A citation checker for South African lawyers: one AI idea, followed all the way through
South African courts have started referring lawyers to the LPC over fake AI citations. I ran OppFinder's full analysis on a tool to catch them, then checked what it got wrong.
Of the 18 ideas from my six-industry scan of South Africa, this one had the best-evidenced reason to exist.
In June 2025 a Johannesburg High Court judge found two non-existent cases cited in heads of argument in Northbound Processing v South African Diamond and Precious Metals Regulator ([2025] ZAGPJHC 538), and referred the legal team to the Legal Practice Council. Earlier that year, in Mavundla v MEC ([2025] ZAKZPHC 2), most of the authorities cited didn't exist. Lawyers are drafting with AI that invents cases, and judges have noticed.
The idea is to upload a draft filing and have every case citation and statutory reference checked against real law reports before it goes to court.
I've scanned the legal industry a few times in the last year. Mistakes there are expensive, especially because people are personally liable for them, and a lot of the work still happens in Word, from templates, late at night.
I ran OppFinder's full deep dive on it. This is what came back, and the parts that didn't survive a closer look.
What the analysis said
Who pays
It named two buyers. The first is advocates working alone, who have no junior to check citations and carry the professional risk personally. The second is litigation partners at small and mid-sized firms, who sign off on work drafted by others.
Pricing it proposed
| Plan | Price | For |
|---|---|---|
| Solo | R349/month | One advocate, 25 documents a month |
| Practice | R899/month | Up to 5 users, 150 documents |
| Firm | R2,499/month | Unlimited users, API access |
What to build first
A drag-and-drop upload, extraction of South African citation formats ([2024] ZACC 12, 2023 (1) SA 450 (SCA)), matching against public case databases, a report flagging anything that doesn't match, and a downloadable verification certificate. It estimated 3 to 4 weeks for a solo developer and R4,500 to R12,000 to get started.
The risks
It called one assumption fatal, and it's the most useful sentence in the whole analysis:
"Litigation partners are willing to pay a recurring SaaS subscription rather than relying on junior staff to perform manual lookups on free SAFLII searches (FATAL ASSUMPTION…)"
I agree. SAFLII is free, so the product has to be worth more than a candidate attorney's hour, and what it would really be selling is the record that you checked.
The other risks it listed were reasonable too. Firms won't upload confidential court papers to a stranger's web app, so it suggested processing documents without storing them, hosted in South Africa. A licensed legislation API would eat the margin: Laws.Africa's paid tier starts at R4,200 per knowledge base per month, and it suggested starting on the free sandbox and public case data instead, which is the right call. And an incumbent practice-management suite could simply add the feature.
What it got wrong
I checked the analysis the same way I checked the scan, and four things didn't hold.
-
The market size was built on the wrong headcount. It started from "roughly 15,000 South African legal practitioners". The Law Society's own figures put practising attorneys at about 35,000, before you count advocates. The error happens to make the market look smaller, but it still carries into every TAM and SAM figure that follows.
-
It recommended AI models that are years out of date. For the parsing and verification steps it suggested "GPT-4o-mini" and "Claude 3.5 Sonnet". That's the same problem that made me take down 61 of my own blog posts: the model only knows its training data, and nothing told it the date. I'm fixing this in OppFinder now. Until then, treat any model name in a build plan as a placeholder.
-
An unverified "fact" turned into a product feature. The scan said the LPC issued a directive in July 2026 requiring direct personal supervision of candidate attorneys. I couldn't find that notice on the LPC's site, and secondary sources mention a change effective August 2026. The analysis had already built it into the plan: the R899 tier includes "candidate attorney oversight logging", justified by that directive. By the time I read the plan, an unchecked claim had become a pricing decision.
-
It quoted competitor prices I couldn't find. It benchmarked against a practice-management product's "published" R299/R599/R999 per-user pricing, and I couldn't find those prices published anywhere. Its revenue estimates lean on that anchor.
None of these sink the idea. All four would have embarrassed me in front of the first lawyer I pitched.
What I'd do with it
I won't be building it, because I'm a solo founder with a day job. Selling to lawyers means long trust-building and confidentiality reviews, and I can't carry that sales cycle. But if you're a developer with legal contacts, this is the best-evidenced idea from these scans.
If you want to test it, don't start with the product. Take five real, public heads of argument, check every citation by hand on SAFLII, and time it. If it takes ten minutes, there's no product. If it takes two hours and you find a mistake, there might be.
Why do this in public
The deep dive did most of what I want from it. It found a real trigger, named the fatal assumption itself and suggested a cheap way to start. It also carried four errors into a plan that looked finished.
AI market research in 2026 gets you a fast first draft, and someone still has to check the handful of claims the plan depends on. OppFinder shows its sources so you can do that checking.
Run the same analysis on your own idea at oppfinder.app.
Related Resources: