Page 13 - EngineerIT June July Issue 2026

LARGE LANGUAGE MODELS

No infrastructure problem The process is not unlike teaching a child through repetition and
Conversations about AI often focus on correction. Models improve when they are exposed to more examples and
hardware, processing power and data given feedback when they misunderstand context or meaning. Hawkins
centres. Hawkins believes the bigger suggests that South Africa's long-term opportunity is in contributing
challenge is less obvious. more local data rather than attempting to solve the problem through
infrastructure alone.
"We don’t have an infrastructure
problem," he says. "It ultimately comes "We are steps behind America and England in terms of language because
down to data." they didn't have to go through that process," says Hawkins.

“Most large language models were After the transcription
trained primarily on American and For decades, organisations have recorded customer interactions, support
international datasets. As a result, they calls and sales conversations. Most of that information has remained
have had far more exposure to those largely untouched because reviewing thousands of hours of recordings
speech patterns than to local languages, manually is impractical. Transcription and AI aggregation tools change that.
accents, and conversational habits. They
just don’t have enough South African Once conversations become searchable text, businesses can begin
context yet.” analysing patterns across large volumes of interactions. Customer
sentiment can be tracked. Recurring complaints can be identified. Coaching
The lack of context becomes more opportunities can be highlighted, and service quality can be monitored at a
obvious when conversations move scale that would previously have required enormous human effort.
between languages, accents and cultural
references. For example, a South African Hawkins explains that reports can identify positive and negative
customer service call may be conducted interactions, allowing managers to review conversations and use them as
in English, Afrikaans, and isiZulu in the coaching opportunities for staff, adding, “The value lies not in the transcript
same discussion. itself, but in the insights hidden within it.”

Industry terminology may sit alongside Responsibility is still human
slang expressions and regional Hawkins cautions that AI is a tool, and not a substitute for accountability.
pronunciation. Human listeners generally "The user is still responsible for the output that the transcription supports,”
make sense of these transitions without he says.
giving them much thought. For AI, each
variation introduces another layer of He compares AI-generated transcription to the work produced by a junior
interpretation. employee. The tool can perform the task, but responsibility for accuracy
remains with the person reviewing the result. That becomes particularly
Hawkins points to a simple example. important in regulated industries such as healthcare, insurance and legal
Before a model can accurately interpret services, where a transcription error could alter the meaning or create
what is being said, it must first determine unintended consequences.
which language it is hearing. Even that
becomes challenging when speakers For businesses adopting AI, Hawkins believes the focus should be on
move between languages or use words augmentation rather than replacement.
that appear in multiple linguistic contexts.
Says Hawkins, “The goal is not to remove people from the process. It is to
Learning through exposure remove repetitive administrative work so skilled employees can spend more
For Hawkins, the solution is neither time solving problems, supporting customers and improving services.”
mysterious nor revolutionary. AI learns
from examples. Transcription technology, it seems, is on the cusp of greatness, but only
we can thrust greatness upon it, because we are responsible for teaching
“The more representative data a model this “toddler on a growth spurt” the facts. Through human interaction, the
receives, the better it becomes at technology will continue improving. For now, while understanding words
recognising patterns and interpreting is relatively straightforward, understanding South African conversation is
meaning,” he adds. something else entirely.

"We have to explain clearly to the AI what Euphoria Telecom is actively working on speeding up the process. Expect to
actually happened in the sentence." hear more South Africanisms from call centres and bots in the near future.

13 | EngineerIT June/July 2026