Advanced Application of AI
Presenters : Nestene Botha CA(SA)
Overview
Generative AI in most practices still means one chat window, one question and one answer copied out. This session moves beyond that. It shows how a structured AI workflow can hold a firm's rules, work from the actual source documents and run the same steps in the same order on every engagement, so that it becomes more reliable each time it is used.
The session opens with a live demonstration: drafting a set of annual financial statements, then repeating the engagement the following year, using a simple input and output folder structure and a single plain-English rules file. It then takes the workflow apart piece by piece. It covers how to write a concise rules file, how to put together the material a task needs, how to scope work into small steps that can each be checked, and how to build verification into the process so that review does not depend on someone happening to notice an error.
The final section deals with professional judgement. It identifies where generative AI is the wrong tool and sets out the South African chain of accountability under SAICA, IRBA, SARS, IESBA and POPIA. The central point is that the tool can speed up the work, but the practitioner remains responsible for the output.
No coding knowledge is required.
Topics covered
- From chat window to workflow: why a tool with access to practice files behaves differently from a chat interface, and how a workflow differs from an autonomous agent
- The rules file: what belongs in it (boundaries, house style, non-obvious practice knowledge, citation rules), why shorter files work better, and how to maintain it like a standard working paper
- The material: working from the actual documents, standards and precedents rather than the model's training data, and the POPIA and vendor data-use checks to complete before any client information is used
- The ask: scoping work so it cannot go wide, asking the blocking questions before drafting, and breaking large tasks into small, reviewable steps
- The check: why models produce confident but incorrect output, building self-verification tests, using independent second-pass review, and the difference between requesting a behaviour and enforcing it
- Judgement: the three common failures (invented, imported and incomplete output), a four-step verification routine, and the professional duties that stay with the practitioner
Learning outcomes
Practitioners will be able to:
- Describe how a practice builds an AI workflow system that improves each time it is used
- Explain why a tool with access to practice files behaves differently from a chat window
- Write a concise rules file that captures how a practice operates
- Assemble the source material a task requires instead of relying on the model's existing knowledge
- Scope professional work into discrete steps that can each be checked
- Build verification into the workflow so that review is structured rather than incidental
- Identify work within a practice where generative AI is the wrong tool
Who should attend
Tax practitioners, accountants, auditors and practice managers who already use generative AI tools and want to move to structured, repeatable and reviewable workflows within their professional and regulatory obligations.
Related Webinars
Integrating Systems for Seamless Workflow and Operations: From Manual to Digital