Contract review and drafting is the single biggest legal-AI use case — and for good reason: a lot of a commercial contract is standard, and extraction, redlining, and first drafts are exactly what the tool is good at. But South African contract law gets two things backwards from what a US- or English-trained model assumes. Penalty clauses are enforceable here (the Conventional Penalties Act), not void. And a restraint of trade is presumed enforceable (Magna Alloys), with the onus on the person trying to escape it — the opposite default. Add the CPA, the NCA, ECTA e-signatures, and a three-year prescription clock, and you have a practice where the drafting help is real and the confident wrong answer is a live risk.
South African contract law is common law — Roman-Dutch in origin, developed by the courts — and its starting point is freedom of contract: parties may agree what they like, and the courts will hold them to it. As a general rule no writing is required for a valid contract; a handshake deal binds. But that freedom sits under an overlay of statute and distinctive common-law rules that a generic tool doesn't carry: the Consumer Protection Act 68 of 2008 (CPA) where a consumer is involved, the National Credit Act 34 of 2005 (NCA) where credit is granted, the Electronic Communications and Transactions Act 25 of 2002 (ECTA) for e-signatures, the Conventional Penalties Act 15 of 1962 for penalty clauses, and the Prescription Act 68 of 1969 for how long a claim survives.
The exceptions to "no writing needed" matter: the sale of land must be in writing and signed (Alienation of Land Act 68 of 1981), a suretyship must be in writing, and long leases and a few other categories carry formalities. These are precisely the edges a fluent model rounds off.
Every commercial contract is a mix of standard machinery and a handful of clauses where the value and the risk actually live. AI is strong on the first and must not be trusted with the second.
| Standard machinery | The negotiated risk clauses |
|---|---|
| Definitions, interpretation, notices | Limitation & exclusion of liability — who carries what, and the cap |
| Term, renewal, general boilerplate | Indemnities — the risk-shifting engine of the deal |
| Governing law, jurisdiction, severability | Warranties & representations — what's promised true |
| Force majeure, assignment, entire-agreement | Restraint of trade & IP assignment — what's protected and owned |
A contract negotiation is almost never about the definitions. It's about the cap on liability, the scope of the indemnity, and the reach of the restraint. Those are judgement, leverage, and risk appetite — a human's job. AI compresses the machinery so the lawyer's time moves to the four clauses that decide the deal. Used the other way round — trusting the model on the risk clauses — it's a liability, not a tool.
Contract work is the top-cited legal-AI use case because it's high-volume and pattern-heavy (see the Legal AI leaf). The tool genuinely helps with review against a playbook (does this draft match our positions?), clause extraction across a portfolio (find every change-of-control, every auto-renew, every uncapped indemnity), first drafts from a precedent, and contract-lifecycle management — surfacing renewal dates, obligations, and expiries buried in a stack of signed agreements. This is real, measurable time saved on work that was always tedious.
The pattern that works: give the model your house playbook and precedents as the ground truth, and use it to draft and check against them — not to invent the law from its training data. Grounded in your own approved templates, the output is a fast first pass a lawyer verifies. Ungrounded, it produces confident, plausible clauses that may be wrong for the jurisdiction — which is exactly the next section.
These aren't gaps in the model's knowledge — they're places where the default it learned is the opposite of the SA position. That's the dangerous kind of error: fluent, confident, and inverted.
| Topic | What a US/English-trained model assumes | The SA position |
|---|---|---|
| Penalty clauses | Penalties are unenforceable — use "liquidated damages" | Enforceable. The Conventional Penalties Act 15 of 1962 makes penalty stipulations enforceable; a court may only reduce one that's disproportionate to the prejudice. |
| Restraint of trade | Prima facie unenforceable unless the employer proves it's reasonable | Prima facie enforceable. Magna Alloys v Ellis (1984): the onus is on the employee to show the restraint is unreasonable or against public policy. |
| Consumer terms | Freedom of contract; the terms are the terms | The CPA 68 of 2008 requires plain language and can strike unfair, unreasonable or unjust terms in consumer transactions. |
| E-signatures | Wet ink for anything important | ECTA 25 of 2002 gives electronic signatures legal validity — except land sales, leases over 20 years, bills of exchange, and suretyships. |
| Time to sue | A long, variable limitation period | The Prescription Act 68 of 1969: an ordinary contractual debt prescribes in three years. Miss it and the claim is dead. |
Two statutes flip a commercial contract into a regulated one based on who you're dealing with. The Consumer Protection Act applies when the other side is a consumer (and, for smaller businesses, when they transact as one): it mandates plain language, regulates unfair terms, and gives cooling-off and disclosure rights — ECTA adds a seven-day cooling-off for distance selling on top. The National Credit Act applies whenever credit is granted: it demands affordability assessment, disclosure, and fairness, and it reaches ordinary commercial arrangements (deferred payment, instalments) that don't look like "credit" at first glance. A generic template that ignores which regime applies isn't just suboptimal — it can be unlawful and unenforceable in the parts that matter.
A model will confidently rewrite an enforceable SA penalty clause into "liquidated damages" to avoid a rule that doesn't apply here — and may weaken the client's position in doing it. The Conventional Penalties Act means the penalty was fine.
Drafting a restraint on the assumption it's presumptively unenforceable produces weak, over-hedged wording. In SA the restraint is presumed valid; the drafting posture is different, and a foreign default gets it backwards.
Ungrounded, a model invents plausible clauses and cites Acts or cases that don't say what it claims. In a contract, a confident wrong clause is a latent dispute. Ground it in your precedents, and verify.
Three years is short. A CLM or advice tool that doesn't surface the prescription risk on an unpaid or breached contract lets a live claim quietly die. The date is the point.
Counterparties, signatories, and often employees appear in the contract set. Running an AI tool over the portfolio is a POPIA processing — and a section 72 transfer question if it processes offshore. See the Data privacy & POPIA leaf.
The clearest ROI: a high volume of recurring contracts — supply, services, NDAs, SLAs — reviewed against a house playbook. Extraction and playbook-review compress the queue; the team's time moves to the negotiated clauses and the CPA/NCA calls the template can't make.
Portfolio-level work: due-diligence extraction (shared with the M&A leaf), lease and contract abstraction, and lifecycle migration into a CLM system. The tool reads and structures; the lawyer decides what the findings mean.
Where the CPA's plain-language requirement bites hardest — consumer-facing terms that must be understandable. AI can help simplify and standardise, but the compliance judgement (is this term fair, is it enforceable, does the NCA apply) is a lawyer's, not a generator's.
Penalties enforceable, restraints presumed valid — both are the reverse of the Anglo-American position a model most likely learned. The lesson generalises: on commercial contracts, the danger isn't the SA rule the tool doesn't know, it's the foreign rule it's sure of. Ground the tool in local precedent and treat every confident statement of law as a claim to verify.
CPA, NCA, and ECTA don't just add clauses — they change which terms are even permissible, when a signature is valid, and how long a claim lives. An SA commercial contract is a common-law document policed by statute, and the policing is invisible to a jurisdiction-blind tool.
A contract set is dense with personal information — signatories, counterparties, employees. Any AI that reads it is a POPIA processing event, and a section 72 question if it processes across the border. The residency answer is the tree's usual one: keep it in country. See Data privacy & POPIA.
Commercial contracting is where legal AI is used most — and where the confident foreign default does the most quiet harm.
The Acts that overlay SA contract freedom — and note the leading restraint-of-trade authority is case law (Magna Alloys v Ellis 1984 (4) SA 874 (A)), so read it in a current practitioner text.