Vodafone to boost TOBi with Gen AI

"Mr. Watson, come here -- I want to see you..." Vodafone is very happy with its new Gen AI-enhanced chatbot, which may not be great news for IBM's Watson. The company says its success is a direct result of Gen AI principles it has laid out as it adapts to a world that might have far-reaching effects for the network.

Vodafone’s customer assistant robot TOBi is set to get a significant injection of intelligence from Gen AI, following work the operator has done in Italy to improve both service quality and customer interactions using new Gen AI technology.

TOBi can be a frustrating experience for consumers, running script templates that have been perhaps too obviously designed to keep users as far away from a human agent as possible. Indeed, Vodafone insiders admit that the key early KPI for TOBi was the number of incoming calls it could offset. For customers with anything beyond a basic query, interacting with the robot agent therefore became a game of bypassing the replies that would send you quickly to prepared website content and FAQs. Under more recent leadership, that KPI has changed to one more focus on customer satisfaction and resolution.

But in Italy, things are going even further as Vodafone’s data team have been putting LLMs to work on vast amounts of customer interactions, call recordings and post-call surveys to make things better.

Ignacio Garcia, Group Director of AI, Vodafone said that the company has multiple million customers in Italy already being served by the application, and claims that has led to a significant improvement in their experience of interacting with Vodafone.

Garcia said that Vodafone started with solving a problem: that it relied on post-interaction analysis of interactions and post-interaction surveys to assess customer satisfaction. These would be correlated with the conversations, analysed by the marketing department and perhaps a month later some action would be taken to solve a campaign or network issue that was raising its head a month previously.

Instead, it now runs all its anonymised customer conversations and touchpoints through its LLMs, running a generic model summarisation.

“And that was a very big thing. Because we started to see everything the customer was telling us. Every single conversation in summary. So we knew when we were having billing issues, or a problem in the network in a specific region. And that was like magic.”

Garcia says Vodafone has been able to do this because it has the data in the right place due to its data ocean strategy that it developed in partnership with Google Cloud.

This enhanced chatbot has been live for a year and Garcia says the value has been “amazing.”

“With millions of customers live we are getting 50% more satisfaction every time they interact with a chatbot – and 20% higher NPS in comparison with the previous AI. And that’s because of the comprehension of the model. With the new brain customers can ask questions and receive answers in natural language. We are changing from having to maintain a lot of templates to letting the LLM do that, and focussing on the experience of the customer.”

In addition, the human care agents are also now being supported by an advanced AI agent – a super agent – giving them additional capabilities and access to key information and decisions for when a customer requests a callback or when a situation needs to be escalated.

The operator now plans to extend this capability to wider markets.

Scott Petty, CTO of Vodafone Group, said it would do so cautiously, not least because TOBi also supports voice interaction, which requires very timely responses. He said that it could take 6-12 months to scale up the GenAI version of TIBi.

“You could do that in six weeks or three months but you run a bigger risk that you don’t deal with the edge cases and you create negative customer experiences that you haven’t thought through.”

He added that, in time, the operator would gradually move away from using IBM’s Watson and Microsoft’s LUIS as its machine learning foundations – although there would still be cases where more “traditional” AI has its uses.

One aspect that has given Vodafone confidence is that it has found a good cost profile for the service.

Initially the enhanced TOBi service was expensive to run, Garcia admits. But within six months the company changed to another model (he wouldn’t say which two models were involved) and were able to run the service ten times cheaper.

Gen AI principles

That, Garcia says, underscores key Gen AI principles Vodafone is adopting. First is the importance of well-structured and available data. Vodafone has an internal capability it calls AI Booster, which structures all its data to be ready to be accessed by its AI apps. This also tracks the LLM and the software version being used which is a regulatory requirement, but one that means it can apply its Booster to any apps across the European footprint.

The second is the ability to swap vendors and retain flexibility in Gen AI models, to be able to move fast. The operator does not see the value in investing in its own GPUs or tying itself too closely to any LLM provider or foundation. Instead it has focussed on partnership, using secure containers in public cloud capabilities to run a selection of LLMs and focussing its capabilities on fine tuning those by training them on its own data.

Speaking of data, Vodafone is also proud of its 24 Petabyte data ocean, which takes in information from all its network systems and probes, as well as billing and charging data, on top of which it was now built over 600 models, from “next best action” models to predictive network apps.

It has also created and recruited a cadre of experts within the organisation, over 50 data scientists and prompt engineers, as well as a wider skills training initiative that runs right up to senior management.

AI and the network

Although Vodafone is using GenAI in internal process – with the key three areas being software development, customer experience and network operations, in the future, Petty said, AI could have a bigger impact on the structure of its network itself.

That’s because on-device AI may well create a demand for network connections with much lower latency, from the device AI interacting with edge or centralised data centres. That could mean a wider requirement for the specific lower-latency capabilities of 5G SA, for which operators may need to roll out wider and even ubiquitous coverage.

“The investment in latencies is going to be critically important. We’re fortunate that 5G Standalone delivers low latency capability – but it’s not deployed at scale. We don’t have ubiquitous coverage. We need to make sure that those things are available to enable those applications.

“It may also lead to a different economic model in the internet, where customers reject the “free” model that sells them advertising based on their data,” Petty said.

“It’s not clear which economic model is going to win in that space. You’ve seen people ike Google announce they’re trialling paid-for search. So how that model evolves will dictate where economic funding in the internet is and then how we build our networks to support that. I think we’ll see an evolution of a two-sided economic model that we probably didn’t get in the growth of the Internet over the last 20 years.”