The GenAI challenge.

Agentic AI, powered by Generative AI (GenAI), has incredible potential to deliver improved outcomes in your contact centres. In fact, contact centres would seem to have the best use cases for AI compared to almost any other business function. Yet statistics repeatedly show that anywhere between 75% and 90% of GenAI proof of concepts (PoCs) fail to make it into production. Why is that?

By Martin Cross, CEO of AI-enabled customer experience specialist, Connect

The voice quality challenge.

GenAI has often been trained on high-quality, clean wide-band audio, such as the audio on YouTube videos or podcasts. Yet in a contact centre environment the audio consists of narrowband telephony calls, often made over mobile phones and affected by poor signal quality and background noise.

GenAI simply struggles to fully understand what is being said.

The contact centre environment has its own peculiar quirks which AI was not trained on or is not well-suited to. This is particularly exacerbated in the voice channel. What are these challenges?

The regional accent challenge.

Automated Speech Recognition (ASR), the AI models that turn human speech into text, have often been predominantly trained on "standard" American English podcasts, YouTube videos, etc. Yet in a real-world contact centre environment particularly outside of the US, those accents are uncommon.

The result is that generic AI often struggles to understand what is being said in UK regional accents such as Glaswegian, Geordie and Scouse or in South African English.

The industry- and business-specific terminology challenge.

Specific industries such as insurance, healthcare and financial services have industry-specific terminology. Your individual business will have its own terminology. Yet GenAI is trained on general information rather than the specifics of your business or industry and makes probabilistic decisions based on the data examples it has seen.

So GenAI often may misinterpret a word given a context, for example:

• ”Pool” in the financial services industry is quite likely to refer to pooling disparate debts into a single pool but probabilistically, a swimming pool, can feel more likely to GenAI.
• ”Centurion” in your business may be a VIP charge card but probabilistically, a fourth-century BCE Roman soldier feels more likely to GenAI.

Techniques like Retrieval-Augmented Generation (RAG) are widely used to inject business-specific content. However, these tools merely provide external reference guides; they do not fundamentally alter how an LLM generates an output based on the probability at its core. Relying solely on RAG is like giving a general historian a financial playbook - they can read the rules, but their core intuition remains unchanged.

Consequently, overcoming a general model's baseline biases requires heavy prompting, which increases token costs and still leaves the door open for subtle, context-defying hallucinations. To achieve true native fluency and operational reliability, organisations must move beyond just guiding a general model.

The Q&A challenge.

GenAI is optimised to understand conversations, identify intent and extract entities. However, significant parts of a contact centre journey involve structured question-and-answer exchanges rather than conversational engagement. For example: What is your name? What is your car registration number? What is your postcode? What is your ID number?

GenAI simply struggles to understand these Q&A sessions, particularly: alphanumeric answers like "EC3M 4BY", spoken dates like "next Thursday at midday" and spoken numbers like "oh treble three, one, double two, double two".

This can be really limiting in a contact centre because these types of Q&A engagements are critically important, specifically for security purposes, i.e. identification & verification (ID&V).

ContactBabel's research shows that 74% of calls into UK contact centre require ID&V and if this has to be performed by live advisers because AI cannot do it, then the proportion of calls that can be handled agentically is vastly reduced and the return on investment lost. ContactBabel's research shows that currently 91% of those calls requiring ID&V are handled by a human advisor.

The security challenge.

If you are expecting GenAI to act agentically, then you are exposing the AI to all sorts of sensitive customer data and critical business systems. Yet through GenAI you are potentially exposing customer data and proprietary data to the internet.

The cost challenge.

GenAI is already expensive, and is destined to increase. According to a recent Gartner paper, GenAI's cost per resolution will exceed offshore adviser costs by 2030.

Contact centres are blessed with a repository of valuable training data. By training on your contact centre's interaction data (voice recordings, your email and chat transcripts), AI will understand lower quality voice calls with background noise, will understand regional accents, will understand industry-specific terminology and your company-specific terminology.

This will create speech recognition models that are proficient in understanding postcodes, car registration numbers and national ID numbers. It will also create a Small Language Model optimised for your business, making it faster and more accurate. Because the model is smaller, it can be run econom ically within your own environment and behind your firewalls, addressing GenAI security concerns and supporting sovereign AI.

An orchestrated ensemble dynamically coordinates multiple AI models, routing each interaction to the model best suited for the task - whether a financial reasoning model, a domain-specific expert model, or GenAI for conversational engagement. And if you or your business is wedded to particular GenAI toolset, such as Google Gemini, Anthropic Claude or OpenAI ChatGPT, the orchestrator can use this as of the ensemble, too.

Very often the use cases that are tested in the contact centre are not a real-world scenario or an edge case, which is another factor why such a large percentage of PoCs don't make it into production.

Data-Driven Insights is the answer. By analysing your contact histories you can identify contact volumes, interaction durations and the use cases that account for the greatest share of demand. Because you know your adviser costs and handling times, you can also calculate the likely return on investment and prioritise the use cases with the strongest commercial potential.