AI debt collection uses AI to reach customers who are behind on payments: voice bots for early reminders, promises to pay and payment links, and AI practice calls that train collectors for the hard conversations. The AI must confirm identity first, follow the Debt Collectors Act, the National Credit Act and POPIA, and hand disputes and hardship to people.


“ShiftMate builds voice AI in South Africa and trains collections teams with AI practice calls. We think AI should take the early reminders, and people should handle hardship and disputes. This guide covers AI debt collection within the National Credit Act, POPIA and the debt collectors' code, and where people must stay.”
AI debt collection is the use of AI in the collections process, mostly in two ways. The first is an AI collections agent on the phone: an AI voice bot that calls customers soon after a missed payment, confirms who it's speaking to, explains the amount, takes a promise to pay or sets up an arrangement within limits, and sends a payment link. The second is AI on the training side: an AI customer that your human collectors practise on, so they're ready for the calls the bot can't take.
Both are real in South Africa. The BPO industry body BPESA republished a November 2026 report in which a local business said its agentic AI collections agent "dealt with more than 100 000 people every day" (BPESA). The same report quoted a global contact centre company saying AI adoption had forced it to extend training "from an average of three weeks in 2024/25 to eight weeks in 2025/26". Those two facts belong together: when AI takes the easy calls, the calls left for people get harder.
ShiftMate's part in this is the training side: we sell collections teams AI practice calls that prepare collectors for exactly those harder calls, and most of this page's first-hand lessons come from building them. If you want the general picture first, start with our guide to AI voice agents.
An AI debt collection voice bot can handle the calls that follow a script and stay within limits you've set: reminders, promises to pay, simple arrangements and payment links. Anything that needs judgement, care or authority (a dispute, hardship, debt review, a settlement) belongs with an experienced person.
| Call or situation | Good for the AI | Keep with experienced people |
|---|---|---|
| A missed payment, days past due | A friendly reminder, a payment date, an SMS payment link | Repeated broken promises |
| Payment arrangements | Arrangements inside set limits (amount, number of months) | Anything outside the limits, settlements, write-offs |
| "I've already paid" | Checking the system and confirming if it shows | Anything the system doesn't show: proof, allocation errors |
| Disputes | Logging that there is a dispute and stopping | The dispute itself |
| Hardship and distress | Spotting it and handing over at once | The conversation: job loss, illness, a death in the family |
| Debt review or legal stage | Recognising it and handing over | Everything after that |
| Someone else answers | Saying nothing about the debt, asking for a good time to call | Any request for details from a third party |
| "I want a person" | Transferring straight away | The call |
Two things decide whether a bot helps or harms. The first is whether it stops in time: an agent that pushes on after "I lost my job" does real damage to the customer and to your complaints record. The second is whether its records are right: a promise to pay that's logged on the wrong account, or a payment confirmed that never arrived, creates a second problem on top of the first.
If you searched for debt collection AI South Africa rules, the short answer is that an AI must follow the same rules as a human collector. Debt collectors must follow the Debt Collectors Act and its code of conduct, credit providers the National Credit Act, and everyone handling the customer's information POPIA. Using an AI doesn't change who is responsible: the business that runs the calls is.
“Our agents say they're an AI and that the call is recorded. People forgive a machine that's honest with them; they don't forgive one that pretends.”
Because the calls AI can't take are the hardest ones, and they're now most of what your people do. When a bot handles the reminder and the simple promise, the human collector's day fills with disputes, hardship, debt review, broken promises and angry customers. That's the shift behind the training jump from three weeks to eight reported in the BPESA article.
Collections is also a job where a bad call has consequences beyond one customer: a complaint, a conduct finding, a customer who stops paying altogether because they felt bullied. New collectors usually learn these calls on real customers. AI practice calls change that: the collector takes the hard call twenty times on an AI customer before a real one hears it.
A specific date and amount the customer can actually manage, not a vague "end of the month" that breaks.
Stop collecting, log the dispute, explain what happens next. No arguing the amount.
Slow down, listen, offer what your policy allows, refer where needed. No pushing for a payment today.
Check, ask for proof calmly, never call the customer a liar.
Say nothing about the debt, leave a neutral message.
Give a way to check that doesn't depend on trusting the caller: the official number, the website, the account letter.
An AI collections practice call looks like a real call: the collector dials, an AI customer answers in a South African voice with a background, a mood and a reason they haven't paid, and the call is scored afterwards against a scorecard, with quotes from the call. Here is a made-up practice call, with the coaching notes a collector would see.
Good morning, this is Priya from Harbourview Furnishers. Am I speaking to Mr Johan Botha?
Ja, speaking.
For security, please confirm your ID number's last four digits.
Identity before the debt. Good.
Four, one, two, eight.
Thank you. Your account is three months behind, R4,800. Can you pay the full amount by Friday?
Look, I lost my job in September. I've got nothing coming in. I can't pay by Friday.
I understand, but if we don't get something this week the account goes further. Could you do R500?
Missed the hardship signal and pushed. This is the moment the scorecard marks down.
I just told you I've got nothing. Are you even listening?
Sorry, Mr Botha, you're right. Let me stop there. Tell me a bit about where things stand, and I'll explain what options we have for people who've lost their income.
Recovered: acknowledged, stopped pushing, moved to the hardship process.
The useful part is what happens next: Johan doesn't suddenly agree to R500 because the collector apologised. A customer with no income who agrees to pay anyway teaches the collector that pushing works. That was a real problem in our own testing.
“We tested whether our AI customers gave in too easily. In 21 practice collections calls, customers who could pay nothing agreed to pay anyway in 9 of 12 calls, as often with a good agent as a weak one. Practice against a pushover teaches the wrong thing, so we measure it before we trust it.”
So we measure our AI customers before we trust them: a customer who can't pay should hold that line whatever the collector says. The score that follows has to be one a team leader can check:
| Scorecard item | Mark | Quote from the call |
|---|---|---|
| Confirmed identity before mentioning the debt | Met | "For security, please confirm your ID number's last four digits." |
| Recognised hardship and stopped pushing | Not met at first, then met | "Could you do R500?" then "Let me stop there." |
| Offered the hardship process | Met | "I'll explain what options we have for people who've lost their income." |
| No pressure or threats | Partly met | "If we don't get something this week the account goes further." |
“When we first had AI mark every practice call against a scorecard, about one review in three came back cut off and unusable. We changed it so every mark has to quote the words from the call, checked in code. After that, the same stored calls reviewed three more times got the same marks: 18 of 18 on one call. A score nobody can check is worse than no score.”
Track whether promises are kept and whether customers were treated fairly, not just how many calls were made. A promise-to-pay rate that rises while kept promises and complaints get worse means your people, or your bot, are pushing too hard.
| Number | What it tells you | Watch out for |
|---|---|---|
| Right-party contacts | Whether you reach the actual customer | Debts mentioned to the wrong person |
| Promises to pay per right-party contact | How often a call ends with a plan | Promises the customer can't keep |
| Promises kept | Whether the plan was real | The best single measure of a good collections call |
| Hardship referrals | Whether people spot hardship | Zero referrals usually means missed signals, not no hardship |
| Complaints and disputes logged | Fair treatment, as customers see it | Rises after a new script or a new bot |
| Compliance score from call reviews | Identity, disclosures, no pressure | Scores without quotes nobody can check |
| Practice pass rate by situation | Who is ready for which calls | Passing the easy situation only |
Start with the four that cost you most: hardship, disputes, broken promises and "I've already paid".
Your identity steps, your hardship process, your arrangement limits and your forbidden phrases.
Check that a customer who can't pay doesn't give in, and that one who disputes the debt keeps disputing it.
New collectors pass each situation before they phone real customers; experienced ones practise the situation their reviews flag.
After a month, look at promises kept, complaints and hardship referrals for people who practised against those who didn't.
“A rule we wrote for outbound calls broke inbound ones: it flagged a perfectly normal caller saying "Hi, I'm phoning about my fibre" as a mistake. The same sentence is right on one kind of call and wrong on another, so we test every call type on its own.”
In collections that means an inbound call from a customer who wants to arrange payment and an outbound reminder call need separate checks. The opening that's right on one is wrong on the other.
Choose an AI collections agent the way you'd vet a new collections partner: by how it handles the calls that go wrong, not by the demo. Ask every supplier these questions and ask to hear recordings, not slides.
For how voice agents go wrong in practice and how to test them, see why AI voice agents fail. For outbound calling rules in general, see our page on AI cold calling.
Collections voice bots are usually custom priced, because the work is in the compliance sign-off, the account system connection and the testing, not the minutes. Published per-minute prices depend on what they include: one developer platform lists US$0.07 to US$0.31 a minute, built from itemised parts (voice engine, voice, AI model and a US$0.015 phone line), while hosting-only building blocks start at about US$0.01 a minute and bill the AI models and phone line separately (vendors' own pricing pages, checked 16 November 2026; see AI voice agent pricing). Listed prices for South African AI voice agent services run from R849 a month plus calls to R7,500 a month, with several bundles around R1,900 to R2,400 a month including a set number of minutes and then about R0.97 to R13 a minute; some are ex VAT (WhichVoIP, a third-party comparison site, 31 October 2026, not confirmed on each supplier's own page). Put the cost next to promises kept and complaints, not calls made.
AI debt collection doesn't solve affordability. A customer with no income can't pay, however good the call. It doesn't fix bad data, a harsh policy or a collections culture that rewards pushing. And it can make some problems worse at scale.
More calls to customers who can't pay produce more broken promises and more complaints, not more money.
A collector who says the wrong thing does it on one call. A bot does it on every call until someone notices.
Grief, anger and shame are where people earn their keep. A bot should hand these over, and it needs to notice them first.
Off-the-shelf speech recognition makes several times more errors on South African-accented English than on standard US benchmarks (AfriSpeech-200, 2023). Account numbers, ID numbers and rand amounts are where it hurts.
AI practice calls build skill, but they don't replace listening to real calls, coaching and fair targets. Measure the live numbers.
We think AI debt collection should take the early reminders, and people should handle hardship, disputes and debt review. That makes trained collectors matter more, not less: the calls left for people are the hard ones. Our view goes further: how a floor treats the customer who can't pay decides whether it keeps a customer or gains a complaint.
So for collections we build AI practice calls with AI customers in South African voices who don't pay, dispute the debt, have lost their job or get angry, scored with quotes from each call, with plans on the Sim Studio pricing page.
We want to work with collections teams and call centres that want collectors ready for the floor. Weighing an AI collections agent too? Tell us how your book splits between early reminders and hard cases, and we'll show where voice AI fits and where people must stay.
“In a call centre, voice AI is support, not a replacement. It trains the people, generates the leads, and makes the low-value calls the operators don't want to make, so they spend their day on the conversations that need a person.”
AI debt collection is the use of AI in collections: voice bots that make early-stage reminder calls, take promises to pay and send payment links, and AI practice calls that train human collectors for disputes, hardship and angry customers. The rules for collectors apply to the AI too.
Automated collection calls aren't banned, but debt collectors must follow the Debt Collectors Act and its code of conduct, credit providers the National Credit Act, and everyone POPIA. The bot must confirm identity first and never pressure customers. Have compliance sign off the script and take legal advice.
Disputes, hardship, debt review, legal-stage accounts, settlements outside set limits, and anyone who asks for a person should stay with experienced collectors. An AI collections agent should recognise these situations quickly and hand over, not try to handle them.
Collections voice bots are usually custom priced. Published prices depend on what they include: one developer platform lists US$0.07 to US$0.31 a minute, built from itemised parts (vendor's own pricing page, checked 16 November 2026). ShiftMate's AI practice calls for collectors have plans on the Sim Studio pricing page.
The common mistakes are mentioning the debt before confirming identity, pushing on after a customer mentions hardship, arguing a dispute instead of logging it, confirming payments the system hasn't recorded, testing only on easy customers, and practising against AI customers who agree to pay too easily.
For collections teams, ShiftMate sells AI practice calls that train agents on promises to pay, disputes and customers in hardship, scored after every call with quotes from the call. Its voice agents are for sales and inbound calls, not collections.
They let collectors take the hardest calls many times before a real customer hears them. An AI customer plays a person who can't pay, disputes the debt or is in distress, and the call is scored against a scorecard with quotes from the call, so team leaders can check every mark.
Yes, for debt collectors: in the US, the FDCPA and Regulation F apply to debt collectors whether a person or an AI makes the call, including limits on call frequency. The FCC ruled in February 2024 (its AI-generated voices robocall ruling) that AI voices are artificial under the TCPA, so prior express consent is needed. Take US legal advice.
AI practice calls for collections agents: promises to pay, disputes and customers in hardship, scored after every call with quotes you can check.