AI is coming for lawyers before plumbers. Not because plumbing is harder than law, but because a leaking pipe refuses to become a paragraph. Much of a lawyer’s day already has: judgments, contracts, affidavits, opinions, submissions. A machine that can read and write has a point of entry into that work before it has learned to turn a wrench.
The change is real. Faced with a dismissal dispute, a lawyer can now ask a system to find decisions with similar facts, compare the employer’s and employee’s strongest positions, and shape a first draft around the authorities it finds. Work that once began with a guessed keyword can begin with the question the lawyer actually means to ask. More of the law can come into view before the first argument has hardened.
But the ability to produce an answer is not the authority to rely on it. That distinction becomes clearest at the place where legal writing appears most reassuring: the citation.
An invented case is easy to describe as a failure. The more unsettling failure is a real case attached to a proposition it does not support, a passage that records counsel’s argument rather than the court’s decision, or a relevant judgment the system never retrieved. The link opens. The citation is correct. The sentence beside it is still wrong. And although a lawyer can click every case the machine found, no one can click the case it left out.
The research is beginning to make this problem legible. A 2026 benchmark of Australian criminal-law research found retrieval quality more consequential to answer correctness than changing the model that wrote the answer. An empirical study of specialist US legal research tools found that source-backed systems could still make unsupported claims. Neither study tested Zalari or Southern African law. Both point to the same design question: how much of an answer’s evidential path can the person who must rely on it actually see?
For a lawyer in Zimbabwe or South Africa, that path runs through jurisdiction, court, time, and context. A South African decision may offer an illuminating analogy in a Zimbabwean matter without governing it. A sentence that sounds decisive may be narrowed a page later. A judgment can be relevant to the facts and yet weak authority for the point at issue. No fluent paragraph can relieve counsel of making those distinctions.
Nor should it. AI can read widely, bring an unexpected case to attention, assemble a chronology, compare lines of reasoning, and write a draft worth arguing with. These are substantial gains. If they allow a lawyer to encounter the best opposing authority earlier, they may improve the work rather than merely accelerate it. The mistake is to treat the speed of preparation as a transfer of responsibility.
A legal position is a commitment: these are the material facts; this is the authority that governs them; this is what remains uncertain; this is the risk the client is being asked to bear. Counsel can be questioned about each part of that commitment and must answer for what is advanced or withheld. A model can propose a reason. It cannot occupy that place in the institution. It has no standing.
This is why a final “human review” button is too small an ambition. A reviewer at the end of a poor research process cannot inspect an authority the process never showed. The lawyer needs the full judgment, the deciding forum, the passage said to support the proposition, and room to ask what the other side would say. Judgment begins with what the system makes possible to notice, not with the final click of approval.
We built Zalari with that conviction. Our Law Reports archive supplies the Southern African sources around which we are developing research and drafting. Zalari is now in early preview, not at the end of that work. Its drafts are experiments to question and check against the judgments, not production-ready legal documents. We will share the improvements as they arrive, and fuller capabilities will come later. The purpose is not to make the judgment for you. It is to give your judgment more of the law to work with.
Take a question with a real opposing argument. Ask it in Zalari. Follow the sources into the judgments; read the passage and what surrounds it; tell us what the system missed. That is a more demanding invitation than “trust the answer,” and a more useful one.
The machine can draft the sentence. Only a person can stand behind it.
