An AI-resolved request only counts if the customer doesn't come back

At Centauro, an after-sales request counts as resolved when the artificial intelligence agent handles the ask and closes it without passing it to a person. By that criterion, the Grupo SBF retailer says more than 60% of the requests made through its app end this way, according to Exame's report of 11 November. The rate was measured between July and August 2026, applies only to the app, and the request volume was not disclosed.

These agents are part of a platform launched in the first quarter of 2026, with more than 900 agents in production. These numbers are reported by the company, and I found no independent verification. Centauro did not reveal how much it invested in the platform. As a gain, it says it has "optimized" more than 3,600 hours of work, in a calculation that, according to Ariovaldo Carmona, vice president and Chief Digital Officer of the group, compares the average time of each activity before and after the agent. The report does not give this gain in reais or say how many of these hours come from after-sales support.

Carmona told Exame that service became faster and that the chain started needing fewer extra hires during peak sales periods. From here on, the reading is mine. In after-sales support, much of the cost is likely to sit at the peak, because demand rises along with sales and the team grows to keep up. When some requests close without a human agent, that reinforcement can be smaller.

Any online clothing store in sale week would have an inbox full of customers asking where their package is or wanting to exchange a size. These requests fit a rule; a defective product or a refund outside policy already calls for someone on the team. The trap is in the word resolved. If the customer closes the conversation with the agent and comes back two days later about the same package, the metric counted one resolved request and the operation did the work twice. The calculation that matters puts spending on the tool and on exceptions next to how many customers had to come back.

The case points to something many companies probably have in house: the repeated request that swells the queue precisely when sales are going well. In my reading, that is where automation has the best chance of paying for itself, because the volume that today turns into extra hiring gets absorbed without the team growing at the same pace. By that reasoning, the rest of after-sales support can wait until that first use shows, with the operation's own numbers, that customers are not coming back.

Sources

Exame, 11 November 2026, Tamires Vitorio. https://exame.com/inteligencia-artificial/centauro-coloca-900-agentes-de-ia-para-trabalhar-e-resolve-60-dos-chamados-sem-humanos/