NexBDM Blog
AI for Bookkeeping: what it can reconcile, and what still needs a human
By NexBDM Team · 2026-08-28
Key takeaways
- AI is reliable at matching bank lines to invoices. It cannot produce a document you never received, and in South Africa the VAT deduction turns on holding that document rather than on the reconciliation looking tidy. That gap is where the human work actually sits.
AI is reliable at matching bank lines to invoices. It cannot produce a document you never received, and in South Africa the VAT deduction turns on holding that document rather than on the reconciliation looking tidy. That gap is where the human work actually sits.
AI for bookkeeping is reliable at matching. It will pair bank lines to invoices, spot duplicates, learn recurring suppliers and never make an arithmetic mistake. What it cannot do is produce a document you never received, and in South Africa the VAT deduction depends on holding that document, not on the reconciliation looking tidy.
That gap is the whole subject of this post. It is not a caution about robots. It is a specific, checkable line between two different problems that look like one problem on a screen.
What a reconciliation actually proves
A bank reconciliation is a matching exercise. You have a list of money that moved and a list of documents that say why, and the job is to join them. When every line is joined and nothing is left over, the reconciliation is complete.
Complete is not the same as deductible. A reconciliation proves that a payment has a document attached to it. It says nothing about whether that document is the kind of document SARS will accept.
This is where bookkeeping automation is quietly oversold. The matching problem is the part software is genuinely good at, and it is also the part that carries the least risk when it goes wrong, because a mismatch is visible. The document problem is invisible until an audit, and no amount of matching detects it.
The line SARS actually draws
SARS states the rule plainly in its own guide for vendors. In the ten principles at the front of the VAT 404 Guide for Vendors, Issue 15:
"You need to be in possession of documentary proof prescribed by or which is acceptable to the Commissioner to substantiate any input tax and/or other permissible deductions which you want to make."
And again in Chapter 10, on calculating the return:
"you cannot deduct input tax or any part thereof unless you are in possession of documents prescribed by or which is acceptable to the Commissioner."
The guide's glossary ties it to the section: input tax may only be deducted "if the vendor is in possession of the relevant documentary proof under section 16(2)(a) to (e)" of the Value-Added Tax Act 89 of 1991.
Read that as a systems requirement rather than a tax rule and it becomes clear what software can and cannot help with. Possession is a state of the world, not a state of the ledger. Software can tell you that you do not have the invoice. It cannot go and get it.
What AI does well here, and why
Take the optimistic half seriously, because it is real and it is most of the hours.
- Matching. Bank line to invoice, invoice to purchase order, payment to statement. This is pattern work at volume, which is exactly where a machine beats a person who has been at it for two hours.
- Duplicate and near-duplicate detection. The same invoice captured twice under two slightly different supplier names is a classic manual-capture error and a trivial machine one.
- Recurring classification. Once a supplier has been coded the same way eleven times, the twelfth is not a judgement call.
- Extraction. Pulling the supplier name, VAT number, date, serial number and amount off a photographed invoice is now ordinary rather than remarkable.
- Arithmetic and completeness. Machines do not skip a line or transpose two digits.
Notice what those five have in common. Every one of them is a comparison between things you already hold. None of them requires knowing anything that is not already in the pile.
The check almost nobody automates, and should
Here is the part worth taking away, because it is machine-solvable and almost always done by eye or not at all.
South Africa has two document standards, and which one applies is decided by the size of the transaction, not by the size of the VAT. SARS sets it out in Chapter 13 of the same guide:
- Consideration more than R5 000 including tax: a full tax invoice is required, under section 20(4).
- Consideration less than R5 000 including tax: an abridged tax invoice may be issued, under section 20(5), except where the supply is zero rated.
- Consideration less than R50: a tax invoice does not have to be issued at all.
The two standards differ by exactly two fields. A full tax invoice must carry the recipient's name, address and VAT registration number, and the quantity or volume of the goods or services supplied. The abridged version drops both. Everything else is common to the two: the words "tax invoice", "invoice" or "VAT invoice"; the supplier's name, address and VAT registration number; a serial number and date of issue; a full and proper description of what was supplied; and the price and VAT.
So the check is mechanical. Is the amount over R5 000? Then are the recipient fields present, and is the quantity stated? A person doing this by eye on the eightieth invoice of the month will miss one. A rule will not. This is the single highest-value thing to automate in a small bookkeeping function, and it is almost never the thing people reach for.
The trap: a commercial invoice is not a tax invoice
The guide separates the two explicitly. A commercial invoice is "any document, notifying the purchaser to make payment in respect of a transaction". A tax invoice "must contain all of the details prescribed in the VAT Act".
A reconciliation cannot tell the difference. Both are a PDF with a supplier name and an amount, and both will match the bank line perfectly. The reconciliation closes, the file looks complete, and the deduction is still exposed. This is the clearest example of the gap: the match succeeded and the requirement was not met.
Field-level checking closes it, because the required particulars are a list and a list can be checked. That is a machine job. Getting the supplier to reissue the document correctly is not.
What still needs a human
Four things, and they are not the four people expect.
- Deciding what the transaction was. A payment to a hardware store is repairs, or it is an improvement to be capitalised, or it is personal. Nothing in the document says which. That is a decision about intent, and intent is not in the data.
- Getting the missing document. Chasing a supplier is a relationship task with a deadline attached. Software can raise the flag and draft the request. Somebody still has to make it happen before the return is due.
- The last-resort route. Where the correct document cannot be obtained, section 16(2)(g) allows a deduction on alternative documentary evidence acceptable to the Commissioner, in terms of a ruling. SARS is explicit that this is a last resort, available only after all reasonable measures have been taken to obtain the correct documentation from the supplier. That is a judgement and an application, not a setting.
- Anything with an exception in it. The moment a transaction is unlike the eleven before it, the pattern that made the machine useful is the reason it is now wrong.
The professional research points the same way from the other direction. A study published by ICAS on 17 March 2026, conducted by Alliance Manchester Business School and Aston Business School across more than 200 accounting professionals at one UK firm, found that 72 percent of respondents worry generative AI could produce errors or reach incorrect decisions, and concluded that the technology "performs best on routine, structured tasks" while work demanding nuanced understanding or ethical judgement continues to rely on people. Those were UK practitioners, not South African ones, and the tax rules above are ours alone. The shape of the finding still travels.
Where the manual work actually stops
None of this is useful without saying what changes in practice. Five mechanisms, in the order they pay off:
- Capture once, at the point of spend. The invoice is photographed or forwarded when the payment happens, not gathered at month end. The fields are extracted then, while somebody still remembers what the purchase was for. Every later artefact reads from that record instead of re-keying it.
- Validate at capture, not at reconciliation. Run the threshold test and the field check the moment the document arrives. A supplier will reissue an invoice from three days ago without complaint. Three months later, at year end, they will not.
- Match automatically, review exceptions only. The reconciliation should present the handful of lines it could not join, not the hundreds it could. Attention goes to the residue.
- Make the missing-document list a live queue, not a month-end discovery. If the system knows a payment has no compliant document attached, that is a task with an owner and a due date from day one.
- Keep the record where it can be found. Vendors must keep their documentary proof and other transaction records for at least five years, per the same SARS guide. A folder nobody can search is a retention policy in name only.
Read together, those five move the work from reconstruction to capture. Reconstruction is what makes bookkeeping expensive: the cost is not the entry, it is the four weeks of distance between the spend and the person trying to work out what it was.
How this differs from the neighbouring problems
Three related posts already cover ground next to this one, so it is worth being clear about the boundaries.
- Automating data entry is about getting information off documents and into systems accurately. This post assumes that has happened and asks whether the result is defensible.
- Tax invoice requirements covers what a compliant invoice you issue must contain. This post is the mirror image: what the invoices you receive must contain before you may deduct.
- Tracking business expenses for SARS is the habit and the filing. This is the validation layer that sits on top of it.
Worth reading alongside: invoicing for small business, why AI projects fail, and the real cost of manual admin.
Frequently Asked Questions
Can AI do my bookkeeping without an accountant?
It can do the matching, extraction and arithmetic without one. It cannot decide what an ambiguous transaction was, obtain a document you do not hold, or make a judgement call where the rules allow discretion. Those remain a person's work.
Does a perfectly reconciled bank statement mean my VAT is safe?
No. Reconciliation proves a payment has a document attached. SARS requires possession of documentary proof that meets the Act's requirements. A matched line supported by a non-compliant document is reconciled and still exposed.
What makes a tax invoice valid in South Africa?
Above R5 000 including VAT you need a full tax invoice under section 20(4), carrying both parties' details, a serial number and date, a proper description, the quantity and the price and VAT. Below R5 000 an abridged invoice may omit the recipient's details and the quantity.
What if a supplier will not give me a compliant tax invoice?
Section 16(2)(g) allows a deduction on alternative evidence acceptable to the Commissioner in terms of a ruling. SARS treats it as a last resort, available only once all reasonable measures to obtain the correct document have been taken. Ask a practitioner before relying on it.
Where should automation start in a small bookkeeping function?
At capture, not at reconciliation. Extract and validate the invoice fields the day the money moves, while a supplier will still reissue a wrong document and while somebody still remembers what the purchase was for.
Where to start
If your month end is a reconstruction exercise, the reconciliation is not the bottleneck. The distance between the spend and the record is. A business autopsy maps where that distance opens up in your specific process and what closing it would actually take. If you would rather talk it through first, start with a discovery call. Our payments and reconciliation work sits under NexPay.
Sources
- SARS, VAT 404 Guide for Vendors, Issue 15: ten important principles, Chapter 10, Chapter 13 and glossary. Value-Added Tax Act 89 of 1991, sections 16(2), 20(4) and 20(5).
- ICAS, "AI can't replace human judgement in accounting, ICAS study shows", 17 March 2026, reporting research by Alliance Manchester Business School and Aston Business School.