South Africa's Best BPO Argument Isn't Cost Anymore. It's What AI Can't Do.

May 28, 2026

There's a version of the AI customer service story that's been told confidently for the past two years. Deflect volume. Reduce headcount. Let the bot handle it. The data is starting to push back.

Gartner's 2024 research shows only 14% of customer service issues are fully resolved in self-service. A separate Gartner poll found that 51% of customers trust human agents most for issue resolution - and only 7% trust AI. By 2027, Gartner predicts that 50% of organisations that planned to significantly reduce their customer service headcount will have reversed those plans.

The AI deflection playbook isn't dead. But it's hitting its limits - and those limits are showing up exactly where you'd expect them to: regulated industries, complex transactions, emotionally loaded conversations. For South Africa's BPO sector, that's a more significant shift than most commentators have noticed.

The cost math is changing

One of the central arguments for AI-first customer service was cost. That argument is getting harder to make.
In January 2026, Gartner published a prediction that stopped a lot of procurement conversations: generative AI cost per resolution will exceed $3 by 2030, once infrastructure, governance, and human fallback costs are properly accounted for. That puts AI-only resolution in the same cost bracket as experienced offshore human agents in South Africa.

The reason costs are climbing isn't mysterious. Complex customer issues consume more compute, require more governance overhead, and generate more expensive error correction when the AI gets things wrong. The hardest cases - the ones that actually matter to customers - are also the most expensive to automate.

This doesn't mean AI has no role in contact centres. Hybrid models work well when the design is right. The evidence from well-run deployments shows meaningful improvements in first-contact resolution and compliance monitoring when AI handles routine volume and supports human agents with real-time knowledge. The argument is specifically about where humans sit in the model - and right now, that question is being revisited by buyers who moved too quickly toward full automation.

Where AI breaks - and what's at stake when it does

The failure modes are predictable. AI struggles with multi-intent queries, policy grey areas, and interactions where a customer's emotional state matters as much as the technical problem. It fails in novel scenarios it hasn't been trained on. It fails when the right answer requires judgment, not retrieval.

In February 2024, the British Columbia Civil Resolution Tribunal handed down a ruling that made these failure modes a legal issue. Air Canada's chatbot had told a grieving passenger that bereavement fares could be claimed retroactively - contradicting the airline's actual policy. The tribunal found Air Canada liable. The airline's argument that the chatbot was a separate legal entity responsible for its own statements was rejected.

The ruling sent a clear message to every business deploying AI for customer contact: you own what your bot says, regardless of whether a human wrote it.

In regulated sectors, that liability exposure is amplified. FCA enforcement action against UK financial firms exceeded £186 million in 2024/25, with individual fines exceeding £30 million, often tied to monitoring failures and inadequate escalation to human supervisors. The EU AI Act now classifies AI in banking and insurance as high-risk, requiring human oversight, transparency, and conformity assessments. South Africa's own National Financial Ombud has reported seeing AI-generated complaints with fabricated legal citations, inflating adjudication workloads in ways that weren't anticipated when firms rolled out automated customer-facing tools.

The pattern is consistent: the interactions that carry the most commercial and legal weight are exactly the ones AI handles least reliably.

What this looks like across key verticals

iGaming player support is an instructive case. Zendesk's 2026 iGaming CX data shows operators using AI to resolve 37–50% of chat volume and reduce first response times significantly. But the operators doing it well aren't running AI autonomously - they've embedded KYC, AML, and responsible gambling rules into the workflow, with mandatory escalation when those rules are triggered. A player flagged for problem gambling behaviour, an unverified withdrawal request, a dispute about in-play settlement: these go to trained humans. Not because the AI can't produce an answer, but because a wrong answer in these scenarios carries regulatory and licence risk that no operator can absorb.

Insurance claims processing follows the same logic. AI can assist agents with knowledge retrieval, summarisation, and compliance prompts during a claims call. It can flag interactions that need supervisor review. What it can't do is make the suitability judgments or regulatory disclosures that a claims adjuster is required to make. One 2024 insurance-sector discussion documented AI in a pure assist role, where FCR, average handle time, and leakage all improved - not because humans were replaced, but because they had better information in real time.

Accounts receivable management and debt collection add FCA and NCR compliance requirements on top of the standard complexity. Vulnerability screening, payment arrangement oversight, and complaint handling in collections all require human judgment that is difficult to train, harder to audit in an AI system, and legally non-negotiable in most jurisdictions.

The sectors where South Africa's BPO operators have built the deepest capability are the same sectors where AI automation reaches its limit first.

What good hybrid looks like

The evidence from well-run deployments isn't discouraging. One documented case from a multinational bank showed that deploying AI for routine queries - balance checks, simple card issues, standard account queries - while keeping complex financial advice with human agents produced a 94% reduction in wait times for common questions, a 37% drop in escalations to specialist teams, and a 23-point NPS improvement. 92% of agents reported higher job satisfaction when AI absorbed the transactional load and humans focused on conversations that actually required their skills.

The key wasn't making the AI smarter in isolation. It was designing the handoff so customers weren't asked to repeat themselves, agents had full context when they picked up, and escalation triggers were based on complexity and risk rather than a failure count.

That design work - the escalation architecture, the compliance wrapper, the QA layer - is where experienced human-led BPO operators have an advantage that pure-technology providers don't. They've been building it for decades.

South Africa's structural position

South Africa's BPO sector was never solely a cost play. The structural advantages - English proficiency ranked joint-first across the African continent with a score of 602 (EF English Proficiency Index, 2025), strong cultural alignment with UK and US clients, time zone overlap with European markets - compound most in the service scenarios AI handles worst.

When a UK financial services firm needs someone to handle a complex complaints call, conduct a vulnerability assessment on a debt collection account, or manage a disputed iGaming transaction, they need an agent who can read the room as well as follow the process. That's a training and culture question as much as a technology one.

"We've never treated human capability as a cost problem to be engineered away," says Kobus Nel, GCEO of Afrishore BPO, a Johannesburg-based business process outsourcing operator with more than 20 years in the market. "The clients coming to us now aren't just price-shopping. They're rebuilding after a period of over-automation. They want people with the training and cultural fit to handle hard conversations - and the compliance controls to do it in regulated industries. That's what South Africa built for."

The buyers making those decisions are mostly UK and US firms. They're the same markets South Africa's BPO sector has been positioning toward for the better part of two decades.

The opportunity

The AI deflection wave pulled a lot of offshore volume back toward automation. Some of that will stay automated - it should. But the complex, regulated, high-stakes interactions are flowing back toward human-led providers. And the operators positioned to capture that flow are the ones who invested in their people during the years when the narrative was running the other way.

BPESA's mandate has always been to position South Africa as a destination of choice for global buyers. The conditions for that argument are, right now, better than they've been in years.

The sector that kept investing in people is the one those buyers are looking for.

---------------------------------------------------------------------------------------------------------------------------

Anton Vukic is the Vice President Global Partnerships at Afrishore BPO, a South African business process outsourcing operator serving UK, US, and EU clients across customer support, financial services, iGaming, insurance, and accounts receivable management.