Most South African SMEs we talk to are asking the wrong AI question. They ask "which model should we use?" before they've asked "where would AI actually move the needle in our business?".
In six years of integration work, we've seen the same pattern repeat: companies pick a flashy use-case (a chatbot, an image generator) when the real ROI sits two doors down in operations or sales-ops. Here's how we map it.
Step 1 — Audit, don't aspire
Start with what the business actually does today. Sit with the ops manager. Count the hours spent on repeatable cognitive work — proposals, reports, support replies, scheduling. Wherever a smart person is doing something predictable, AI can probably help.
Step 2 — Pick the smallest possible first win
The goal of project one is not impressive technology — it's a paid-back business case in 60 days. Reply drafting for support, proposal first drafts, internal Q&A on policy documents — these are dull but they pay back fast.
Step 3 — Build the guardrails before you build the agent
Approval gates, audit logs, escalation paths. We will not deploy an AI action that touches money or signs anything without human review. POPIA expects it, your auditors will expect it, and frankly customers expect it too.
Step 4 — Train on YOUR data
The public LLM doesn't know your products, your tone, your customers. RAG (retrieval-augmented generation) layers your private docs over a public model so the answers sound like you and reference your reality.
Step 5 — Instrument everything
Every AI action gets logged. Token costs, response times, satisfaction scores, override rates. You can't manage what you don't measure, and you cannot defend a budget for an AI feature you can't put a number on.
Step 6 — Expand only after the audit
After project one ships and earns its budget back, repeat the audit. The second win is always easier because you now have your own playbook.