A customer behind on a phone bill can agree to pay through an AI agent at 11 p.m., long after human collectors have clocked out. At TIM Brasil, about one in four completed negotiations with its collections agent happens outside standard business hours. That gives customers another way to engage with TIM when its human operation is unavailable. The bigger challenge is connecting those gains to the P&L while deciding when a person still needs to be part of the conversation.
Claudio Luiz da Silva is the Director of AI and Transformation for Customer Relations at TIM Brasil, one of Brazil’s largest mobile operators. His work at the company has included the youth mobile program TIM beta, and he now leads the team putting AI agents into collections and customer service. He treats each deployment as a decision about how TIM does business, with the technology question coming second.
“This is the first technology that interacts the same way humans do, through conversation. We’re at the very beginning of this, because for every business leader, the conversation should be about how we change the way we do business and make it better for everyone. That means profits, but also employees and final customers,” says Silva.
Collections gained a second operating model
If anyone had asked TIM’s collections team what it needed, the answer would have been better collections results. Silva sees a larger opportunity in changing how the work gets done. His team built an agent-driven process alongside TIM's existing collections operation, leaving the original process running. “I can improve the results with the same ways of work, but I can bring new results with a new way of doing business. That’s the difference,” Silva explains.
The new process changes more than the way collectors contact customers. In Brazil, Silva says, companies commonly outsource collections to external partners. An AI agent gives TIM the option to handle that work internally and observe the customer experience interaction by interaction. “In Brazil, most companies do collections through an external partner. When I have a solution like AI agents, I don’t need to give that work to a partner. I can do it myself, and I can see point by point what happens in the customer experience.”
Bringing that work in-house also meant rebuilding the way the operation was measured and managed. Within roughly three months, Silva’s team reworked its KPIs and reset commitments with the people involved. The original collections process remained in place, while the new agent created another route to results. The agent has reached more than 2 million customers on WhatsApp, and the combined approach has held the original results while lifting collections recovery by roughly 7 percentage points. The agent also extended the operation beyond commercial hours. That additional availability gives customers a way to engage with TIM outside the human operation's working day, though the share of interactions outside those hours doesn't, by itself, establish the financial impact.
Productivity only counts when the business captures the value
TIM’s other agent works for the employees. TAIS Real Time listens to human operators on live calls and suggests the best next step, and it has supported more than 1.5 million calls with a double-digit reduction in handling time. By the usual measure, that’s a success. But shorter calls don’t show whether the business has captured the value of the time saved. Until it shows up as revenue or cost, it’s an operational win still waiting for a business case. “If we only measure productivity, we can get lost. The KPI should connect with the business case, but in the end it needs to bring P&L impact, because that’s how we connect with the board and build commitment,” says Silva.
The difficulty of translating AI adoption into financial results gives that approach broader relevance. Only 39% of respondents in a global survey attribute any EBIT impact to AI, and most of them put AI’s share below 5% of their organization’s EBIT. A shorter call can demonstrate that an AI copilot has changed an employee's work. It doesn't, by itself, show how that change affects the economics of the operation. TIM's approach starts with the business case and ties its measures to the outcome the company needs.
For Silva, that connection also helps secure the commitment of the board and employees. AI changes how people work, and the business needs to understand both the results it expects and the contribution employees make in a redesigned process. That can require reconsidering the wider process, not simply measuring the task an agent or copilot performs.
Agents own only the steps that need a model
Before a company builds an agent, it should ask whether the use case needs one at all, since many use cases labeled agentic don’t require agentic implementations. Silva applies that same skepticism inside a single workflow. Much of a collections conversation is scripted, and he sees no reason to hand that part to a language model.
Knowing where that line falls is a discipline in itself. Philip Samuelraj, Founder and CEO of Techjays, which reimagines mid-market companies’ operations around AI, says most companies start the whole thing from the wrong end.
"Everybody starts by asking which tasks they can hand to a model. That gets you the same answer everyone else gets, which is, by definition, average. The real question is what the work should even look like if you built it today, and then deciding, piece by piece, what to rebuild, what a model should run, and what to leave alone," says Samuelraj. "Knowing which is which is the whole skill. There’s a taste to it you don’t get from a checklist."
At TIM, that decision shows up in the small mechanics of a call. Silva points to the opening of a routine collections conversation as a place where the judgment is easy.
“When the collector says good morning, you have a bill, and your bill is this amount, that doesn’t need to be an agentic solution. That’s a deterministic approach,” he says. Placing the LLM only where a conversation calls for judgment shrinks the territory guardrails have to cover. For the steps that run autonomously, TIM uses supervisor agents that monitor other agents against business rules.
“We have supervisor agents that work side by side with the other agents. Their role is just to see what those agents are doing, and if an agent does something outside the policy or the business rules, the supervisor intervenes,” says Silva. Some of the most revealing decisions to keep a person involved have less to do with risk than with the experience TIM wants to offer.
Choosing a human can be the customer strategy
The decision to involve a person doesn't always come down to whether an agent can handle the interaction safely or successfully. Sometimes, the company's customer experience strategy calls for a human even when the technology is capable of doing the job. “I have the technology. I can do this. That’s really good, but is it good for the customer? Sometimes no. When you have maturity in this approach, you understand that being able to do something doesn’t mean you should do it,” he says.
He points to sensitive situations and particular groups of customers, including older people, as cases where a human-led interaction may be the better choice. In those situations, a copilot can help the employee do the work without taking the conversation away from the person. “If I put an agent to talk with my mother by text right now, it doesn’t work. My mother wants to talk with people. Sometimes it’s better to connect a human with a copilot, and not an agent that does the transaction.”
Rather than a limitation, Samuelraj frames that preference as a business asset. “You can automate the transaction. What you can’t automate is the person on the other end feeling like somebody actually cared. Some customers want a human, full stop, and that connection is the thing you’re really selling. I’d automate everything around it to protect it, not replace it,” he says.
A supervisor agent monitoring an autonomous system and a human employee leading a customer interaction serve different purposes. In the latter case, the employee leads the conversation while AI provides support in the background. The technology changes how the person can do the job without determining that the person should disappear from it.
Silva expects the boundary between human and automated interactions to evolve as customer preferences change. For now, the decision rests on what the customer needs and what TIM wants the experience to deliver.
