How Vodafone Gave Its Chatbot a Brain Transplant and Reached 70% Resolution on 60 Million Conversations a Month
Vodafone rebuilt its seven-year-old TOBi chatbot on Azure OpenAI, held it back until accuracy passed 90%, and rolled it out across every European market. The group now reports 60 million conversations a month with 70% resolved end to end, and its CTO spent the whole rollout calling generative AI overhyped.
The starting point
Vodafone serves more than 330 million mobile and broadband customers across 15 countries, with around 100,000 employees. Its chatbot, TOBi, had been in service since 2017, and by 2024 was fielding nearly 45 million customer questions a month in 15 languages.
TOBi was a first-generation bot: intent recognition built on keyword matching, running on Microsoft's LUIS in some markets and IBM Watson in others. It handled simple, high-volume queries and routed the rest to humans. Its ceiling showed in the journey-level numbers. On appointment booking in Portugal, first-time resolution sat at 15%.
In January 2024, Vodafone signed a 10-year partnership with Microsoft, committing $1.5 billion to cloud and AI, and later reallocated €140 million in a single financial year to the customer experience transformation. Rather than launch a new bot, the team replaced TOBi's brain: same channels, same routing, new large language model core on Azure OpenAI. Group CTO Scott Petty described the industry mood at the time with unusual candour for someone spending this much money: "GenAI is probably the most overhyped technology for many years in the telecom industry."
What they built
SuperTOBi, for customers
SuperTOBi understands full sentences rather than keywords, and handles complete journeys: booking engineer appointments, explaining bills, troubleshooting routers. When it hits its limits, it transfers to a human automatically. It launched in Italy and Portugal in early 2024, reached Germany and Turkey by July 2024, and was live in every European market by late 2025.
The accuracy story is the detail worth keeping. In early testing SuperTOBi answered correctly 75 to 78% of the time. Vodafone held the rollout until prompt engineering and knowledge work pushed accuracy to around 90%, and Petty publicly estimated it would take up to a year to deploy everywhere, deliberately slower than the technology allowed. The launch gate was accuracy, not capability.
SuperAgent, for the humans
Alongside the customer-facing assistant, SuperAgent supports human agents on the 30% of digital contacts escalated to them: conversational search over complex queries and automatic summaries of chat history so customers don't repeat themselves. It draws only on Vodafone's private knowledge base, a deliberate hallucination containment choice. The pairing is the operating model: the agent resolves the routine, the human gets better tooling for the rest.
The platform underneath
None of this sits on a one-off build. Vodafone's AI Booster platform runs 600+ machine learning models in production and provides a templated retrieval setup, multi-cloud model hosting, content moderation, and prompt-injection protection, built for EU AI Act compliance. Proof-of-concept to production fell from around five months to four weeks. Vodafone also runs a second 10-year hyperscaler deal with Google, signed October 2024, and credits both partnerships in its results statements.
The results
| Metric | Detail | Source |
|---|---|---|
| Conversation volume | ~60 million customer conversations per month | Vodafone H1 FY26 results (Nov 2025) |
| Resolution | 70% resolved end to end across all European markets | Vodafone H1 FY26, FY26 results |
| Customer satisfaction | +8 points NPS improvement group-wide; Portugal online NPS +14 to 64 | Vodafone IR, Vodafone press release |
| Journey-level gain | Appointment booking first-time resolution in Portugal up from 15% to 60% | Vodafone (July 2024) |
| Call handling | Calls shortened by at least one minute on average | Vodafone (May 2024) |
| Agent assist | 61% improvement in agent "helpfulness" rating in Germany | Vodafone H1 FY26 |
| Contact centre NPS | From low single digits to high 30s | Scott Petty (Sept 2024) |
| Employee AI | 68,000 Microsoft 365 Copilot licences; HR agent resolving 86% of requests across 69,000 employees | Microsoft, Vodafone IR |
| Agentic operations | Procurement agent covering 90% of tenders, 30% faster sourcing | Vodafone H1 FY26 |
What makes this case interesting
The rebuild reused everything except the intelligence. Vodafone kept the TOBi brand, channels, and escalation routes, and swapped keyword matching for a language model. Seven years of conversation design, routing logic, and customer familiarity carried over. Companies treating generative AI as a reason to start from scratch are paying twice for assets they already own.
Accuracy was a launch gate, not a press release. The 75 to 78% starting point would have been shippable by most vendors' definitions. Vodafone held for 90% and slowed a group-wide rollout for it, while the CTO called the technology overhyped in public. I think this pairing, sceptical language and heavy investment, is what serious adoption looks like. The opposite pairing, breathless language and thin deployment, is what you should screen for in every vendor pitch.
The famous numbers come with footnotes, and the footnotes are the lesson. The widely quoted "15% to 60% resolution" is one journey in one market, not a group figure. The "1 million conversations a day" number circulating in coverage is the old TOBi's 2018 figure; the current verified rate is around 60 million a month. Vodafone's own materials are accurate about scope. The retellings are not. Any metric travelling without its denominator and date will inflate in transit, including yours.
Autonomy is being granted by tier. The estate now runs three levels: a voice agent for simple high-volume calls, SuperTOBi for complex journeys, and fully agentic deployments where the risk is contained and measurable, like procurement tenders. Customer-facing autonomy came only after the assistant proved itself, and internal agents got the most freedom where errors are cheapest to catch.
The challenges
The workforce context is uncomfortable and worth stating precisely. Vodafone announced 11,000 role reductions in May 2023, before SuperTOBi existed, as a turnaround measure. Vodafone Germany followed in March 2024 with 2,000 cuts explicitly linked to automation of manual tasks. The AI programme did not cause the restructuring, but it lands inside one, and any claim the two are unrelated would strain belief. Leaders running similar programmes should expect their AI story and their headcount story to be read as one story, whatever the org chart says.
Dependency is the second friction. A $1.5 billion Microsoft deal plus a $1 billion Google deal buys capability and lock-in at the same time, and Petty has warned publicly about big tech lock-in with generative AI while signing both. The dual-vendor structure is itself the hedge.
And most of the published numbers flow through Vodafone and Microsoft marketing channels. The investor relations figures (60 million monthly conversations, 70% end-to-end resolution, +8 points NPS) carry the most weight because misleading investors has consequences press releases don't.
Lessons for your programme
Set an accuracy threshold before customers see the agent, and hold it. Vodafone's gap between "works in testing" at 78% and "ready for customers" at 90% took months of unglamorous prompt and content work. Define your own bar and the measurement method before the pilot, using the experiment design in Section 07: The Experimentation Framework.
Build the platform once, deploy many times. AI Booster's shared retrieval, guardrails, and hosting cut deployment from five months to four weeks and made each subsequent use case cheaper. If every AI initiative in your business starts from bare infrastructure, fix this first: Section 05: The Tooling Landscape covers the platform choices involved.
Pair every customer-facing agent with better tooling for the humans behind it. The 30% of contacts SuperTOBi escalates arrive at agents equipped with SuperAgent summaries and search. Designing the human side of the handoff is change work, not technology work: Section 11: Change Management, Scaling, and Adoption covers it.
Grant autonomy by risk tier, not by ambition. Simple voice queries, complex assisted journeys, full agency only in contained domains like procurement. Map your own tiers before any agent goes live, with the classification approach in Section 08: Risk and Governance.
Sources
- Meet SuperTOBi, Vodafone's new generative AI virtual assistant (Vodafone, 2024)
- Vodafone supercharging customer experience with Microsoft's gen AI tools (Vodafone, 2024)
- Vodafone Group: transforming customer experience with Azure (Microsoft Customer Stories, 2025)
- Vodafone H1 FY26 results announcement (Vodafone Investor Relations, 2025)
- Vodafone and Microsoft sign 10-year strategic partnership (Vodafone, 2024)
- Vodafone: GenAI way overhyped, but will spend $151M to enhance its chatbot (IEEE ComSoc, reproducing Light Reading, 2024)
- How Vodafone plans to create an AI-powered tech comms powerhouse with Microsoft (Microsoft UK Stories, 2024)
- Vodafone touts AI progress at FY26 results (RCR Wireless, 2026)
- Vodafone Deutschland to cut 2,000 jobs (Broadband TV News, 2024)
- GenOps: Vodafone's generative AI platform on Google Cloud (Datatonic case study, 2024)