Getting comfortable with AI at work, without the hype
A digestible orientation for any South African business owner who has heard about AI for three years and not yet built a single workflow around it.
If you have read three explainer pieces about AI in the last six months and still do not know where to start, you are in good company. Most of what is written about AI for business is either tech-bro fantasy ("agents will replace your workforce") or a recycled ChatGPT tutorial that does not survive contact with how a real South African business actually works.
This guide is the orientation a senior operator gives a partner before any real workflow gets built. Five minutes here saves you four weekends of reading the wrong things.
The mental model that matters
Forget the word "AI" for the next sixty seconds. The thing you are actually buying is a fast, capable, slightly forgetful junior who works for under R400 a month and never sleeps.
That junior is excellent at three things:
- Reading and summarising — long documents, email threads, contracts, transcripts.
- Drafting and rewriting — replies, proposals, briefs, social posts, the first version of anything.
- Pattern matching — spotting outliers, categorising, comparing, asking the questions a senior would ask.
That junior is terrible at three things:
- Doing maths reliably — they will confidently get the VAT calculation wrong.
- Knowing what they do not know — they will invent a SARS form number that does not exist.
- Remembering anything from yesterday — every conversation starts from zero unless you tell them otherwise.
The whole game is learning which tasks fit the strengths and avoiding the weaknesses.
Where to start (in this order)
1. Pick a paid plan
If you are evaluating AI on the free tiers of ChatGPT, Claude, or Gemini, you are evaluating last year's models. The paid plans (around R350-R400 a month each) get you the current frontier models and, just as importantly, the longer memory and document upload features that make real work possible.
Pick one. ChatGPT Plus is the safe default for most. Claude Pro is the safe default if your work is heavily writing-led. Gemini Advanced is the safe default if you live inside Google Workspace and want AI sitting inside Docs and Sheets. There is no wrong answer. Pick one and use it for thirty days before you compare.
2. Pick one task you do every week
Write down the five tasks you do every single week that you wish you did not have to do. The list usually includes:
- Drafting the same kind of email reply for the fifth time this month.
- Summarising a long document for someone who will not read it.
- Writing the first draft of a proposal that is 80% boilerplate.
- Categorising a stack of incoming things (CVs, invoices, emails, leads).
- Pulling key facts out of a contract or a SARS letter.
Pick the one that lands hardest. That is your first AI workflow.
3. Brief the model like a junior
The single biggest mistake new AI users make is treating the model like a search engine. You do not type "summarise email" — you brief the model the way you would brief a junior.
A good brief has four parts:
- Who you are ("I am a partner at a Sandton-based commercial law practice.")
- What you want done ("Draft a reply to the attached client email.")
- Constraints ("Keep it under 150 words. South African English, not American. Do not promise a timeline.")
- The raw material (paste the email or upload the document.)
4. Repeat until it is boring
The compounding effect of AI at work only shows up when one task becomes ten tasks becomes a hundred. A single one-off use does not move the needle. A workflow you run every Friday for three months does.
After thirty days of daily use on one task, add a second. After sixty days, add a third. By day ninety you will know more about what AI actually does inside your business than any consultant trying to sell you on it.
What POPIA actually requires you to do
This is the part everyone glosses over and it is the part that actually matters in South Africa.
POPIA (the Protection of Personal Information Act, 4 of 2013) governs how you process personal information about identifiable people. Pasting a client's name, contact details, ID number, or sensitive personal information into ChatGPT or Claude is processing, and the same rules apply that would apply to any other vendor.
What this means in practice:
- Anonymise where you can. Replace real names and identifiers with placeholders before pasting.
- Use the paid plans that offer data processing terms. OpenAI, Anthropic, and Google all publish Data Processing Agreements for their paid business plans. Their free consumer tiers do not.
- Disable training-on-your-data where the option exists. Both ChatGPT and Claude offer this toggle inside settings.
- Treat AI outputs the same way you treat a junior's draft. Reviewed by a human before it goes to a client.
The hard part nobody tells you
The hard part is not the technology. The hard part is the moment, three weeks in, when you realise the workflow you wanted AI to fix is actually a process problem dressed up as a technology problem. The reason your proposals take four hours is not because writing is slow. It is because nobody has documented what a good proposal looks like.
That is fine. That is actually the most useful thing AI does. It surfaces, very quickly, the parts of your business that were never properly documented in the first place.
Most of our augmentation work is exactly this. A senior operator embeds, watches the function for two weeks, then writes the brief that should have existed all along. After that, half the work runs itself.
Where to go next
- If you run an accounting partnership, the AI for month-end close guide is the next natural read.
- If you would rather skip the reading and talk through your specific operating shape, the discovery call ends with a one-page brief on which function to start with.