Vivo (Telefônica Brasil)
Vivo gives 11,000 call-centre agents a generative AI copilot, I.Ajuda
Vivo, the Brazilian telecom operator, built I.Ajuda, a chat assistant that answers call-centre agents' questions from its internal systems and process documents. By June 2024 all 11,000 agents used it, and Vivo reported shorter average handling times.
Key figures
Copied exactly from the sources. Copy any figure with its citation.
The problem
Vivo's call-centre agents had to look up answers for customers across more than 30 systems and 2,000 processes. Searching these during a live call slows the conversation and can lead to uneven answers.
What they built
I.Ajuda is a chat tool for agents. An agent types a question about the customer's issue, and a generative AI model, fed with content from those systems and processes, replies with a short, direct answer. In effect it is a copilot that sits beside the agent during the call.
How they did it
Vivo built it with Microsoft on Azure OpenAI Service, with Telefónica Tech, the group's business-services unit, doing the implementation. The work followed a Vivo and Microsoft collaboration announced in April 2023. Vivo says it processes and stores all information inside its own closed environment. Agents rate each suggested answer as positive or negative. A team called CT Bots, made up of 26 people who were among the best agents, reviews those ratings and the AI's answers to keep the tool accurate. Vivo says it started the project in mid-2023.
Results
By June 2024, I.Ajuda was used by all 11,000 of Vivo's call-centre agents. Vivo reports that average handling time fell 9% for individual customers and 4% for business customers, and that the operation uses the assistant in 87% of customer interactions. Agents can reach the information they need on a single screen in 10 seconds at most. Vivo also reports more first-contact resolutions and a better service rating, without figures. Next, it wants the copilot to suggest offers to agents and to transcribe voice calls.
All reported figures
| Measure | Figure | Citation |
|---|---|---|
| Reduction in average handling time, individual customers (Vivo internal data) | 9%1 | |
| Reduction in average handling time, business customers (Vivo internal data) | 4%1 | |
| Share of customer interactions in which the operation uses the assistant (Vivo internal data) | 87%1 | |
| Time for an agent to reach the needed information on one screen | 10 seconds at most1 | |
| Internal systems and processes feeding the assistant | More than 30 systems and 2,000 processes1 |
Lessons for your project
Put the answers where agents already work. Vivo gathered content from many separate systems behind one chat window, so agents stop searching several tools during a call.
Staff a team to train the assistant. Vivo kept 26 of its best former agents reviewing answers and agent feedback to keep the tool accurate.
Collect agent ratings on every answer. Thumbs-up or thumbs-down feedback gives the review team a steady stream of corrections.
Track operational and customer measures together, such as handling time, first-contact resolution and satisfaction score, so efficiency gains are not bought with worse service.
Caveats
- All figures are Vivo internal data, reported in a joint Vivo and Microsoft press release. No audit or independent measurement is cited.
- Vivo also reports higher first-contact resolution and a better service rating but gives no numbers, so they are not included as metrics.
- The model is not named. Sources say only that the solution runs on Azure OpenAI Service.
- The start date is given only as 'mid-last year' in a June 2024 statement, so the deployed date is shown as the year 2023. The release does not say when the assistant reached all agents.
Sources
- Vivo, em colaboração com Microsoft, cria IA Generativa para assistente do atendente de call center
Microsoft News Center Brasil · 2024-06-20 · Press release · news.microsoft.com
- Vivo adota copiloto com IA generativa para seu call center, o I.Ajuda
Mobile Time · 2024-06-20 · News · mobiletime.com.br
Summarized in our own words from the sources above. Spotted an error? See corrections.