Monday morning in Centurion. A collections manager at a Debt Collectors Act-registered firm has four thousand arrears accounts across RCS, Mr Price Money, Capitec revolving and two medical-aid gap-cover portfolios. Debit orders bounced over the weekend. Twelve default-judgment matters sit in the magistrate's court diary for the week. Four Section 129 letters are stuck in the drafts queue because the admin manager was on leave. A collection agent in Alberton is on WhatsApp with a consumer who has promised R500 on Friday, and the agent is not sure whether that counts as a Section 129(3) arrangement or a voicemail-level undertaking. Three debtors have moved address since the last trace, and the bureau report the firm pays for arrived with forty percent of the records flagged "insufficient match".
That is the week in most SA debt-collection and legal-recoveries practices running mid-size books. Not the credit decision. Not the attachment application. Not the court appearance. The thousand small admin motions in between, every one bound by regulation that the Debt Collectors Council and the Information Regulator will enforce if a complaint lands.
AI will not collect a debt for you.
Any provider selling you an autonomous collection agent in the South African market has either not read the National Credit Act or is hoping you have not. What AI can do, placed carefully, is the paper and the chasing — the part that eats headcount but produces no money by itself.
Where the week in a SA recoveries practice actually goes
Open a well-run collections desk and the admin breaks into roughly the same buckets, whether the firm is a Debt Collectors Act registrant or an attorney running recoveries under the Legal Practice Act.
- Trace work when debtors have moved, changed phones, changed employers
- Section 129 drafting for every account about to escalate
- Letter-of-demand follow-up and the paper trail of return receipts
- Payment-arrangement capture and tracking against the next debit-order run
- Court-roll preparation for default judgments and Section 65A proceedings
- Debit-order resubmission across the various DebiCheck mandates
- Consumer contact across phone, SMS, WhatsApp and email
- Monthly reconciliation back to the mandator — retailers, medical aids, bond originators, municipalities
Each one is structured, repetitive, high-volume. None of it is where the money comes from. All of it has to happen correctly, in a legally defensible way, every month.
Trace work: where AI quietly earns its keep
A trace file at a mid-size SA practice is three or four bureau reports plus the agent's own notes. TransUnion, Experian, Compuscan and XDS each report differently. The agent reads three bureaus, cross-references employer, address and known phone numbers, flags inconsistencies, and decides which trail is worth phoning.
An hour per difficult trace. On a four-thousand-account book with roughly five percent untraced, that is two hundred hours a month before anyone has picked up a receiver.
AI reads structured bureau output faster than a person and holds four reports in context without losing track of the field names changing between formats. A well-built trace workflow ingests the bureau extracts, reconciles the fields, flags where employer, address or phone has shifted since the last trace, ranks the leads by freshness and consistency, and hands the agent a one-page tracer brief in place of three dense PDFs. The agent still phones, still verifies, still decides whether to escalate.
What it does not do: pull a fresh bureau report on its own. The consent for a credit-bureau inquiry has to be documented, the purpose has to be legitimate under POPIA, and the inquiry must show up in the audit trail with a human agent's name against it. "The AI pulled it" is not a defence the Information Regulator will accept, and I have not seen that argument survive a response letter from Johannesburg.
Section 129 letters: drafted by AI, signed by a person
Section 129(1)(a) of the National Credit Act is where most SA collections files either stand up or fall over in court. The letter has to go to the correct address — the debtor's chosen address under Section 65 of the NCA — and it has to be sent by a method the court will accept as service (registered post is the standard, with the slip retained for the file). The content has to match the account. The dates have to be right. The in duplum rule under Section 103 has to be honoured, which means the calculation is the firm's own, not the retailer's statement.
Most SA firms still draft Section 129 letters from a Word template with fields replaced by hand. The per-letter error rate is low. It still corrupts a few files a month, each of which can collapse an application for default judgment the magistrate was otherwise going to grant.
An AI draft reads the mandator's statement, the firm's own ledger, the debtor's chosen service address and the current in duplum ceiling. It produces the letter populated correctly, flags any mismatch between the mandator's balance and the firm's ledger, and marks the account for human sign-off before dispatch. In my experience, firms that put this in run their Section 129 cycle in a day instead of a week, and the court-roll fallout drops to near zero.
What it does not do: sign or post the letter. The letter leaves the office under the signature of a Debt Collectors Council-registered collector or an admitted attorney, and the registered-post slip or sheriff's return is kept with the file. Automate past that signature and the regulator starts asking why.
WhatsApp and SMS: compliance before cleverness
WhatsApp is where South African debtors actually read messages. Any practice ignoring the channel is leaving contact rate on the table. Any practice using it badly is building a file for the Information Regulator.
The Council for Debt Collectors Code of Conduct and POPIA together draw lines you cannot blur:
- No contact before seven in the morning or after seven in the evening unless the consumer has consented in writing to extended hours
- No contact on Sundays or public holidays
- No messages to the debtor's workplace in a way that reveals the debt to colleagues
- No messages to third parties (a spouse, a parent, a neighbour) about the account
- Clear sender identity in every message, with the firm's CDC registration number available on request
- A working opt-out that stops the automated stream the moment the consumer uses it, logged with a timestamp
A WhatsApp workflow built with those rules as hard constraints is a reasonable use of AI. The system schedules messages inside allowed hours, routes replies to a human agent when the consumer uses words like "lawyer", "complaint", "ombud" or "overpaid", and keeps the audit log that will matter if the Information Regulator comes asking. A payment arrangement captured in the thread flows to the firm's system as a tentative promise, triggers an automated reminder two days before the agreed date, and escalates to the agent if the date passes without payment.
What it does not do: handle a dispute. The moment a consumer disputes the debt, raises harassment, or mentions legal advice or an ombud, the thread moves to a person. A bot answering a Section 134 harassment complaint is the kind of file that ends up before the Debt Collectors Council, and the lawyer's bill arriving afterwards is the second-worst part.
In duplum, prescription, and the parts AI should not touch
The maths of a collections file is deceptively simple until it is not. The in duplum rule caps unpaid interest, charges and fees at the amount of the capital still outstanding at the moment the account went into default. Section 126B of the NCA changed how prescription works for credit agreements. The Judicial Officer in a default-judgment application will check both.
AI can calculate the ceiling. It should not sign off on the figure that goes on the summons. The firm's collector or attorney is personally accountable to the court for every number in the papers, and a model that quietly hallucinated an R12,000 interest line on a R45,000 balance is not something the magistrate will forgive on the day. In practice, the split that works looks like this: the AI produces a one-page in duplum and prescription summary with the maths shown, and the collector or attorney signs off before the papers go to the sheriff.
The decision to issue summons, to apply for an emoluments attachment order, to oppose a Section 86 debt-review referral, to accept or reject a settlement: none of that is for a model. It is for the person whose name goes on the warrant.
Where to start: narrow, auditable, defensible
Firms that try to automate everything on day one end up with a system that handles seven things badly and that nobody in the practice trusts. Six weeks later it is quietly switched off, and the trial-and-error cost sits on the P&L for the rest of the year.
The right starting project for most SA recoveries practices is tracing plus Section 129 drafting, in that order. Both are high-volume and structured. Both have clear audit trails. Both leave the final sign-off with a registered person. The trace layer produces leads the agent can phone the same day, and the 129 layer produces letters the collector signs the same morning. Compliant WhatsApp capture of payment arrangements follows once the first two are running cleanly.
The firms in Centurion, Bedfordview and the Southern Suburbs that have built it in this order report the same shift. Trace coverage up from fifty to eighty percent of previously untraced accounts within a month. Section 129 cycle time down from a week to a day. The admin manager stops dreading Monday mornings. And the collectors spend their calls on the files that will pay, not on the paper that proves they tried.