This two-day course for legal professionals explores how AI works, where it fails, and how to evaluate and implement it responsibly. The course is taught by a practitioner who has built and deployed grounded AI systems in a regulated industry.
When and where?
22 to 23 February 2027, 9:00 to 11:00 SAST
This course will be held remotely, most likely on Zoom - exact details will be sent to registered participants a few days before the course.
Course outline
Throughout the course, a popular LLM and a grounded, orchestrated AI implementation are compared side by side. The popular LLM answers from patterns in its training data. The grounded implementation retrieves from verified sources and answers only from what it retrieves, within defined guardrails. The comparison is about architecture, not products: the same principles apply to any AI tool participants encounter or evaluate in practice.
Day 1: How AI Works and Where It Fails
Why This Matters for Lawyers
• AI is already in legal practice, used formally or informally by staff and opposing counsel.
• South African courts have already dealt with AI-fabricated case citations; this is a present risk, not a future one.
• Responsible use starts with understanding the technology, not avoiding it.
How Large Language Models Actually Work
• LLMs predict the most likely next word based on patterns in training data. They do not look up facts.
• Fluency is not accuracy: a confident answer carries no guarantee of a real source.
• The witness analogy: a popular LLM is a witness testifying from memory.
Where Your Data Goes
• Consumer, enterprise and self-hosted tools treat your data very differently.
• Some tools retain inputs or use them for training; this directly engages confidentiality and privilege.
• Hosting location matters for cross-border data transfer under POPIA.
What AI Can and Should Not Do
• Strong: drafting, summarising, first-pass document review, structuring information.
• Weak: citing authorities, reaching legal conclusions, anything requiring verified fact.
• AI output is a starting point for professional judgment, never a substitute for it.
Live Demonstration: Memory vs the Bundle
• The same legal question put to a popular LLM and to a grounded, orchestrated AI implementation, side by side.
• Retrieval-augmented generation (RAG) is the witness bound to the bundle: the model answers only from verified, retrieved sources.
• A well-designed system declines to answer when its sources are silent, rather than guessing.
• Grounding reduces hallucination significantly but does not eliminate it; retrieval quality still matters.
The Danger Map
• Key risks: hallucination, data leakage, bias in automated decisions, IP uncertainty, over-reliance.
• Each risk has an architectural mitigation, illustrated through a grounded, orchestrated implementation.
• The question shifts from whether to use AI to how to evaluate and govern it.
Day 2: Evaluating and Implementing AI Responsibly
The Regulatory Frame
• POPIA applies fully to AI processing, including section 71 on automated decision-making.
• Professional conduct duties of competence and confidentiality extend to the tools practitioners use.
• The EU AI Act sets the global benchmark for risk-based AI regulation and signals where local regulation may go.
Evaluating AI Tools
• A practical due-diligence checklist: data use and training, retention, hosting location, audit logs, accuracy and contract terms.
• Vendor claims should be tested contractually, not accepted at face value.
• The right tool depends on the risk profile of the task, not the popularity of the product.
Worked Evaluation: A Grounded Implementation
• A grounded, orchestrated AI implementation assessed against the full due-diligence checklist.
• Where a grounded system performs well, and where its honest limits lie (retrieval errors, source quality).
• Knowing a tool’s limits is as important as knowing its strengths.
Implementing with Guardrails
• Core safeguards: grounding in verified sources, scope limits, access controls, human-in-the-loop review and logging.
• Real build decisions from deploying a grounded AI implementation in a regulated industry.
• Guardrails are designed in from the start, not added after deployment.
Where This Is Heading
• Agentic AI: tools that take actions, not just produce answers.
• AI in courts and regulatory processes.
• Emerging liability questions: who is responsible when a grounded system still gets it wrong?
Q&A and Close
• Open discussion and key takeaways.
• Take-home resources: an AI tool evaluation checklist and a firm AI use policy template.
Who will benefit from this course?
In-house lawyers and law firm practitioners seeking a practical understanding of AI.
Presenter
Nicholas Pillay
AI engineer, business consultant and strategist. Founder and CEO of Emma, a bespoke, fully independent AI orchestration platform deployed at multiple enterprise-level companies. Founder of Viginti Digital, where he leads AI and automation, and co-owner and CTO of Pipeline Concierge, an AI-powered platform for US real estate agents. Combines hands-on engineering with business strategy to build secure, privacy-conscious AI systems that solve real commercial problems. Holds a Postgraduate Diploma in Business Management from the University of Essex (merit).
How much?
R2,870 per person
Certificate
A digital certificate of attendance from UCT will be issued to those who attend the full course.
Please note that the digital certificate can only be viewed on a secure portal. It cannot be downloaded or printed. You will have the option of ordering a hard copy of the certificate at your own cost, including the cost of the courier fee. More information is available here.
No certificate will be issued without the full course fee having been received. Please allow up to three weeks after the end of the course for certificates to be processed.
How to sign up
Complete and submit the registration form. You will then be given the payment information. Please note that registrations will not be accepted until payment has been made.
One or two days before the course, we will send you the Zoom link. You will need to register and use a password to enter the virtual classroom.
Registrations close three days before the course starts.
Download the brochure.
