Abstract

Artificial Intelligence is rapidly changing what organisations can learn from customer conversations. The September 2026 edition of Coffee with Callbi provided an interesting glimpse of where this evolution is heading, with the launch of AI Enhanced Callbi and a broader discussion about the transition from speech analytics towards Interaction Intelligence.

But beneath the excitement surrounding AI was a rather more important message. Technology does not remove the need for human expertise. In many respects, it makes that expertise even more important.

AI can summarise conversations, interpret context, identify call drivers and outcomes, detect sentiment and interrogate interactions in ways that traditional keyword-based analytics cannot easily achieve. Yet transcription quality remains fundamental, AI interpretation must be validated, prompts must be carefully constructed and humans must ultimately determine what the resulting intelligence means for the business.

Perhaps the most significant development is therefore not simply what AI can do. It is how the people who once listened to and scored individual calls are increasingly becoming Insights Analysts, interpreting interaction data and carrying customer intelligence far beyond the contact centre and, increasingly, into the boardroom.

Speech analytics is becoming something much bigger

For many years, speech analytics was fundamentally about finding things in conversations.

Did the agent say the required words? Was a regulatory statement made? Did the customer mention cancellation? Was a prohibited phrase used? Was the prescribed script followed?

These remain enormously valuable applications, particularly in regulated environments where organisations need precision and repeatability.

What is changing is the ability to move beyond searching for specific words and phrases and begin asking questions about the meaning of the conversation.

The newly introduced AI capabilities within Callbi illustrate this transition. For eligible calls, the platform can generate a summary together with additional data points including the call driver, call outcome and enhanced sentiment information showing not simply whether sentiment was positive or negative, but how it developed through the interaction.

That distinction is important.

A traditional analytics query might determine whether a particular phrase appeared in an interaction. An AI query can potentially address questions such as whether the agent adequately explained a product, whether the customer appeared hesitant, whether the agent was misleading or whether the customer is likely to pay a monthly premium.

We are moving from What words were used? towards What actually happened?

And that is a substantial shift.

AI and text analytics are not competitors

One of the most useful discussions during the webinar concerned where traditional text-based queries remain preferable and where AI adds something fundamentally different.

Whenever a new technology appears, there is a temptation to assume it must replace whatever preceded it. That would be a mistake here.

Text queries remain exceptionally powerful where precision matters. If an organisation needs to establish whether prescribed words were used, whether a compliance statement appeared or whether particular terminology was present, a carefully constructed text query provides considerable control.

Words can be added or excluded. Synonyms can be incorporated. Proximity can be defined. The organisation can establish very precisely what constitutes a positive or negative result.

AI is strongest elsewhere.

Its advantage lies in interpretation and context.

Consider something as apparently straightforward as identifying rude behaviour. Searching for swear words or obviously aggressive phrases will find some examples. But rudeness does not necessarily require profanity. The meaning of an apparently innocuous sentence can change completely according to the surrounding conversation.

Nosihle Mbatha used this example precisely during the webinar, explaining that AI can examine the wider context rather than relying solely upon the presence of particular words.

The sensible future is therefore not AI or traditional analytics.

It is AI and traditional analytics, each applied where its particular characteristics are strongest.

There is no AI magic wand

This was perhaps the most important caution to emerge from the discussion.

It is very easy to present AI as something approaching magic. Ask the question, press the button and receive the answer.

Real-world deployment is considerably more complicated.

Large Language Models interpret language. Different models can interpret the same material differently and even repeated interrogation of the same information can produce variations. For compliance applications, where an organisation may require an extremely precise and repeatable answer, that matters enormously.

Prompt construction consequently becomes a genuine business discipline.

Henriette Potgieter described testing in which changing a single word in an AI prompt changed the outcome produced by the model. Her analogy was particularly useful. An organisation would never recruit a new QA or Business Intelligence analyst, point them towards thousands of calls and simply say, “Go and find something interesting.”

People are trained. They are taught what the organisation considers important, what to listen for and how findings should be interpreted.

AI requires much the same discipline.

Queries need to be constructed, tested against known examples and validated. The results need to be compared with human assessment. Only when the organisation is satisfied that the query is producing sufficiently reliable results should it be allowed to operate at scale.

Corey Springett described pilot queries beginning at around 40% accuracy and, through testing and refinement, reaching 85% to 90%. His point was not the percentages themselves. It was that there is no silver bullet. Reliable analytics requires work.

That principle applied to speech analytics before generative AI arrived.

AI hasn’t made it disappear.

Garbage in. Garbage out.

All the intelligence generated downstream ultimately depends upon something rather less glamorous: the transcription.

If the speech recognition engine incorrectly converts what was said into text, everything that subsequently analyses that text works with corrupted information.

As Corey put it during the webinar, an organisation might deploy one of the best AI engines in the world, but if the transcription is inaccurate, the AI is still interpreting noise.

This issue is particularly significant in South African contact centres.

A language model trained predominantly on broadcast-quality English faces a very different environment when exposed to actual contact centre conversations. Broadcast speakers generally have good microphones, relatively controlled acoustic environments and clear diction.

Contact centres have background noise, variable microphones, imperfect connections, different accents, mispronunciations, industry terminology, product names and conversations that don’t politely wait for one person to finish speaking before the other begins.

South Africa adds another dimension: language diversity and code-switching.

A conversation may begin in English, switch into isiZulu or Sesotho and move backwards and forwards between languages. For organisations operating at scale, this has historically presented a considerable challenge to QA and analytics teams.

The AI capability discussed during Coffee with Callbi introduces an interesting development. A query can be asked in English against supported vernacular conversations and the resulting summary, call driver, outcome and other intelligence can be returned in English.

That potentially removes an important barrier to extracting intelligence from multilingual customer conversations.

From samples to big data

Traditional contact centre quality management has always suffered from a basic mathematical problem.

There are simply too many interactions and too few people available to listen to them.

Consequently, QA departments historically assessed a tiny percentage of the total interaction population. Valuable work could certainly be done with that sample, but it remained precisely that: a sample.

Interaction analytics changes the equation.

When summaries, call drivers, outcomes and other data points can be generated across every eligible interaction, organisations begin accumulating something much closer to a genuine dataset of customer behaviour.

That opens up possibilities far beyond conventional QA.

Why are customers actually calling?

Which contact types have unusually long handling times?

Where is silence occurring?

Are the disposition codes selected by agents consistent with what actually happened during the conversation?

Which issues are driving repeat contacts?

Where are processes creating unnecessary customer effort?

The value lies not simply in identifying an individual problematic call, but in aggregating thousands of interactions and discovering patterns.

The next stage discussed during the webinar is the ability to interrogate those aggregated interaction datasets using AI. Callbi currently allows the underlying information to be extracted and analysed externally, while bringing large-scale AI interrogation of multiple calls directly into the platform is on the development roadmap.

This is where the phrase Interaction Intelligence starts to become particularly meaningful.

Human in the loop

Throughout the webinar one phrase kept resurfacing: human in the loop.

It deserves to.

AI can identify patterns humans could never practically discover by manually listening to hundreds of thousands of conversations. But AI does not understand an organisation’s strategy, customers, risk appetite and operating environment in the same way experienced people do.

Humans therefore remain involved at several critical points.

Someone must decide what questions are worth asking. Someone must construct or refine the queries. Someone must validate whether the results are sufficiently reliable. And someone must interpret what those findings mean for the organisation.

This becomes even more important when interaction intelligence starts influencing significant business decisions.

A dashboard showing that a particular customer complaint is increasing is interesting.

Understanding why it is increasing, whether the cause lies in the contact centre, product design, pricing, fulfilment, digital self-service, marketing promises or another business process requires analysis and judgement.

AI accelerates discovery.

Humans provide meaning.

The QA is becoming an Insights Analyst

Perhaps the most encouraging discussion of the session concerned the changing role of the traditional Quality Assessor.

When speech analytics first entered contact centres, some QA professionals understandably viewed it as a threat. If technology could automatically examine thousands of interactions, what happened to the person whose job was to listen to and score calls?

What we are increasingly seeing is evolution rather than simple replacement.

Henriette described how the role increasingly requires critical thinking. Instead of merely assessing a call and recording a score, the QA professional must now ask whether the data makes sense, whether it is reliable, what it reveals and how it should be communicated.

The title appearing increasingly frequently is Insights Analyst.

That is much more than cosmetic job-title inflation.

The value of the role moves upwards.

Instead of telling an Operations Manager that an agent achieved 82% on a quality scorecard, the Insights Analyst might identify that thousands of customers are contacting the organisation because of a particular process failure.

That intelligence might influence training, operations, digital design, product development, marketing or corporate strategy.

During the webinar I made the observation that we are beginning to see these insights reaching the boardroom, rather than remaining an operational contact centre metric.

That may ultimately prove to be one of the most important consequences of Interaction Intelligence.

Responsible AI requires responsible data management

As AI becomes embedded within interaction analytics, questions about privacy, security and data residency inevitably become more important.

The webinar addressed these issues directly.

Callbi retains call recordings for a rolling three-month period at its standard tier. Corey also outlined security measures including encryption, two-factor authentication and ISO 27001 certification. He explained that customers can have particular data-storage requirements, including South African hosting where regulatory requirements demand that information remain within the country.

AI introduces an additional infrastructure consideration because large-scale processing requires significant GPU capacity. In some configurations data may therefore need to be processed temporarily within EU infrastructure while remaining stored within the customer’s designated region. The webinar stressed that this is governed through specific data-processing arrangements.

This is an important reminder for the wider industry.

Responsible AI is not simply about whether the model produces a sensible answer.

It encompasses how information is collected, transmitted, processed, stored, secured and ultimately used.

The destination is intelligence, not technology

Towards the end of Coffee with Callbi, Corey made an observation that neatly captured the direction of travel.

The first release of an AI capability will not necessarily do everything every customer eventually wants. User feedback, experimentation and practical deployment will continue shaping development. Callbi’s roadmap includes further AI capabilities, enhancements to traditional text queries and the ability to derive insights across larger populations of interactions.

That is how this technology should evolve.

The objective should never be to deploy AI simply because AI is fashionable.

The objective is to understand customers, agents, processes and business performance better than we could before.

Speech analytics gave us the ability to search enormous volumes of conversation.

Interaction analytics allowed us to quantify what was happening within those conversations.

AI now gives us the ability to ask increasingly sophisticated questions about why things are happening and what they might mean.

But none of this eliminates the human contribution.

Quite the opposite.

As machines become better at processing interactions, people need to become better at asking questions, testing assumptions, challenging outputs, interpreting patterns and converting information into action.

That is the real transition taking place.

We are not simply adding AI to speech analytics.

We are moving towards Interaction Intelligence, where technology can examine virtually every customer conversation, but experienced humans remain responsible for deciding what the intelligence means and what the organisation should do about it.

And that, I suspect, is where the real value will be found.