Rebuilding a design org around AI
New-feature delivery went from 2-4 weeks to 1-2, with the same human design and code gates. Screens from days to hours. 9 of 9 designers adopted the model, including its strongest initial skeptics.
- Role
- Design Manager, senior design leadership of the company
- Team
- 9 product designers across 3 squads
- Period
- July 2025 to present
- Context
- meutudo, a regulated Brazilian credit fintech serving millions of customers
- What I personally did
- the diagnosis and the operating model; the pitch that convinced leadership; the platform that makes the model possible, which has its own case on this site; and the change management, one conversation at a time.
The thesis
Code is just another prototyping tool, the way Figma is and paper once was. And drawing screens was never the job. The designer designs the experience and owns the patterns: defines them, evolves them, and knows when and how to break them. AI is excellent at creating what already exists. So I taught it our standard, and freed the team for the part that does not exist yet.
The situation
Nine designers stretched across too many squads, with operational work eating the hours real design needed. A new feature took 2 to 4 weeks from concept to deploy. The queue kept growing, and the obvious answer, hiring more people, would not change how the work happened. In a regulated market, speed without quality is a liability, not a win.
Convincing
The diagnosis I took to leadership fit in one sentence: what slows the team is not talent or effort, it is the operating model. I pitched an eleven-slide model: each designer as the guardian of a journey rather than a ticket handler, individual missions with clear indicators alongside a controlled backlog, and AI as a first-line executor that is always supervised by whoever owns the journey. Minimum governance, success metrics defined upfront, ninety days to roll out. The initiative was mine, from diagnosis to plan, and it was approved.
Building
For AI to execute inside our standard, the standard had to become something a machine reads. I built that platform myself: contracts, guardrails and automated review, described in its own case on this site. What matters here is what it changed in how the team operates: anyone can materialize an idea on-pattern, drift is caught by tooling instead of by memory, and the design and code gates stayed human.
It opened the same door from both sides. People who had never drawn a screen, and would not touch Figma out of fear, got a safe path to materialize ideas on-pattern. Designers, who master Figma but would not leave it on their own, got bridges into code starting from their own comfort zone.
Converting
The hardest resistance came from inside. Three fears, in their words: "am I supposed to become a programmer?", "you are looking at technology, not at design", and "we are handing our design to other people".
My answers became culture: technology has always been part of digital design, and code is a prototyping tool, not a career change. What we hand over is not the design, it is the execution of aesthetics inside a system we designed, govern, and are the only ones who know when to break. And if AI creates what already exists well, the designer's time moves up: from drawing screens to deciding what should exist.
Arguments alone convert nobody. What converted was proof and familiarity: the first deliveries landing in days instead of weeks, and bridges to the tools the team already knew. Today designers materialize their own solutions without waiting on the engineering queue, inside the same review gates as any other code. And the strongest skeptic, the one of "these things do not mix", is now among the people who deliver the most: what used to die in a backlog, he unblocks himself.
| Results | Before (Jul 2025) | Today |
|---|---|---|
| New feature, concept to deploy | 2-4 weeks | 1-2 weeks max |
| New screen | 3-5 days | under 1 day |
| Prototype | disposable artifact | production ready |
| AI adoption on the team | none | 9 of 9 designers, all committing to production |
Limits and learning
The model has deliberate borders: user research stayed human, AI is always supervised, and critical decisions belong to people. The lesson I would carry forward: emotional conversion only came after visible proof. The right arguments were not enough. Fear only gave way when the first delivery landed in days. If I did it again, I would engineer that first visible win in week one, before any slide.