Commercial and contracts, disputes, corporate and M&A, employment, IP, and data privacy — the practice areas where getting it wrong isn't an option. That single fact sets the bar for the whole domain: verification, grounding, and data residency come before capability. The branch opens where the pressure is greatest right now — Legal AI, read as a category rather than a product pitch, with the POPIA and legal-privilege reality that the global studies leave out.
Seven sub-categories under leg/, all live — the tool (Legal AI), the law (Data privacy & POPIA), the deal (Corporate & M&A), the workforce (Employment & labour), the contract (Commercial & contracts), the dispute (Disputes & litigation), and the idea (Intellectual property). The Legal branch is complete.
Where the value actually lands — the tasks (drafting, DD, extraction), the honest read on the evidence, the landscape (Harvey, Legora, CoCounsel, Lexis+), why the business model is the real disruption, and the POPIA / privilege / residency reality for SA firms.
LiveThe top legal-AI use case — and the two SA rules a global tool gets backwards: penalty clauses are enforceable (Conventional Penalties Act), and restraints of trade are presumed valid (Magna Alloys). Plus the CPA, NCA, ECTA e-signatures, and a 3-year prescription clock.
LiveAI summarisation lands hardest on the discovery bundle — and the catastrophic failure mode is real: two SA courts (Mavundla, Northbound) found fabricated AI citations and referred the lawyers to the Legal Practice Council. Where the record work is gold and the generated citation is a conduct event.
LiveThe deal, the data room, and what AI can't decide. Companies Act fundamental transactions, where AI due diligence does its heaviest lifting and where it must not, Competition Act merger control and the public-interest test, the Takeover Regulation Panel, and POPIA in the data room.
LiveHigh-volume drafting over a floor you can't contract out of. The BCEA and LRA, fair dismissal (substantive and procedural), the CCMA, and the 2025 employment-equity overhaul — and the one thing a US-trained model gets dangerously wrong: SA has no at-will employment.
LiveThe Act 4 of 2013 read for the AI era: the eight conditions, special personal information, the section 72 cross-border transfer test every AI decision runs through, the operator agreement, and how POPIA compiles into an AI-use policy. The legal substrate under the whole tree's residency work.
LiveWhere AI collides with the law most sharply. SA is the only country to have granted an AI-inventor patent (DABUS) — but because it doesn't examine, not because it decided a machine can invent. The computer-generated-works authorship rule, and the fair-use Bill now before the Constitutional Court that decides whether you can train on protected works.
Five positions that hold across the branch. Legal is the domain with the lowest tolerance for hype, because the cost of a confident error lands on a client.
Legal AI is judged on verifiability and grounding, not fluency. A tool that produces a plausible, unsourced answer is a liability in this domain, not a shortcut.
The capability is becoming undifferentiated across vendors. The return lives in the power-user cohort and the workflows around the tool — not in which model you licensed.
AI is prising apart the billable hour — fixed-fee, subscription, managed services. The technology change is smaller than the commercial one it forces.
Under POPIA, client matter data is special personal information and often privileged. Where the documents sit and who can reach them is decided before any tool is chosen.
Most of the impressive legal-AI data is commissioned by the vendor and scoped to its own customers. Read the methodology; keep the structural findings, asterisk the numbers.
Legal sits on top of the tree's residency and agents work. Strongest links to the Agents and Tech domains (grounding and in-country inference), the People branch (the roles legaltech creates), and the briefings that translate POPIA and pricing for a decision-maker.