Situation
S1 had made GitHub Copilot available to some developers in 2024, with little adoption. In late April 2025 the CEO decided S1 would move to Windsurf and drive real adoption, owned by the AI org. It was treated as a top-priority program, and I was called in. I came to it with hands-on use of the AI coding tools of the day: Cursor, Cline, Roo Code, Windsurf, Copilot, bolt.new and bolt.diy.
Objective
Acquire and onboard Windsurf fast, get it into daily use across the company, keep the token budget under control, and measure the productivity effect instead of asserting it.
Approach
- Onboarding in under a month. Onboarding a new vendor at S1 typically took 3 to 6 months. As the CEO initiative’s assigned driver, I engaged the Windsurf team directly and pushed 30+ people across finance, compliance and governance, IT, engineering, procurement and security, up to the CISO, to move fast. Acquisition and onboarding were complete by May 27, and engineer onboarding began in early June.
- The rollout machine. We had about 800 seats, so access was gated by request. I wrote the onboarding process, the mandatory training required before a seat was granted, the tracking, the invitation flow and the approval flow, then engaged the teams directly. Once a user’s name was submitted, they got an invitation to complete a survey. When they finished it, if their manager approved and they were in the right groups, we onboarded them, and they got an email on how to log in and get started, with pointers to support in Slack and on the web. June went to training and onboarding.
- A forcing function. In July, with the CTO’s backing, S1 mandated 90%+ unit-test coverage for all code. We spent the month showing engineers who doubted the tool, live, how much faster Windsurf got them to coverage.
- The opener. Many engineers doubted Windsurf’s value and worried it was there to replace them. I opened group trainings with my own history: learning to code before IDEs existed, writing C in vi, then the graphical IDE, then integrated debuggers. Framed that way, Windsurf was the next step in developer tooling, a sharper instrument for the same craftsman.
- Beyond engineering. I led MCP adoption and MCP creation. Windsurf with MCPs proved potent for non-engineering work, so I started training for Product Management, IT, SecOps and Sales Engineering.
- Token governance. Each developer had a token budget backed by a shared pool of one-time, non-expiring overage tokens. Engineers reaching for frontier models would stall at their limits or burn 10x their allocation and drain the pool. I built usage reporting and tracking, automated outreach to over-consumers, and projected overage-credit burn, so the purchase-order process had time to buy more credits before anyone got blocked.
- Ownership and instrumentation. I was the sole Windsurf app owner: licenses, credit monitoring, support, onboarding docs, and coordination with the vendor TAM on roadmap features such as Devin and the CLI. I built the Windsurf usage-data collectors, a Kestra-based framework pulling prompt-credit, lines-of-code and acceptance-rate metrics from the Windsurf admin API into the S1 data lake. Training ran with the Coralys training team and the AI Champions program, through office hours and small-group sessions.
- Measuring the gain. I asked engineers to put their estimate to complete each issue without AI into Jira’s “Est. Story Points” field, and let “Story Points” reflect the actual points burned at close. That gave a per-issue with-AI and without-AI comparison.
Net results
- Procurement compressed from S1’s typical multi-month cycle to under a month.
- Uptake climbed to nearly the full seat allocation, with millions of lines of AI-generated code accepted over the following months.
- Within weeks, the org hit the mandated unit-test-coverage bar across most of the codebase, including the monorepo of an acquired company.
- GitHub change metrics shifted. Before Windsurf, nearly all code changes were additive; after rollout, the mix moved to a far more balanced blend of additions and deletions. Windsurf was driving substantial refactoring and a net flattening of total non-comment lines of code: the codebase was growing faster and getting leaner at once.
- The first productivity read looked like a decline, because story points per engineer per sprint fell. The data showed the opposite: issues resolved per sprint rose substantially for AI-enabled developers, and against their own non-AI estimates they were realizing a large multiple on delivery productivity. The point drop was an artifact of estimation.
- The clearest adoption signal: engineers who hit token exhaustion and had to go back to coding by hand. Once people felt the loss, the argument about the tool’s value was over.
Takeaway: instrument the counterfactual, or you will read the gain backwards.
Scope and technologies
Windsurf, by Codeium, later Cognition: acquisition, onboarding, rollout and sole app ownership; MCP adoption and creation; frontier LLMs; GitHub change-metrics analysis of additions, deletions and net lines of code; unit-test-coverage tooling; a homemade token-usage reporting, outreach and overage-projection layer; Kestra-based usage-data collectors into the S1 data lake; Jira with-AI and without-AI estimation; the onboarding, training and approval workflow, with Coralys and AI Champions.