For a South African accounting or bookkeeping practice, the useful AI is the boring AI: the one that answers "what do you charge", "what do I need to bring", "when is my provisional tax due" and books the meeting — plus the one that turns a quote into a compliant invoice without anybody retyping it. Nothing here does the accounting. It removes the admin around it.
Where does a practice lose time?
- The same eight questions, every February and every August. Provisional tax dates, what a first meeting costs, whether you handle payroll, whether you do trusts, whether you take on a business that is behind on submissions.
- Chasing documents. Bank statements, invoices, the logbook, the thing the client promised on Tuesday.
- The enquiry that arrives at deadline season and gets answered a week late — which is precisely when a prospective client is most likely to be phoning three practices.
- Quoting and invoicing. Rekeying a quote into an invoice, then into a statement, then chasing it.
Which of those does a chatbot fix?
| Job | Website chatbot | Quoting and invoicing | A person |
|---|---|---|---|
| "What do you charge for a small company?" | Answers from your own published fees | — | Bespoke scopes |
| "When is provisional tax due?" | Answers | — | — |
| "What must I bring to the first meeting?" | Answers, with your checklist | — | — |
| Booking the meeting | Yes | — | — |
| Quote out the same day | — | Yes | Approves it |
| Quote becomes an invoice | — | One tap, no retyping | — |
| Anything about a client's actual tax position | Never | — | Always |
Why must it refuse tax questions?
Because a confident wrong answer about a deduction, a threshold or a deadline is worse than no answer, and a language model will produce one fluently. There is a clean line: publishing a SARS deadline is information anyone can look up, while telling a particular business what it owes is advice about their facts. The bot stays on the first side of that line and escalates the moment a question crosses it — configured as a refusal, not as a suggestion.
What does it cost?
A small practice is a Starter chatbot: R299 a month plus R250 once-off — R549 at checkout. The quoting and invoicing side is separate and cheaper than most people expect: our Bookkeeping & Invoicing product starts at R80 a month with no setup fee, so the first payment is R80 and there is nothing to pay up front. Both are month to month.
Does it help with the document chase?
Partly, and it is worth being precise about how. It can send the reminder reliably, on schedule, to the right client, in their language, which is the part that currently depends on somebody remembering. It cannot make a client open their banking app. Practices that measure this usually find the reminder is not the binding constraint — consistency is, and consistency is exactly what a system supplies and a busy February does not.
What about client data and POPIA?
Financial records are personal information and often sensitive. The safe shape is the same as everywhere else: collect the minimum at first contact, keep it in a store belonging to your practice rather than pooled, tell the client what you are collecting, and be able to delete it. A first-contact bot does not need a client's numbers — it needs a name, an entity type and a contact detail. See AI and POPIA for what the Act actually asks.
Nothing here does the accounting. It stops a partner answering the fee question for the ninth time in a week, and it stops a quote from waiting three days for somebody to find a moment.
What would I do first?
Look at how long a quote currently takes to leave the office. If it is same-day, start with the chatbot and the after-hours enquiries. If it is three days, start with quoting — it is R80 a month and it removes the retyping entirely.
Where we do this
If you are reading this from Soweto, Midrand and Paarl, those pages go through the local version of it: which languages the bot needs, what the busy hours look like, and which sectors ask what.