Nubank's five support agents: card-delivery agent gains 37 points of transactional NPS in A/B tests

Nubank's KDD '26 paper describes five customer support agents in production and the evaluation pipeline behind them. One A/B test showed a 37-point NPS gain and a 29-point self-service gain.

Organization: Nubank
Industry: Fintech
When: Jun 2026
Evidence: Research paper
Agents and tools: LLM judges GEPA prompt optimization

Reported outcome

In A/B tests on the card-delivery agent, AI transactional NPS rose 37 points and self-service rate rose 29 points over prior agent variants.

+37 points (paper)
AI transactional NPS gain
+29 points (paper)
Self-service rate gain
5 (paper)
Agents in production

Caveats

Described in a paper with a stated method. Peer review status varies.

The headline gains are for the card-delivery agent only, compared with Nubank's earlier agent variants, not with human agents.
The paper is by Nubank's own staff. We read the abstract page, not the full paper, so we have not checked the method or the authors' stated limits.
The abstract says AI satisfaction on most use cases is within a few points of human experts, but gives no exact figure.

What the agents do

Nubank runs five support agents: card delivery, debt management, credit-limit support, card management, and product explanation. The paper is mostly about how the team builds and tests them.

Why this works

The authors' central claim is that the quality of the evaluation pipeline sets how fast a team can improve its agents. Their framework has four parts: structured context engineering, human-in-the-loop prompt iteration, LLM-judge evaluation with measured agreement between raters, and a path from idea to production validation.

What to copy

Test an LLM judge against human raters before you trust it.
Run live A/B tests against your previous agent version, not only offline tests.
Track a customer-reported score (like NPS) and a task score (like self-service rate) together.

Starter prompt

Written by us, not by Nubank.

You are grading a customer support conversation.
Conversation: {transcript}
Goal of the customer: {goal}

Score 1-5 on each: resolved the goal, accurate, clear, appropriate tone.
Give one sentence of evidence per score, quoting the conversation.
If a score is below 3, name the exact turn where it went wrong.