It is early August. The 31 August provisional tax deadline sits three weeks out. A tax practitioner in Rondebosch, seventy clients on IRP6, opens her inbox and finds the same three questions repeating: "how much am I paying?", "did you get my July bank statement?", "is that the final number?". None of them needs her judgement. All of them will get it, over the next fortnight, at eight minutes each. Or she picks which ones she reaches before Sunday night.
This is the shape of a tax practice in South Africa. Deadlines are fixed. Client communication is dense, repetitive, and time-boxed. The actual tax work — the interpretive judgement your Practitioner Reference Number was issued for — is a small share of the hours you bill. Everything around it is chasing documents, forwarding SARS letters clients did not read, and explaining why an ITA34 came in higher than they hoped.
What follows is a practical read on where AI helps in that environment, and where it does not. I am writing this for the smaller practices: the sole practitioners, the two- and three-partner tax firms, and the tax desks embedded inside SAIPA or SAIT bookkeeping practices, where nobody has ever had time to properly automate anything. That is where the gap between how the work runs today and how it could run is widest.
Tax practice work is not bookkeeping work
A brief but useful distinction. A bookkeeping practice's largest time sink is usually source-document coding into Xero, Sage or Pastel. A tax practice's largest time sink is chasing the twelve provisional-tax clients who have not sent their six-month management accounts, then explaining the estimate calculation four different ways, then re-uploading the IRP6 because eFiling has logged the partner out again.
The workflow shape is different. The automation shape follows.
Once you accept that, the useful projects fall out naturally.
The provisional tax cycle: where automation earns its keep first
If you are going to automate one thing in an SA tax practice, automate the provisional tax cycle. Twice a year, February and August, every practitioner runs the same sequence: identify which clients are on IRP6, request their latest management accounts or turnover figures, calculate the estimate, get sign-off, submit through eFiling, and confirm payment. On a book of a hundred provisional clients, that is several weeks of near-identical work every six months.
A structured workflow, built around whatever practice management tool you already use, handles the parts that do not need you. Six weeks out from the deadline, each provisional-tax client gets a personalised WhatsApp or email listing exactly what you need for their entity: management accounts to the interim date, any once-off income they already know about, confirmation of medical scheme and RA contributions if it is an individual. Two weeks out, the ones who have not responded get a firmer follow-up. One week out, the partner sees a short escalation list on a single screen. Three clients still outstanding. Call them.
The tax calculation itself stays with you. That is not the work AI should be signing off, and no controlling body will thank you for outsourcing it to a chatbot. What automation removes is the twenty small acts of communication around each calculation. That is where the hours are.
In my experience, practices that get this right stop dreading provisional weeks within one cycle. Not because the work is less. Because it stops arriving in the last four days.
SARS auto-assessments changed the shape of filing season
SARS's auto-assessment programme has expanded every year since 2021 and now covers a meaningful share of individual taxpayers. If your practice serves individuals, this has changed what filing season actually is. It used to be about collecting IRP5s, medical certificates, and RA certificates, then filing. Now, for many clients, an ITA34 has already landed in their SARS profile before the practitioner has finished her kettle in July.
The client's question has changed too. It is no longer "when are you filing my return?". It is "SARS says I owe R14,600. Is that right?". You have forty business days to accept, edit, or reject each auto-assessment, and the clients who do not understand what SARS has done often panic on WhatsApp at 20:00 the day the ITA34 lands.
This is a genuinely useful spot for AI. Not to review the auto-assessment — that is your work — but to handle the first-round client communication. A trained assistant, given the client's ITA34 and your prior-year working papers, can produce a plain-language explanation of what SARS pre-populated, what changed year-on-year, and what you would check before advising the client to accept. The client reads that within an hour of the auto-assessment landing. Their anxiety drops. The follow-up call, when it happens, is about substance instead of reassurance.
Do not let it draft the objection. Do not let it commit to a position. The tax practitioner drafts, reviews, and signs.
SARS correspondence: a triage problem, not a drafting problem
Every tax practitioner receives a steady drip of SARS letters on behalf of clients. Verification requests. Information requests. Revised assessments. PAYE reconciliations gone wrong. VAT audit letters. Some are urgent, some administrative, and some just want a supporting document uploaded to eFiling within twenty-one days. The problem is not writing the response. The problem is spotting the ten letters in a busy week that actually need work in the next 72 hours.
An AI triage layer on your SARS-notice inbox, one that reads each incoming notice, extracts the reference, the client, the type, and the deadline, and puts it on a dashboard sorted by urgency, takes an hour a week off the partner's desk. More importantly, it catches the letter that would otherwise sit unread in a colleague's inbox until day nineteen.
The drafting itself stays yours. SARS responses are technical, they build a paper trail, and the wording matters for future disputes. AI-drafted SARS letters are only ever useful as a starting point for the routine, factual replies: an IRP5 upload, a correction of banking details, a proof-of-payment. For anything touching interpretation, position, or dispute, you write it.
Where AI does not help in a tax practice
Worth being clear on this. There are parts of the work where AI is not the answer today, and pretending otherwise costs you time and credibility.
- Notices of Objection and Appeal. A NoO or NoA lives and dies on your reading of the Tax Administration Act, the underlying legislation, and the specific facts of the assessment. AI does not draft one that will stand up in front of a SARS reviewer, and you do not want to test it on a client whose money is at stake.
- Voluntary Disclosure Programme applications. The scope of disclosure, the wording of the facts, and the timing all carry real consequences. This is partner work, in the room, with a coffee that has gone cold.
- SARS in-person audits. If a SARS auditor is at your client's premises, the practitioner is on the phone or on-site. No bot.
- Signed tax opinions. A written tax opinion carries professional-indemnity implications. It is not something to speed up with an AI drafter.
- Advising on aggressive structuring. Anything that could reasonably attract the General Anti-Avoidance Rule needs a human sitting across from another human, thinking about a specific set of facts.
Being honest about these limits is what separates a durable automation programme from one that quietly gets rolled back after the first mishap.
POPIA, the Tax Administration Act, and where client data actually sits
Tax practitioners hold some of the most sensitive personal data any professional handles. Full ID numbers, banking details, medical scheme certificates, income breakdowns, third-party dependants, sometimes divorce settlements and estate documents. POPIA obliges you to protect that data. The Tax Administration Act layers further confidentiality on top: you cannot disclose taxpayer information without authority, and your controlling body cares about how you handle it in practice, not only on paper.
Any automation you introduce has to respect three things. Client documents do not flow through consumer AI tools whose data terms you have not read. Access to client data is logged, per user, so that a future POPIA request or an Information Regulator query can be answered without a scramble. And your written authority to act on the client's behalf — the SARS Special Power of Attorney, the RCB-required engagement letter — is captured and stored inside whatever intake flow you build.
If a prospective provider cannot walk you through where each of these lives on their infrastructure, that is the answer.
Where to start, honestly
I would not begin with SARS correspondence triage, tempting as it is. It is a smaller relative win, and it depends on plumbing into your notice inbox that varies by practice. Start with the provisional tax cycle, because the calendar is fixed and the win is measurable within six months. The February and August peaks are the ones that break small practices. Moving the client-communication load off the partner in the run-up frees them to do the actual estimate work with a clearer head, which is what clients actually pay you for.
The second project, once provisional is stable, is client-facing auto-assessment explanations during filing season. It is a July and August problem, largely, but a growing one as SARS keeps extending auto-assessment coverage.
The tax practitioners in Cape Town and Gauteng who have built this out in roughly that order report the same pattern. Provisional weeks stop being brutal. Filing season feels less like a hostage situation. And the interpretive work, the reason clients pay you at practitioner rates rather than doing it themselves on eFiling, gets the attention it always deserved.