MartechARTICLE

Cielo Cuts Time Between Customer Complaint and Product Launch by 76%

Cielo Cuts Time Between Customer Complaint and Product Launch by 76%
Image: Vanessa Guerra
Article · Martech
Reading
6 min
In short

Cielo cut the average development time for new features from 50.5 to 12.3 days by using AI agents as a quality filter before the product team decides what to build. The model, in use since July 2026, shows a path to capacity gains without headcount growth.

The bottleneck Cielo attacked wasn't the code, it was the decision

Cielo, one of Brazil's largest payment processors, solved a problem every founder recognizes: code is rarely the bottleneck, the decision about what to build is. Since July 2026, the company has used artificial intelligence agents in a product management division to turn complaints, satisfaction surveys, and usage signals into feature proposals ready to enter development, according to a report published by Exame, a Brazilian business magazine, on November 18, 2026.

The angle matters because this isn't a case of generative AI writing code faster. It's a case of AI applied to the phase no staff engineer controls alone: discovery, the stage where a complaint turns into a requirement, a business rule, and a prototype. It's there, not in the development sprint, that Cielo says it found its biggest speed gain.

The number that matters to the board: from 50.5 to 12.3 days

The metric that goes into any board presentation is straightforward: the average time to develop a new feature dropped from 50.5 to 12.3 days, a 76% reduction, with the median falling from 39 to 12 days. More than 20 initiatives have already gone through the new workflow, including refunds for Pix, Brazil's instant payment system, in the Cielo Gestão app (ready in under a week) and changes to the receivables anticipation flow handled by Assistente Cielo, a chatbot that answered questions about statements and products.

Two points deserve the attention of whoever approves the innovation budget:

  • Cielo also disclosed the picture of the previous problem: 51.2% of requests spent more than 30 days just in the preparation phase, and 23.3% exceeded 45 days.
  • After the agents, no tracked initiative stayed above 30 days, and about 15% were completed in under five days.

In terms of delivery capacity, that's equivalent to more than tripling the product team's throughput without hiring a single additional analyst. It's this argument, throughput per head, that justifies investment in process AI better than any discourse about innovation for its own sake.

How the funnel works: two agents, two roles

The design is simpler than Cielo's 500-agent figure suggests. One agent, named Assistente de Discovery, coordinates the work of the others: it gathers customer reports from satisfaction surveys, support, and product usage, and organizes everything into a single interface for the team. Another agent acts as a quality gate: before a project moves to development, it checks whether there is evidence supporting the proposal, a defined business rule, a ready prototype, and a criterion to assess whether the feature is ready to go live, as well as verifying whether a post-launch monitoring plan already exists.

If something is missing, the system flags what's missing and how severe the gap is, but it doesn't decide on its own. This separation of roles, the agent proposes and audits, the human decides and is accountable, is the point any founder should copy before copying the technology itself.

Technology only changes the outcome when it changes the way people work. The gain didn't come from creating agents, it came from redesigning how a product decision is prepared, so that it reaches the customer with more evidence and less waiting time.

Carlos Alves, Vice President of Technology and Business at Cielo

"The agents research, consolidate, and compare alternatives, but the decision still rests with the product team, which is accountable for the outcome," Alves adds, in a statement to Exame. The legal and reputational risk of a wrong product decision still falls on a person, not a model, and it's this accountability standard that sustains the board's confidence in the process.

What Cielo didn't disclose (and why that's the right question for the CTO)

Exame's report is honest about the gaps in the material disclosed by Cielo, and a founder should be equally honest when evaluating the case. The company doesn't specify the exact dates used to build the historical baseline of 50.5 days, nor the precise start and end markers of the measurement that produced the 76% reduction. In the receivables anticipation project, Cielo states that the volume anticipated by Assistente Cielo doubled after the workflow change, but doesn't disclose figures in reais nor the time interval used in the comparison.

That doesn't invalidate the case, it only invalidates the temptation to treat the number as a closed benchmark. The question a CFO should bring to the CTO before replicating the model isn't "how much time does this save," it's "what's the exact baseline, measured over what period, and does doubling a channel's volume mean doubling revenue or just redistributing operations that already existed elsewhere." AI hype without these closed numbers is vendor talk, not a board decision.

Replicating without 500 agents: what a founder takes from the case

Cielo states it operates more than 500 AI agents across different areas, with a product system built in-house and more than 150 updates since the first version. It's tempting to think the model only works at that scale. It doesn't: the 76% gain came from a single product management division, with two well-defined agents, not from the entire fleet.

For a startup or scale-up that wants to apply the same logic without the budget of a payments incumbent, the replicable roadmap has three questions, not three tools:

  1. Where's the real bottleneck, in the code or in preparing the decision of what to build?
  2. What minimum evidence (usage data, complaints, prototypes) should the product team already require before any line of code, with or without AI in the middle?
  3. Who remains responsible for the final decision, and how is that recorded if the outcome goes wrong?

Cielo had also already used the same type of agent to cut internal IT tickets in half, according to another report from Exame itself, a sign that the company's bet is recurring, not an isolated pilot meant to generate news.

The risk left out of the announcement: does the gate become a bottleneck?

Every automatic auditing mechanism carries an inverted risk: the agent that checks whether there is sufficient evidence can, in practice, block genuinely new ideas that don't yet have historical data to support them. Cielo describes the system as something that flags gaps and lets the team decide, which is the correct design, but the announcement doesn't include a single case of a project blocked or delayed by the gate itself.

For the founder who's going to copy the model, this is the metric that's missing from day one: how many proposals the quality agent rejected, and how many of those, even rejected, turned out to be right. Without this counterbalance, a system designed to speed up discovery can, over time, turn into just another layer of compliance disguised as agility.

Translated from the Brazilian Portuguese original · Read the original

More from Vanessa Guerra
View profile →