We’re putting $5.5M behind research and development of smaller, specialized AI models for real-world deployment — Read our manifesto →
    All posts

    [Case Study] 60 Million Conversations a Month: Vodafone's Enterprise AI Operating Layer

    How Vodafone scaled AI across customer service, network operations, employees, software engineering, sourcing, marketing, and HR.

    Conscious Engines

    Industry: Telecommunications
    Organization: Vodafone
    Use case: Customer, network, employee, and enterprise-process AI
    Evidence basis: Vodafone H1 FY26 investor presentation
    Disclosure: This is an independent analysis by Conscious Engines based on company investor reporting. Results are company reported and may use internal definitions.

    1. Outcome at a Glance

    Vodafone reports approximately 60 million AI-supported customer conversations per month. Its SuperTOBi assistant operates across European markets and reaches 70% end-to-end resolution, with an eight-point NPS improvement over the previous AI experience.

    Key Outcomes

    About 60 million per month

    Customer conversations

    Reported in the cited source.

    70%

    SuperTOBi end-to-end resolution

    Reported in the cited source.

    +8 points

    NPS versus previous AI

    Reported in the cited source.

    +61%

    German agent-assist helpfulness

    Reported in the cited source.

    43% reduction

    Network zero-touch MTTR

    Reported in the cited source.

    AreaReported result
    Customer conversationsAbout 60 million per month
    SuperTOBi end-to-end resolution70%
    NPS versus previous AI+8 points
    German agent-assist helpfulness+61%
    Network zero-touch MTTR43% reduction
    Employee assistantMore than 50,000 users, over 90% monthly adoption
    Employees reporting productivity benefit93%
    Coding support2,800 engineers, 30% acceptance, 12% productivity gain
    Sourcing cycle improvement30% reduction
    Campaign analysis time85% reduction
    HR agentMore than 69,000 users, 86% resolution

    The breadth is as important as the scale. Vodafone is using different AI capabilities for conversations, network repair, internal knowledge, code, procurement, marketing, and HR. That is a portfolio of task-specific systems, even when some share foundation models.

    2. The Operational Problem

    Telecom operators run two tightly coupled businesses: a high-volume customer service operation and a complex physical network. A customer issue may involve account data, device configuration, billing, coverage, service incidents, or field repair. Resolution often crosses systems and teams.

    Traditional bots can answer simple questions but fail when they must authenticate the customer, inspect service state, execute a change, or explain an exception. Network tools can alert operators without assembling cause, impact, history, and the next approved action.

    At Vodafone scale, cost, latency, and reliability become model-selection constraints. Sending every customer turn, log line, and employee query to the largest model would be expensive and hard to govern. The enterprise needs routing, specialized models, deterministic tools, and shared observability.

    3. What Was Built

    Vodafone's investor materials describe an enterprise AI layer across several domains.

    System at a Glance

    SuperTOBi

    Conversational model plus customer and service actions.

    Agent assist

    Real-time transcript, retrieval, recommendations, and summary.

    Network operations

    Event correlation, root-cause support, and zero-touch remediation.

    Employee assistant

    Enterprise RAG over internal knowledge and productivity tools.

    Coding

    Code models, repository context, tests, and developer review.

    HR and sourcing

    Process agents with policy retrieval and workflow integration.

    DomainLikely system pattern
    SuperTOBiConversational model plus customer and service actions
    Agent assistReal-time transcript, retrieval, recommendations, and summary
    Network operationsEvent correlation, root-cause support, and zero-touch remediation
    Employee assistantEnterprise RAG over internal knowledge and productivity tools
    CodingCode models, repository context, tests, and developer review
    HR and sourcingProcess agents with policy retrieval and workflow integration

    End-to-end resolution implies action capability. The assistant must move beyond conversation into authenticated tools while respecting permissions and business rules. Network remediation needs even stricter action tiers, simulation, rollback, and human override.

    At 60 million conversations per month, routing economics matter. Small intent, language, sentiment, and extraction models can process early stages. Retrieval and deterministic APIs can answer known requests. Stronger generative models can handle ambiguous cases. This cascade reduces cost and isolates failures.

    4. How It Reached Production

    Vodafone's portfolio points to a platform operating model rather than isolated proofs of concept.

    Create shared model services. Identity, retrieval, evaluation, observability, safety, routing, and cost attribution can serve many use cases.

    Keep domain ownership. Network, service, HR, sourcing, and engineering teams need their own workflows, data owners, and release thresholds.

    Measure resolution, not contact deflection. A 70% end-to-end resolution rate is more useful than a conversation count if it excludes rapid repeat contacts and incorrect actions.

    Use action tiers. Informational answers, reversible account actions, commercial changes, and network controls require different authority.

    Manage model cost by outcome. Track cost per resolved contact, accepted code change, resolved HR request, and recovered network incident. Token cost without outcome attribution obscures economics.

    The reported 30% code acceptance and 12% productivity improvement also show why output volume is a weak measure. Adoption, acceptance, quality, and delivered work matter together.

    5. What Telecommunications Leaders Should Take Away

    Vodafone provides a mature example of enterprise AI as an operating layer. The company did not stop at a customer chatbot. It applied models to the network and internal processes where telecom-specific data creates an advantage.

    A production telecom program combines reusable private infrastructure with bespoke domain models: telecom ASR, multilingual TTS, customer and field voice agents, network-event models, enterprise RAG, and task-specific SLMs. A central router should select models by risk, language, latency, and complexity.

    The key metric differs by workflow, but all should connect to an accepted outcome. For service it is cost per resolved contact. For the network it is safe MTTR reduction. For coding it is accepted, tested software change. Portfolio scale is justified only when each model has an owner and measurable value.

    Sources