SoftNet: from AI experiments to an organizational capability and competitive advantage.
A software house serving 100+ banks. One year, told from the articles we published along the way and the talk we gave together.
January 2026 · People, Process
A 2-day AI-augmented workshop on a clone of SoftNet's stack: 6 teams, 83 stories, 315 Specs By Example. Speed showed up after the guardrails existed, not before.
Read the article: AI-First Development Isn't About AI
April 2026 · Process
A real financial feature from PRD to production in 2 days. The honest part: review discipline decayed once the team got comfortable with AI output.
Read the article: From PRD to Production in Two Days
June 2026 · Structure, Strategy
The first H³ cycle with 15 to 20 of 50 to 60 developers. 4 outcomes left with named owners, among them a skill-evaluation system on main, the first AI guidelines repo and automated E2E tests in CI/CD. The COO committed code the morning after.
Read the article: The Hard Part of Enterprise AI Was Never the Models
September 2026 · Process, Rewards
After the summer: support adopted AI inside its own tools, E2E tests on the core system went from close to zero to 50+, the skill-eval check moved into CI/CD, and the guidelines repo was killed with reason. A rule nobody runs is a wish. Outcome measurement for milestones went from 0 in May to 100% in September.
Read the article: The AI Rules That Survived the Summer
October 2026 · On stage
At OrgDesignConf 2026 with COO Adam Piekarczyk, the year told through Galbraith's Star Model.
The base came first. LeSS gave SoftNet feature teams, then mob programming, TDD and agreed test levels, before any AI. AI speeds up the good and the bad alike.
Then the initiatives spread on their own: AI code and test reviews in the build pipelines, comments only at first and blocking today, and AI architecture reviews that move new code towards the target architecture. Support adopted AI in its own tools, and infrastructure now hosts a model for sensitive data. SoftNet is hiring, and AI capability is part of the interviews.
By the second H³ cycle, initiatives started to run into each other on the same problems: domain language and customer support. Leaders set boundaries, not topics, and stayed visible.
We analysed the merge requests with a deterministic check: output up 42%. Output is not value, so every milestone now measures its outcome: from 0 in May to 100% in September. That changes how financial controlling is done.
The strategy moved too. End-to-end tests on a very old system, too expensive before AI, now make modernizing it simpler. And SoftNet is opening its core system to partners through APIs: a feedback loop larger competitors can't buy.
See the slides: AI Augmented Orgs in Action
What I built alongside the teams
The first skill and agent designs, a central skills marketplace, an eval system that checks skills don't degrade over time, and a skill improver driven by merge-request comments. SoftNet's teams own and run them.
How I work with AI: the patterns behind this work.
Next
The next cycle starts on 14 December 2026.