Global business process outsourcing (BPO) and credit-lifecycle management company Nutun launched its new AI Lab in Johannesburg last night.
The team introduced Zoey, an agentic AI debt collector that is already recovering close to R10 million a month.
“Zoey is live. Zoey is working. Zoey is operating at scale,” said Nutun South Africa CEO Robert Amoils.
Zoey has been operational on Nutun's R30 billion principal portfolio for the past year, working across more than 80 000 accounts.
“We are not slightly ahead, we are far ahead,” said Amoils. “We built it ourselves, we’ve tested it on ourselves and we are ready to put it to work.”
The numbers
Frikkie Craucamp, Nutun chief analytics officer, led Zoey’s development. Conceived roughly two years ago and built entirely in-house, last month, for example, Zoey made more than 1.8 million calls and collected almost R7 million, he noted.
Since the start of its financial year, Nutun has collected more than R40 million through Zoey.
During Craucamp’s 30-minute presentation, Zoey was running in real-time, and attempted almost 6 000 calls, logged about 15 hours of live conversation and collected more than R20 000.
By the end of the day, Craucamp estimated that Zoey would have attempted approximately 90 000 calls, logged 230 hours of talk time and collected around R330 000.
How Zoey works
The agentic AI debt collector verifies a customer’s identity, negotiates a payment arrangement and concludes it without a human agent on the line, drawing on what Nutun calls its ‘master data universe’ of behavioural and credit data built up over years of collections work.
Zoey isn’t AI bolted onto collections, but collections rebuilt around AI, Craucamp explained.
The system predicts how likely a customer is to pay and how much they can realistically afford, then adjusts its tone and approach accordingly in real-time. When a call needs a human touch, such as a complex arrangement, Zoey hands it to a live agent along with a full summary, so the customer doesn’t have to repeat themselves.
“As we’ve changed and evolved Zoey’s architecture, we’ve found that the propensity-to-pay measure can be far better in terms of conversion rates, but the pipeline is far lower,” said Amoils. “It’s a careful calibration as to what we’re trying to achieve.
“About five months ago, we changed Zoey’s harshness, and in certain parts of our portfolio that resonated very well – we secured a bigger pipeline, slightly lower conversion, but a better net outcome,” continued Amoils. “In other areas of the portfolio it didn’t resonate as well. That suitability and flexibility per campaign is a very big part of Zoey.”
Humans in the loop
Zoey’s second function is listening to and scoring every call made by Nutun’s human agents, assessing each one against more than 100 measures, including whether the right party was reached and overall call quality.
Although the coverage currently spans Nutun’s own acquired debt book in English, Craucamp said it will extend to outsourced collections and additional South African languages by July.
According to Nutun CIO Hans Zachar, the build took around three years, beginning shortly after the launch of OpenAI’s ChatGPT. Rather than training a model from scratch, the team assembled existing large language models and layered Nutun’s own collections analytics on top.
A bigger challenge, according to Zachar, was modernising the legacy systems sitting underneath Zoey so they could operate in real-time rather than the batch processing – something common across the collections industry.
“The hardest part isn’t the AI, it’s the genetics underneath it,” said Zachar. “You can build a world‑class AI engine, but if it’s bolted onto 20‑year‑old batch systems, you have to change your own DNA before you see the real impact.”
Nutun plans to offer Zoey to its outsourced collections clients from July, and to international clients from October as an outbound collection tool running on its cloud platform.
“Most of our industry does one of two things: they buy books or they run call centres,” said Craucamp. “We build what doesn’t exist. We saw where this was going, and instead of waiting for it, we made it.”
Share
