Discover the bottlenecks faced by South African organisations in deploying advanced AI voice agents and how these challenges impact customer experience.
Image: AI / Sora
When a consumer calls a service line, they do not care about the size of an organisation’s Large Language Model (LLM) or the neural networks running behind it. They want their query resolved.
Many South African organisations deploying generative AI are encountering a bottleneck in this regard.
They have built or deployed AI voice agents that are advanced and sound remarkably natural. Yet, when these polite bots are asked to execute a transaction – such as verifying a payment, updating account details, or checking an order status – they hit a digital speedbump.
Building a bot that can hold a superficial conversation has become remarkably straightforward. Integrating that bot into backend enterprise databases to solve real-world customer problems remains the true frontier of corporate innovation.
This is somewhat of a new operational hazard for customer experience (CX) teams. Historically, the primary customer complaint was waiting in long queues. Today, organisations risk replacing those queues with automated agents that politely waste a caller’s time.
A voice bot without deep backend integration is simply a high-tech roadblock. If an automated assistant cannot securely read from and write to a central Customer Relationship Management (CRM) platform, its utility is severely limited.
It is forced to guess, hallucinate, or unnecessarily hand the call over to a human employee way too often, kind of defeating the purpose of automation.
In other words, the AI agent should not be guessing.
It should be retrieving the right information from the right place, and using that to answer accurately because a customer needs the correct answer over something that sounds like an approximation of the truth.
Every call centre agent knows the phrase: “Just a moment, my system is slow today.”
Customers know it too with first-hand experience of waiting on the phone while an agent tries to find a record or check in with another department for relevant information.
Given the possibility of replacing one kind of frustration with another, businesses should be asking what customer journey they want to improve and whether there are the systems and data to support it. That foundation takes longer to build, but it is what separates a novelty from a real CX asset with measurable ROI.
To prevent automated agents from hallucinating, organisations are turning to Retrieval-Augmented Generation (RAG).
RAG serves as a secure, real-time index for the AI. Instead of allowing the model to rely on its general public training, RAG forces the AI to ground its responses strictly in the organisation's verified datasets.
When a customer queries a specific contract clause, product parameter, or pricing tier, the AI queries the internal database, retrieves the exact document, and translates that data into speech.
However, this architecture changes when applied to transactional environments.
Instead of using RAG to just answer general FAQs, it now also becomes about verifying real-time, personalised account parameters before executing an action. If the underlying data is siloed across disconnected departments, the AI inherits those structural flaws.
Managing risk in these automated journeys requires strict operational guardrails. An AI voice agent should never operate with absolute database autonomy.
At 1Stream, we address this through a system of continuous confidence scoring, recommending a strict threshold of 80%. In a transactional context, this score serves as a data validation gateway rather than just a trigger for human intervention. If the voice bot’s understanding of a customer’s instruction (such as a change of address or a payment arrangement) falls below 80% statistical certainty, the system freezes database write permissions.
This safeguard prevents corrupted, misspelled, or misaligned data from being written directly into the CRM. Instead of executing an unverified transaction, the system routes the call to a human specialist, passing along the transcribed context so the customer does not have to repeat their story. It’s a triage model and it protects the organisation from database corruption and regulatory errors, while freeing up human employees to handle complex, high-value conversations.
But businesses also need to think about where their AI processing and customer data sit.
Today, many AI services rely on international infrastructure because that is where the large models and processing power are available.
In many cases, that is unavoidable, but over time this is likely to change as local capacity will be needed.
There are several reasons for this. Speed is one. Data sovereignty is another. Security is a third. For many organisations, particularly those handling sensitive customer information, it matters where data is processed, stored and trained.
We believe strongly in building more local capability for AI voice agents.
That includes investment in local GPU capacity and local data centre capability, because the future of customer-facing AI cannot rely only on overseas infrastructure. South African businesses need AI solutions that are fast, secure, cost-effective and built for the environment they operate in.
Every speech-to-text, LLM and text-to-speech interaction carries a cost. If an AI agent rambles, goes off topic or takes too long to solve a problem, it wastes money and frustrates customers, and strong integration and clear guardrails are necessary to help control that cost.
Starting with the value AI can bring to CX will win over deploying it “because it’s fashionable”. Above all AI voice should genuinely make the customer journey easier and more consistent.
Bruce von Maltitz, CEO of 1Stream.
Bruce von Maltitz, CEO of 1Stream.
Image: Supplied.
Follow Business Report on Facebook, X and on LinkedIn for the latest Business and tech news.