Will agents be replaced by bots? Few people today believe that AI is capable of largely taking over customer contact. Empathy and complex conversations? You need people for that. But for how much longer?

 

There are at least three reasons to keep thinking about whether AI will be able to conduct conversations with customers independently. Technological developments show that we have come a long way; there is a continuing drive for cost savings within contact centers; and the revenue model of CCaaS suppliers is ready for a new boost. In this analysis, we will delve deeper into these three factors.

The technology

First of all, whatever is technologically possible will be used – unless the potential is too limited or the countervailing forces are strong enough. Over the past twelve months, the world’s largest technology companies, including Amazon, Microsoft, Anthropic, Meta and Alphabet (Google), have collectively more than 200 billion dollars invested in new and low-threshold applications such as Retrieval-Augmented Generation (RAG) and Large Language Models (LLMs). Tens of billions of dollars have also been invested in a multitude of AI startups; almost twice as much as in 2023. Investment rounds are still accompanied by large amounts: in October 2024, OpenAI raised more than 6,6 billion dollars from Microsoft and Nvidia, among others.

LLMs develop through

In the past period, LLMs and RAG have improved considerably. By integrating LLMs with RAG, the chance of hallucinations and the use of outdated information has decreased. Also new methods such as Iter-RetGen ensure that LLMs work with more relevant information, which significantly improves the quality of the output. LLMs are also now able to process longer and more complex texts. Multimodal models have also become much better in the field of images, audio and video. Knowledge graphs are also playing an increasingly important role in organizing the information properly with which GenAI must work to achieve greater reliability, transparency and accuracy.

The above factors play a role mainly in the text-based side of customer contact: the dialogue. When it comes to performing actions or carrying out transactions, also an important part of the work of customer contact employees, RAG in combination with LLMs is not always sufficient. Other forms of AI are needed for that.

Copying the art of real agents?

Today, contact centers use AI systems that listen in and summarize call data or analyze sentiment, which are largely support tasks. Customer-facing AI is still limited to chatbots and the occasional voicebot. Many vendors believe that with continued investment and technological advancements, it is likely that AI systems will be able to handle more complex customer interactions autonomously in the next three to five years. CCaaS vendors have recently invested in R&D programs focused on AI agents or Agentic AI, where the autonomous execution of actions is the starting point.

The Agentic AI solutions can learn a lot from the huge amounts of customer conversations that CCaaS vendors can access via their customers. These contain information about the customer question and the solution offered, sentiment, empathy and often also context – especially when compared to data from knowledge bases. Supervised learning is very possible with these types of datasets, but only if the customer service is of a very good standard. Poor customer service is nothing more than a bad example for AI.

Obstacles and limitations of AI

In complex and emotional situations, AI cannot do it alone and human intervention is necessary, so the belief is. This involves both an empathetic response and an adequate substantive response that fits the situation. At Ziptone, we have previously argued that AI can come across as empathetic. In AI, that empathy is simulated; we may assume too easily that empathy in people is authentic and intrinsically driven. Agents can also 'act' on autopilot.

Until recently, ambiguity in human speech has been a tricky phenomenon for AI: think of a customer sarcastically remarking “I’m really happy with your service.” This limitation around sarcasm and humor was largely due to the rules-based models behind sentiment analysis: words are categorized as positive or negative, and machine learning is used to train algorithms to classify text.

Sarcasm and other complicated expressions

AI models are now increasingly able to recognize ambiguous information such as sarcasm and include it in the output. The basis remains regular sentiment analysis, but this is supplemented with information about 'sentiment shifts': subtle changes in the emotional charge of a text, where successive parts of text are compared with each other. This allows a model to identify contradictions: 'what a great service' versus 'I've been on hold for an hour'. In doing so, a model uses specifically trained applications to pay attention to specific stylistic figures such as hyperbole. These AI models can also be trained on existing, marked examples of sarcasm.

These kinds of 'abilities' can be trained via supervised learning and reinforcement learning based on human feedback (RLHF). If these conditions are met, AI can learn iteratively from failures and thus become increasingly better at handling complex situations. There will still be situations that AI cannot handle, for example because they have not been sufficiently prevented.

Ethical side of smarter AI

As the field of AI moves more towards complex, sensitive and emotional topics, the ethical discussion also comes more to the fore. AI Act stipulates that it must always be clear that the customer is 'talking' to a robot. As robots increasingly display empathic behavior while it is known that it is a robot, the chance of a negative review increases, according to research, including from Tilburg university. In other words: as soon as the interaction becomes so 'human' that it no longer feels completely like technology, users can feel uncomfortable or experience it as fake and manipulative. This phenomenon is often related to the so-called "uncanny valley". Consumers tend to trust a robot more if they know that they are really dealing with a robot. In the coming period, companies will have to learn to balance cost savings and customer trust even better.

'New questions' remain complicated for chatbots

It is true that AI has difficulty interpreting situations that deviate from what it has seen before. New or unusual questions or emotions require improvisation and creativity that AI is not (yet) fully capable of. AI is also not yet able to adapt in real time to unique or new circumstances in a conversation. Think of a follow-up question that is not standard, sudden turns in a conversation or a change in emotion. For now, humans are better than AI at dealing with these kinds of, sometimes subtle, variations (such as hesitations in the voice, a silence, or unexpected changes in tone) and come up with creative or intuitive solutions for unique situations. Human agents, thanks to their empathy and moral awareness, can make decisions that go beyond protocols, such as going the extra mile, making exceptions or finding workarounds. People are also expected to be better at understanding cultural nuances. But they can also make immoral decisions and discrimination is not only an activity of algorithms.

Meanwhile, language and speech models are being further developed. A new program was recently started in the Netherlands: HOSaN, or High-quality Speech Models for all Dutch people, the speech variant of GPT-NL and relevant for the customer contact sector. This model focuses on better understanding the voices of the elderly and children, people who use a dialect or second-country nationals (EU citizens with a nationality other than Dutch). HOSaN can contribute to more inclusive voicebots.

AI and cost savings in contact centers

Cost reduction has contributed to the rise of IVR systems, self-service environments, FAQs, web forms and chatbots. AI in contact centers promises to be the next cost killer. The use of automated summarization is often presented as a solution that can make the work of agents more pleasant. But do contact center managers measure the reduction in AHT or the increased employee satisfaction? The same applies to automated quality monitoring: in addition to 'reviewing all conversations', AQM also saves time (or FTEs of coaches). Do employees also receive more personal attention from coaches after the introduction of AQM?

With new technological capabilities such as real-time translation, outsourcing and offshoring in a different light. With AI that first listens, then learns and then independently conducts conversations, companies have the opportunity to have customer conversations handled automatically.

In real-time speech translation, latency will continue to play a role, not only due to technical factors, but also because good translation often requires the client to finish speaking. However, a shrinking labor market with increasing competition and rising labor costs will further increase the pressure to develop and implement technological solutions.

Implementation costs

However, implementing AI-based solutions also comes at a cost. In many cases, it will be necessary to put your data management in order and infrastructureThere are also developments that could have a dampening effect on infrastructure costs in the long term. For example, the rise of optical connections in data center hardware (instead of copper-based chip-to-chip connections) will accelerate training of AI models and reduce infrastructure costs, IBM expects.

Then come the costs of implementing and integrating AI systems, for which license fees must also be paid. The usage costs of solutions such as AI agents are currently not always easy to predict. The underlying data traffic can also become a significant cost item. This makes drawing up a positive business case complicated, apart from the possible negative impact of AI on the customer experience. In addition, people are needed to take care of the ongoing monitoring and management of AI systems.

Some of that control can, however, end up outside your own organization: for example, with (multiple) specialized AI suppliers. This can lead to organizations not building their own 'intelligent layer', with which the technical debt grows unnoticed. It can also lead to fragmentation in the cost structure, with the disadvantage of less overview.

The existing revenue model of CCaaS suppliers is being challenged

Third, the existing business model of CCaaS vendors is a driver for the further development of autonomous AI systems for customer service. CCaaS players currently earn their money based on the maximum number of employees logged in simultaneously – a variation on the license model – sometimes with a multiplier of the number of hours.

A decline in the number of human agents poses a threat to the revenue model of CCaaS players. Although they are currently partly cannibalizing their own market with their additional AI agents and other AI services, AI agents will mainly become a new revenue model in the medium term. Currently, most AI agent pricing structures are based on the number of AI interactions, on the complexity of tasks or on a certain outcome.

What will bring about a change in the future?

AI has already made significant technological progress in customer contact, but fully autonomous customer conversations remain a challenge due to complexity, ethical concerns and the need for human empathy in emotional situations. The question is, however, what is needed in terms of cost pressure and technological development to ensure a change. It will not be up to the technology suppliers. (Ziptone/Erik Bouwer)

 

Thanks to Carla Verwijmeren, partner and CCO of Y.digital

Featured, Technology