Vodafone's SuperTOBi: a generative-AI brain transplant for a telco chatbot now clearing 70%+ end-to-end resolution across every European market

Last Updated: November 10, 2026

The problem: a chatbot that could only answer chatbot-shaped questions

By May 2024, Vodafone's virtual assistant TOBi was not a pilot. It was infrastructure. Vodafone's own newsroom put it at nearly 45 million customer questions per month in 13 countries using 15 different languages. That scale is the whole story: at that volume, a percentage point of containment is worth more than most companies' entire support budget.

The constraint was architectural. TOBi was built the way chatbots were built before 2023 — classify the intent, match it to a flow, run the flow. Vodafone described the limitation bluntly when it launched the replacement: existing chatbot technologies typically can answer only simple questions based on a few key words. A customer who opens with a paragraph about a disputed charge, a device upgrade and a roaming add-on gets routed on whichever keyword wins.

The second-order damage is worse than the wrong answer. When the bot gives up, the customer repeats the whole story to a human who has no context, and often to a second human after that. Vodafone's European markets CEO later framed the target state as the inverse of exactly this failure.

support should resolve issues without customers being passed between chatbots, different agents and FAQs - where problems are solved first time, and customers only need to ask once.

— Ahmed Essam, CEO of Vodafone's European Markets, Vodafone newsroom, April 2026

What Vodafone actually built

In January 2024 Vodafone entered what its newsroom describes as a far-reaching 10-year strategic partnership with Microsoft. The customer-service workstream produced three components, shipped together rather than sequentially.

SuperTOBi — the customer-facing assistant

SuperTOBi is TOBi with a generative reasoning layer. Vodafone states it is powered by Microsoft Azure OpenAI and can interpret entire sentences and phrases rather than keyword fragments. Critically, the brand, the channels and the back-end integrations stayed put. This was a brain transplant, not a new product launch.

SuperAgent — the assistant for the humans

SuperAgent is built on Azure OpenAI's Agent Copilot and is deliberately walled: Vodafone says it only uses information from Vodafone's companywide and private knowledge database. Its job is retrieval over internal knowledge plus conversation continuity on escalation.

TOBi sends a summary of its customer conversation to the agent, so they can jump right in.

— Beverley Bartlett, Head of Digital Care, Vodafone, Vodafone newsroom, May 2024

The platform underneath

Vodafone's H1 FY26 investor deck names the substrate: a Vodafone AI Booster platform with GenAI reusable templates, a GenAI app catalogue, a “Data Ocean”, multivendor cloud and LLMs, and a testing, monitoring and responsible-AI layer. That is the mechanism by which one assistant becomes twelve national deployments instead of twelve rebuilds.

The measured outcome

The earliest hard number Vodafone published was a single journey in a single market. In Portugal, SuperTOBi took over appointment booking and, per Vodafone, first-time resolution rose from 15% to 60% while online net promoter score improved by 14 points to 64 — with the company noting that anything above 50 points is considered a strong result. Before that, internal call-centre testing had shown what Vodafone described as a 50% improvement in first-time resolution on critical journeys such as complex billing enquiries.

What makes this case worth studying is that the numbers then migrated from the newsroom to the investor deck, which is a materially higher bar.

Disclosure drift: H1 FY26 vs FY26

Two readings are available and the public record does not settle between them. The generous one: the programme is stable and Vodafone has moved to rounded “greater than” framing because the precision no longer adds information. The sceptical one: precise numbers become imprecise numbers when they stop improving. Note also that the headline metrics for adjacent systems changed category between the two decks — mean-time-to-repair in November, automated resolution in May — which makes year-on-year comparison impossible by construction.

What the public record does not establish

This is the part most write-ups skip, so be clear about it.

There is no euro figure. Vodafone has not published a cost saving, a cost-per-conversation, or an FTE equivalent attributable to SuperTOBi. The €140 million that circulates in coverage was, in Vodafone's own words, a reallocated investment for that financial year covering its wide-reaching transformation of the customer experience — not a SuperTOBi line item.

“End-to-end resolution” is undefined. No methodology note appears in either deck. It could mean no human handover, no repeat contact inside a window, or customer-confirmed resolution. Those are very different numbers, and a 70% under one definition could be a 40% under another.

The NPS benchmark is the old bot. Vodafone's own footnote states the improvement is measured against the NPS rating of conversations handled by the previous AI agent. That is a fair and transparent disclosure — and it is not the same as beating a human agent.

Nothing is published on failure behaviour. No hallucination or grounding-failure rate, no latency figures, no model versions, no evaluation harness, no safety or privacy incident data, and no information about what happens inside the roughly 30% of conversations that are not resolved. For a system touching tens of millions of customers a month, that is a large blank.

Nothing is independently audited. These are management-selected figures on investor slides, not assured disclosures.

And the employment question is open. On 9 November 2026, Vodafone Deutschland announced a restructuring reducing headcount by about 2,300 positions over two financial years while folding customer service into its sales units and widening AI use, as reported by PPC Land from the German-language Vodafone Deutschland release. That release quotes no executive, gives no current headcount, no savings target and no breakdown by function. Separately, Vodafone Group's FY26 deck reports 11,000 role reductions against a 10,000 target. It is tempting to draw a line from deflection rates to headcount. The public record does not support drawing it, and we are not going to.

Rollout timeline

What an integration team should take from this

1. Transplant the brain, keep the body

The expensive, risky parts of a support assistant are the integrations: identity, billing, provisioning, appointment systems, 15 languages, regulatory handling per market. Vodafone kept all of it and swapped the reasoning layer. If your current bot has those integrations and a bad brain, you are closer to a good outcome than a greenfield rebuild would suggest. This is the pattern we see most often in production workflow automation work and it is consistently underrated.

2. Ship the agent-assist in the same release

Deflection creates a harder residual queue by definition — the easy conversations stop reaching humans. Vodafone launched SuperAgent alongside SuperTOBi and restricted it to internal knowledge. Teams that ship deflection alone usually see handle time rise on the remaining tickets and conclude, wrongly, that the AI failed.

SuperAgent mirrors Microsoft Copilot to help our agents look up detailed information, ensuring they can rapidly resolve issues and provide more personalised customer care.

— Beverley Bartlett, Head of Digital Care, Vodafone, Vodafone newsroom, May 2024

3. Build once, deploy per market

Vodafone's reusable AI Booster templates are why a Portugal pilot became an all-markets rollout inside roughly eighteen months. Without a shared platform layer, each market negotiates its own vendor, its own prompts and its own evaluation — and the group number never materialises. Our model and platform selection notes cover the trade-offs between a single shared stack and per-unit autonomy.

4. Decide what “responsible” means before procurement, not after

Vodafone's framing here is unusually explicit. Its AI governance framework was developed with the University of Oxford, and the vendor choice was partly a governance choice.

Going beyond just the technology, Microsoft’s commitment to responsible AI was a differentiator. As we work with new technologies, this is a critical foundation.

— Ignacio Garcia, CIO Italy & Global Director Data Analytics and AI, Vodafone Group, Vodafone newsroom, May 2024

5. Be honest about where the humans go

Vodafone's public position is that automation should remove unnecessary interaction rather than human interaction. Its group CEO has put it this way:

AI will make real impacts to customer experience across businesses - leveraging the best that technology can offer, while at the same time relying on experienced and engaged customer care professionals.

— Margherita Della Valle, Vodafone Group CEO, Vodafone newsroom, April 2026

Seven months later, the German unit announced 2,300 role reductions in a restructuring that explicitly widens AI use. Both things are on the public record. Any organisation running a deflection programme should expect to hold both at once, and should decide now how it will talk about that internally — because the works-council conversation arrives whether or not you prepared for it. For more deployment write-ups of this kind see our case study library, or start a scoping conversation.