Market · AI in practice
How Nubank speeds up support testing with simulated customers
AI programs simulate customers so the team can compare agent versions before testing with people, and the bank reported cutting average development time from 21.2 to 4.4 days.
What the sources report
On November 5, 2026, Nubank described a way to test support agents with more care. The team uses AI programs that simulate customers and talk to the agent. The bank calls these “synthetic customers.”
That lets the team compare agent versions before asking real people to join the test. Simulation comes first. Testing with people comes next. According to Nubank, this step helped cut the average development time for each version from 21.2 to 4.4 days.
Adriano's reading
For teams that put support agents into production, that pace changes the calendar. Less time stuck on each version leaves more room to adjust the product based on what people actually need.
I see a useful example of AI helping improve support itself with more speed. What interests me is the process: simulate first, test with people next, and only then widen usage.
Sources
- https://nu.com/pt/sala-de-imprensa/companhia/nubank-reduz-de-tres-semanas-para-menos-de-cinco-dias-o-ciclo-de-atualizacao-de-agentes-de-ia-ao-testa-los-primeiro-com-clientes-sinteticos (opens in a new tab) Nubank
- https://epocanegocios.globo.com/inteligencia-artificial/noticia/2026/11/de-3-semanas-para-5-dias-nubank-usa-clientes-sinteticos-para-acelerar-testes-com-agentes-de-ia-antes-de-implementacao.ghtml (opens in a new tab) Época Negócios
- https://arxiv.org/abs/2609.30137 (opens in a new tab) arXiv