What is generative AI?
Generative AI is AI that produces something new. Earlier AI in contact centres mostly classified or predicted: is this call about billing or a fault, how many calls will we get on Monday? Generative AI can write a reply, summarise a conversation, answer a question in natural language or draft a report.
The text side of generative AI is powered by large language models (LLMs). An LLM is trained on enormous amounts of written text and learns the patterns of language well enough to generate fluent, relevant responses to almost any prompt. The same family of models increasingly handles speech as well, which is why AI voice and chat have improved so quickly.
How large language models work, in plain terms
An LLM works by predicting the next word (strictly, the next piece of a word) given everything before it, over and over. Because it has learned from so much language, those predictions capture grammar, facts, reasoning patterns and tone. Given instructions and context, such as “summarise this call for a CRM note” plus the transcript, it produces a response that fits.
Two points follow from this:
- It is flexible. The same model can summarise, classify, extract details and write replies, just by changing the instructions.
- It is not a database. It generates likely text, not looked-up facts. Unless it is given the right information, it can produce something plausible but wrong.
Where generative AI is used in contact centres
- Conversational AI, including voice assistants and chatbots that understand free-form requests and respond naturally.
- Agent assist: suggesting knowledge, improving written replies and capturing details.
- AI call summarisation and wrap-up notes.
- Understanding and routing: modern natural language understanding is often LLM-based.
- Analytics: finding themes and answering questions across thousands of conversations in speech analytics.
- Agentic AI, where models plan and carry out multi-step tasks.
Limits and risks
Generative AI is powerful, but it needs to be used with care, particularly with real customers.
- Hallucination. Fluent but incorrect answers, such as a made-up policy or a wrong restoration time. This is the main reason AI guardrails matter.
- Inconsistency. The same question may get slightly different answers, which matters for regulated information.
- Data privacy. Customer conversations contain personal, financial and sometimes health information. Where it is processed, who can see it and whether it is used for training are critical questions.
- Over-reliance. It is tempting to automate conversations that need human judgement and empathy.
Questions to ask a provider
- Is the model a private deployment, or are conversations sent to a public AI service?
- In which region is data processed and stored?
- Is our data isolated from other clients’ data?
- Is our data used to train or improve any model?
- How are answers grounded in our approved information, and how are errors caught?
- When does a person take over?