From using AI to operating with AI: What changes in product management when the bottlenecks disappear
B · Useful, limited A Nubank engineering-blog account of one lead product manager's three concrete AI workflows, with enough process detail to copy; credible practice, but every outcome is self-reported by the person who built it.
Summary based on the full text·Original: about 9 minutes to read
In short
A post on Building Nubank, written with Cinthia Scarlatti, Lead Product Manager for Growth in Nubank's US expansion, based on her May 2026 talk at the PM3 Summit in São Paulo. Her method: map where work waits in someone else's queue, and point AI agents at those handoffs. She describes three cases: a code fix she specified for a coding agent, a data question answered in one session, and an agent that keeps a regulated, multi-stakeholder project documented every day. Decisions and validation stay with the PM.
Key numbers
- more than a week to less than a daylead time for a small production fix once the PM had an AI coding agent prepare the pull request
- around 1 hour per dayPM's estimate of time saved by a daily project-digest agent
- more than 60open project questions tracked by that agent
- 6:30 p.m.daily run time of the digest agent, which follows seven steps
As reported by the source.
English·Web page·opens building.nu.com
What it reports
- Code fix: an audit found that guidance for customer service agents handling deceased-customer cases in Mexico pointed to the Brazilian support email. Text was organised by language, so editing the Spanish copy would also change Colombia. Scarlatti wrote a Cursor prompt structured like a Jira story (problem, current behaviour, desired behaviour, constraints). The agent confirmed the change had to be made by country, offered two options, implemented the chosen one and opened a pull request that passed checks and was approved by an engineer. Lead time fell from more than a week to less than a day.
- Data question: to check a suspected bug after a partial rollout, she gave Cursor some chat IDs. The agent set up her Databricks authentication using company documentation found through Glean over MCP, joined several datasets and switched away from a table that needed personal-data access. In a few minutes it showed agents were not using the new flow: a training problem, not a bug.
- Project digest: for a regulation-driven product in discovery with more than ten stakeholders and more than eight Slack channels, a Claude Cowork agent runs each workday at 6:30 p.m. It reads a Confluence page of open questions, the day's Slack threads, her journal, calendar and meeting transcripts, then updates the questions, writes a daily summary and drafts a Slack update. More than 60 questions have been mapped; she estimates it saves around 1 hour per day.
- Tool roles: Cursor for code, Claude Cowork for recurring automations and analyses, Gemini and ChatGPT for reviews and second opinions, Figma Make and Claude Design for prototypes, with MCP connections to Glean, Databricks, GitHub, Atlassian, Slack, Miro and Google Drive.
- Advice: start with whatever tools IT already allows, write prompts like requirements, build reusable systems rather than one-off chats, and validate column names, filters, sample size and table sources before using an analysis.
How they got it
First-person practice from one PM, written up with Nubank's editorial team. No measurement method for the time savings.
Limits
- Company blog; outcomes are the author's own estimates.
- One person's workflows in one company with broad tool access; she says the examples are meant to inspire, not to be copied.
- The code case was a small, well-bounded change reviewed by engineering.
Why it matters for companies
The useful part is the method, not the tools: find where a role waits on another team's queue and let an agent prepare a reviewable result instead of a ticket. A PM arriving with a tested pull request, or with a query whose sources she has checked, still uses engineering and data teams for approval but no longer waits for their capacity. This depends on agents having governed access to code, data and documentation through connectors, and on reviewers staying in the loop.
Summary based on the full text. Read the original for the full argument and data.Read the original on building.nu.com