14 specific claims from the scans (dates, figures, named cases, prices) and checked each one against sources I could find independently.
| Result | Count | Example |
|---|---|---|
| Confirmed | 4 | The Consumer Goods and Services Ombud logged 15,355 complaints in its 2025/26 year, up 25.8% |
| Right story, wrong detail | 6 | The National Opt-Out Registry did launch in October 2026, but the Trade Minister and the National Consumer Commission launched it, not the Information Regulator |
| Couldn't confirm | 3 | A citrus-industry "R1.6 billion cold storage" figure. The closest real number is a BFAP study's R1.56bn direct cost component of a R5.27bn total |
| Wrong | 1 | The Council for Medical Schemes' 2027 increase guideline is 3.8%. The scan said 3.3%, which was the 2026 figure |
That's about what I expected. Grounded AI research is good at finding the story and worse at the details. It slips on the regulator, the year and the percentage, and those are the details you'd put in a sales email to a customer who knows better.
So treat what follows as leads, each with a reason to look closer.
The ideas worth a second look
I've kept one idea per industry, the one whose "why now" held up best in the fact-check.
Legal tech: a citation checker for court filings
The trigger is recent. In Northbound Processing (Pty) Ltd v South African Diamond and Precious Metals Regulator ([2025] ZAGPJHC 538, 30 June 2025), heads of argument cited two cases that don't exist, and the court referred the legal team to the Legal Practice Council. Earlier that year, in Mavundla v MEC ([2025] ZAKZPHC 2), most of the cases cited were fictitious and the firm paid costs (Michalsons summary).
The idea is to upload a draft and have every case citation and statutory reference checked against real law reports before it goes to court.
The scan also spotted why it might fail. Advocates can check citations on SAFLII for free, so what you'd be selling is peace of mind and an audit trail.
I liked this one enough to run the full deep dive on it, which is its own post.
Logistics: transfer-pricing customs paperwork
On 11 September 2026, SARS published amendments to the customs rules (Government Gazette 55355, effective 14 September). Among other things they introduce a Customs Value Adjustment Calculation (CVAC) spreadsheet that importers must submit when transfer-pricing adjustments change the value of goods.
The idea is to pull the numbers from intercompany invoices and import declarations and produce the CVAC submission.
The scan said importers have 30 days to submit. I couldn't confirm that from the gazette excerpts, so read the notice yourself before building around a deadline. The scan also flagged the weakness: adjustments are periodic, so a customer might only need you once a year.
Retail: handling complaints before they reach the Ombud
The CGSO received 15,355 complaints between March 2025 and February 2026, up 25.8% on the year before, mostly about defective goods and failed cancellations.
The idea is a tool for mid-sized retailers that triages these complaints and drafts responses in line with the Ombud's guidelines before they escalate.
The scan's own objection is one I agree with. Most of these complaints come from broken deliveries and poor stock, and software that writes a better apology doesn't fix a parcel that never arrived.
Agriculture: plant variety registration paperwork
The Plant Improvement Act 11 of 2018 came into force on 1 December 2025, replacing the 1976 Act, with new rules for registering plant varieties and propagation facilities.
The idea is to check seed companies' and nurseries' registration submissions against the new requirements before they're filed.
The scan pointed out that faster filing doesn't help if the department's backlog is the bottleneck. I couldn't confirm the size of that backlog, but the logic holds. It also couldn't estimate revenue and said so ("Unknown - would need published consulting fee schedules"), which I'd much rather see than an invented number.
Property management: governance tools for sectional title schemes
The Community Schemes Ombud Service had 37,613 registered schemes at 31 March 2025, many of them run by volunteer trustees.
The idea is a portal that gives owners the financial disclosure and maintenance tracking they're entitled to, so disputes don't end up at CSOS.
I couldn't confirm the scan's dispute figure (16,791, up 7.7%); the closest I found was "over 16,000". The scan did name the real obstacle itself: selling to a body corporate means waiting for a committee and an AGM.
Healthcare: skipped
The healthcare scan had the weakest evidence of the six. Its headline figures were the 3.3% benchmark (wrong year) and an "R43 billion" out-of-pocket figure I couldn't find anywhere. The ideas might still be good, but I'd start the research again before building on them.
What I'd do with this
I would probably choose the citation checker. The buyer carries personal risk, and attorneys have already been caught citing non-existent cases in court. It's also cheap to test on public court documents before writing any code.
Before building anything, I'd check the two or three facts the idea depends on against the primary source (the gazette, the judgment or the regulator's own report), then talk to five people who'd have to pay for it.
Why I'm publishing the mistakes
OppFinder lists its sources for every idea so you can do this kind of check. Writing about its output without doing the check myself would have been odd, and hiding the parts it got wrong would have been worse.
If you run your own scan, click the sources. Most of the time they'll hold up. When they don't, it's usually a date, a number or a name, which is what you'll be quoted on later.
You can run a scan on your own industry and region at oppfinder.app. Every idea lists the sources behind it.
Related Resources: