A strong new AI adoption case landed on June 7, 2026, when Nubank researchers published a paper on building customer-support AI agents at 100 million-plus user scale. The headline result is hard to ignore: in one card-delivery deployment, the company reported a 37 percentage-point improvement in AI transactional NPS and a 29 percentage-point gain in self-service rate versus prior agent variants. On most use cases, the paper says AI satisfaction now sits within a few percentage points of expert human agents.
That would already be notable. What makes the case stronger is that it lines up with earlier production metrics Nubank shared with OpenAI. In that customer story, Nubank said its AI-powered assistant resolves 55% of Tier 1 inquiries, handles more than 2 million monthly chats, and cuts chat response times by 70%. The same story says customer queries are resolved 2.3x faster with higher accuracy, while internal enterprise search is used by more than 5,000 employees per month.
Taken together, that gives us one of the clearest recent business cases for AI in customer support: not because the models are flashy, but because the company treated evaluation, retrieval, and escalation as production disciplines.
Successful customer-support AI is not a bot with a better personality. It is a measured operating system for deciding what the model should answer, what context it should see, and when a human should take over.
What Nubank Actually Built
The June 2026 paper describes five production deployments spanning card delivery, debt management, credit-limit support, card management, and product explanation. That matters because many AI support programs still live as one narrow pilot with one carefully chosen use case. Nubank is already past that stage.
OpenAI's customer story fills in the operating details. Nubank first built a custom enterprise-search layer using GPT-4o and GPT-4o mini so employees could retrieve FAQs, policy documents, and brand guidance quickly. That search foundation then fed a Call Center Copilot for live agent assistance, including suggested replies, chat summarization, and step-by-step guidance for harder cases. Nubank also deployed an AI assistant that speaks directly with customers before escalating when necessary.
This sequence is the important part. The company did not start by throwing a model into the contact center and hoping for good outcomes. It first improved retrieval, then improved agent assistance, then expanded customer-facing automation. That is a much more defensible path than trying to automate the entire frontline in one step.
Why This Looks Like a Real Business Case
Three things make this case stand out.
First, the numbers are operational rather than theatrical. A 37-point uplift in AI transactional NPS and a 29-point lift in self-service rate are not vanity engagement metrics. They indicate customers are completing more of the support journey successfully while rating the experience more positively.
Second, the published approach explains why the gains happened. The paper argues that evaluation-pipeline quality determines iteration velocity. In practice, that means Nubank linked offline simulation, structured context engineering, prompt iteration, and online measurement instead of treating them as separate projects. That is the difference between a support bot that looks promising in a demo and one that survives real traffic.
Third, the OpenAI case study shows the gains are not confined to one research result. Nubank says over 45% of agents use the copilot's key features, the assistant handles more than 2 million chats each month, and response times are down 70%. Those metrics suggest the system is already embedded in the company's service model, not sitting beside it.
There is also a scale signal here. OpenAI says Nubank now serves more than 114 million customers across Brazil, Mexico, and Colombia. That does not automatically prove ROI, but it does show the system is being operated under serious demand conditions rather than boutique traffic.
What Other Companies Should Copy
Most companies do not have Nubank's scale, but the design logic transfers well:
- Fix retrieval before fixing conversation. If the model cannot reliably access current policy and case context, the conversation layer will drift.
- Measure against prior production systems. Nubank's paper compares new agent variants with earlier ones instead of only presenting absolute scores in isolation.
- Automate simple tiers, not everything. Resolving Tier 1 issues well is a stronger business move than forcing full automation into edge cases too early.
- Use AI to help agents before replacing them. Copilots, summaries, and guided next replies reduce cognitive load and raise consistency.
- Link offline evals to online results. If test metrics do not predict production behavior, iteration stays slow and trust stays low.
This playbook is relevant well beyond banking. Insurance, telecom, healthcare administration, travel, utilities, and ecommerce all run support operations where retrieval quality, escalation discipline, and repeatable issue categories matter more than chatbot novelty.
The Caveats
This is still a case that requires careful reading. The June 2026 paper is a strong primary source, but it is authored by people involved in the deployment rather than by a neutral third-party auditor. Nubank also does not publish a full economic model showing cost-to-serve reduction, labor savings, or exact revenue impact from higher satisfaction and self-service.
There is also a transferability issue. Nubank has a digital-native operating model, large support volume, and a broad stream of repeated issue types. Businesses with weaker data hygiene, fragmented internal documentation, or less mature escalation design will not get the same results by simply purchasing a model API.
Still, this remains one of the most credible recent AI support cases because the evidence goes beyond claims of "better service." It includes production outcomes, scaling evidence, and a published explanation of how the team turned evaluation into a repeatable engineering capability.
The Business Takeaway
Nubank's latest case shows customer-support AI works best when the model is only one part of the system. The stronger business result came from combining retrieval, agent assistance, measured escalation, and continuous evaluation. That is why the company could move from faster response times and Tier 1 automation to a 37-point jump in AI transactional NPS and a 29-point increase in self-service rate on a real deployment.
If you are building your own AI adoption case, the lesson is simple: do not start by asking whether the model is smart enough. Start by asking whether your knowledge base is current, whether your evals predict production outcomes, and whether your workflow tells the model when to answer and when to hand off. That is where customer-support AI stops being software theater and starts becoming operating leverage.
Sources & Further Reading
- arXiv: Building Customer Support AI Agents at 100M-User Scale: An Evaluation-Driven Framework — June 7, 2026 paper providing the 37-point transactional NPS lift, 29-point self-service gain, five production deployments, and the core argument that evaluation quality drives iteration velocity
- arXiv PDF: Building Customer Support AI Agents at 100M-User Scale — Full paper with deployment framing, methodology, and additional technical context behind the production rollout
- OpenAI Customer Story: Nubank elevates customer experiences with OpenAI — Official deployment metrics covering 114 million-plus customers, 5,000 monthly enterprise-search users, 55% Tier 1 resolution, 2 million monthly chats, 70% faster chat response times, and 2.3x faster query resolution