AI CodeFix
Strategic launch of AI-powered code fixes at enterprise scale
Overview
When SonarQube flags a code issue, fixing it is still the developer's job. Read the problem, understand the context, write the fix, test it. , multiplied across thousands of flagged problems in any large codebase. By 2025, 42% of committed code was AI-assisted, accelerating the volume of new issues while the time to resolve each one stayed the same.
61% of developers say AI code looks correct but is not reliable. 88% report at least one negative impact of AI on technical debt. AI was accelerating creation while quietly making the maintenance side harder. The problem being designed for was not code generation. It was what happens after.
AI CodeFix was SonarSource's answer to the other side of that equation: surfacing a fix the moment an issue is detected, so developers can resolve it in seconds instead of hours. I led the design for its launch, figuring out how to introduce an AI feature to an audience that is to their code.
That skepticism is not irrational. It is the result of years of experience with tools that break things at the wrong moment. The challenge was not overcoming the distrust. It was designing something that earned its place by respecting it.
Challenge
The typical way enterprise software introduces a new feature is through announcements, onboarding modals, and tutorials. That approach does not work with developers. They dismiss anything that interrupts their flow or .
Developers have strong filters against product noise. A modal asking them to try a new AI feature in the middle of a review is not just ignored. It actively damages trust. Interruption signals that the product does not understand how they work. The bar for earning attention is not lower than for any other tool. It is higher.
The challenge was designing something that felt like a natural extension of how they already work, not an add-on they had to learn, and doing it in a way that never made them feel like the product was trying to do too much.
Process
Understanding the developer mental model
I ran generative research with developers across different seniority levels and tech stacks to understand how they evaluate new tooling and what triggers skepticism toward AI suggestions. The key finding was behavioral: despite 96% of developers not trusting AI-generated code, only 48% always verify it before committing. The gap between what developers believe and what they actually do under pressure was the real design space.
Mapping activation patterns
I audited how enterprise B2B products typically drive feature adoption, then stress-tested each pattern against developer behavior. Most failed: too interruptive, too abstract, or front-loading commitment before value. The insight was that developers do not explore new features, they encounter them. The design had to be present at the moment of need, not before it.
Designing progressive disclosure
I prototyped multiple activation variations and tested them with developers, iterating on wording, placement, and visual hierarchy. The principle behind every iteration was the same: each step could only ask for as much trust as the user had already earned. Nothing was revealed before it was relevant.
Building shared design foundations
Running this alongside the Remediation Agent meant two teams could easily diverge. I created shared design guidelines and reusable components covering AI suggestion states, confidence indicators, and opt-out affordances. This gave both products a consistent language and prevented months of duplicated design work.
Solution
Developers do not discover new tools through marketing. They discover them at the exact moment they hit a problem. So instead of onboarding flows or feature announcements, CodeFix surfaces in one place only: when an issue is flagged, right where the developer is already looking.
The activation is built around a single first step: fix one issue, see it work, then decide if you want more. No broad commitment before experiencing value. Once a fix is generated, applying it takes , so the whole experience feels like one action rather than two separate tools.
IDE integration was not a nice-to-have. It was the difference between CodeFix feeling like part of the workflow and feeling like an interruption. If applying a fix required leaving the editor and navigating somewhere else, the cognitive cost would be high enough that most developers would skip it. One click means the experience lives entirely inside the flow they are already in.
Consistency across AI features mattered for the same reason trust does: developers notice when the same company's products make different promises about how AI behaves. I built shared design guidelines and components across CodeFix and the Remediation Agent so that the language of AI transparency stayed coherent no matter which surface a developer was using.
Presentation
Use arrow keys to navigate
1 / 3
Impact
4,000
engineering teams activated
1,000+
fixes generated every day at peak
32%
still using it weekly months after first activation
4,000 engineering teams activated. Over 1,000 fixes generated every day at peak. 32% still using it weekly months after first activation. The thing that took 30 minutes per issue became a single click. That is the gap that closed.
Research sources
In the media
Ask the AI
Next Project
Hubli →