15 million chats per year: Lessons in AI-driven customer conversations from Lloyds Banking Group with Cainan Wright and John Young
VUX WorldShow Notes
Once switched off during busy periods, Lloyds Banking Group’s chatbot now handles more than 15 million conversations a year. Learn what it really takes to build AI solutions that customers trust with their money.
We sit down with Cainan Wright, Senior Product Owner, and John Young, AI Lead - Chatbots and Generative AI, from Lloyds Banking Group, to explore their remarkable nine-year journey in building one of the UK's most successful banking chatbots. From handling 3 million conversations in 2021 to over 15 million in 2025, they share the hard-earned lessons of transforming a rather "unloved" virtual assistant into a mission-critical service for almost 27 million customers.
We cover how the team stabilised their service by focusing on the basics, investing in design expertise, and securing organisational buy-in through demonstrating clear value.
Cainan and John also share valuable insights into their pragmatic approach to generative AI, explaining how they combine deterministic processes with generative language capabilities to deliver the best customer experience while maintaining brand consistency and ensuring regulatory compliance.
The team's journey offers important lessons about perseverance and the importance of striking a balance between innovation and responsibility when deploying AI in a highly regulated industry.
Timestamps 00:00 – Intro 01:34 – Meet Lloyds Banking Group 03:18 – The unloved virtual assistant era 05:30 – AI before ChatGPT 06:50 – Stabilising the service and earning trust 14:23 – From underloved to mission critical 16:55 – Buy-in from stakeholders to scale AI capabilities 18:41 – Defining new AI roles in banking 20:51 – Scaling from 3 to 15 million conversations 25:15 – How to handle a 5x increase in three years 27:43 – Experiments with generative AI in banking 30:24 – Using an LLM classifier 31:07 – Improving tools within teams 32:45 – Using generative AI for value and impact 34:56 – Valuable design research sessions 38:56 – Meeting customer expectations with generative AI 46:04 – Maintaining consistency in probabilistic landscape 49:30 – The right tool for the right job in conversational AI 55:26 – How do you measure the impact of generative AI? 56:13 – Responsible AI and proving business value 1:01:10 – The next 12 months 1:07:02 – Magic wand moments 1:11:34 – If the team could start over again 1:14:03 – Closing remarks from Kane
Show notes
Follow Kane Simms on LinkedIn Discover VUX Consulting Take our updated AI Maturity Assessment Subscribe to VUX World Subscribe to The AI Ultimatum Substack Hosted on Acast. See acast.com/privacy for more information.
More from VUX World
Automation Anywhere has agreed to acquire boost.ai. What could that mean for customer experience, personal AI agents and the autonomous enterprise?In this exclusive interview, Kane Simms speaks with Jerry Haywood, CEO of boost.ai, and Ankur Kothari, co-founder and COO of Automation Anywhere, about the proposed combination. They discuss how conversational and voice AI could connect to work across enterprise systems, where people remain involved and what customers need to prepare for.The conversation covers the rationale behind the deal, customer and employee operations, the changing relationship between CX and IT, personal agents, agent-to-agent interactions and the foundations of successful AI adoption. Kane closes with his view on why the combination matters.Acquisition status: the companies announced a definitive agreement on 7 October 2026. Completion is expected in Q4 2026, subject to regulatory approvals and customary closing conditions. Show notes Follow Kane on LinkedIn: https://linkedin.com/in/kanesimms Find out more about VUX: https://vux.ai Subscribe to The AI Ultimatum Substack: https://open.substack.com/pub/kanesimms
Reliable voice AI comes from rigorous testing and evals, and Rohan Vasishth from Bluejay sees trust as the fundamental bottleneck holding back enterprise deployment. Rohan is Co-founder and CEO of Bluejay, which tests and monitors voice and chat AI agents for companies including Google and DoorDash. We work through what to test, when to test it and what each approach costs. We discuss why testing voice AI introduces a whole new layer of complexity, from accents and background noise to interruptions and speech recognition errors. Rohan also shares how customers can stress-test agents, why starting with text-based evaluations can save time and money, and how businesses can identify and reduce the latency customers actually experience. Rohan closes by exploring the widening gap between rapidly advancing AI capabilities and slower enterprise deployment, and why evaluation is essential to building the trust needed to close it.
Stacey Kyler and Guy Tonye of Zammo return to the show for the first time since 2021. Back then they were building for Google Assistant, Alexa and Cortana. ChatGPT arrived soon after and reshaped what customers expected. They cover the move from NLU and diagram flows to agentic systems, why customers now want both the freedom of Gen AI and firm control, how Zammo enforces determinism at the code level while Gen AI handles the conversation, and why evals are the thing Guy would add to any conversation design course today.
The contact centre holds a record of almost every customer interaction. Every conversation carries a signal about why customers are calling and what is behind it. Processing those conversations at scale took the power the contact centre never had before. Ping Wu, CEO of Cresta, joins us to work through what that shift means for enterprise contact centres, from cutting the calls that should never happen to automating the ones that can be. We get into how AI changed what a contact centre can see. Every call, every screen, every exchange can now be processed rather than sampled. Human effort never scaled the way silicon does, so for years most of that record went unread. Large language models bring processing power and language understanding to work through millions of conversations, see what is driving them and turn that into insight, coaching and automation.
Voice AI is the channel everyone is talking about in 2026. One of the biggest things slowing it down is often already sitting inside your stack. In this episode, I’m joined by Endre Davids from Seamly to explore what really happens when businesses take an existing chatbot and move it into the voice channel. We discuss why latency often starts deeper in the conversational platform than teams expect. Endre explains how companies can spend time optimising speech recognition and text-to-speech while the chatbot itself is responsible for much bigger delays. The conversation gets into why voice needs a different interaction model, because people speak differently from how they type. That puts far more pressure on getting the conversation design right. We also cover how to manage perceived latency, when it makes sense to reuse existing chatbot infrastructure and how sentiment, multimodality and customer context could shape richer voice experiences. Endre and I keep landing on the same point. The success of voice AI depends heavily on conversation design, and voice punishes the shortcuts that chat lets you get away with.