Almost a third of the legal community says it would never let an AI agent act without human approval, even on low-risk work, while nearly half would allow it to act only under strict oversight. Given the hallucination risks inherent with any AI system, that scepticism is rational. However, it is also the single biggest constraint on whether agentic AI delivers anything beyond a faster first draft. The same audience ranked security and governance as the deciding factor when selecting legal technology and put AI functionality last, on zero. For that audience, what an agent can do matters less than whether anyone can show how it did it.

This session looks at autonomy from the build side: what it takes to get an agent trusted with real legal work in a South African firm, and what breaks when you try. Drawing on the build of Juta Integra and on delivery experience with firms across South Africa and Namibia, it examines why layered review so often reinstates the bottleneck the technology was meant to remove, how risk-based oversight can be designed rather than improvised, and why data quality and knowledge structure (rather than model capability) usually decide whether an agent is safe to let run. An agent is only ever as defensible as the content it is allowed to work from and how quickly someone can open the source behind an answer.

The session moves past pilot-stage enthusiasm to the operational questions: where judgement must stay human, how to evidence that an agent behaved defensibly, and what a firm needs in place before autonomy is worth attempting at all. The worked examples are drawn from Juta Integra, where agentic steps run over Juta’s licensed legislation, case law and commentary, and every proposition returned resolves back to the source it came from.

Learning 1: A practical test for separating work that can be delegated to an agent from work that cannot.

Learning 2: The data and knowledge foundations that determine whether autonomy is safe, and how to adequately assess yours.

Learning 3: What “defensible” means for agentic work in practice: traceability, provenance and audit trails that hold up under client and regulator scrutiny.