How Satispay Generates 75% of Its Code with Claude

Satispay entered 2025 with an engineering team weighted toward earlier-career engineers and a mature payments codebase that had accumulated legacy services difficult to navigate without deep context. When CTO Fabio Rapposelli joined, the default workflow showed the strain: an engineer would pick up a ticket, spend significant time orienting in an unfamiliar service, draft a change, then wait for a senior engineer whose review was carrying three jobs at once — teaching, quality gate, and context transfer. Roadmap velocity suffered, and senior time was going to enabling others rather than building.

Previous experiments with autocomplete-style AI coding tools had stayed in pockets and never addressed the orientation gap. Satispay ran a structured 30-day evaluation in mid-2025 comparing Claude Code against another AI coding tool on real tickets across its Java and Spring services. The rubric covered code generation quality, refactoring, adherence to internal patterns, multi-file context understanding, test generation, and documentation. Claude won on every dimension except IDE integration. Rollout took 30 days from decision to full coverage, handled by IT support, with Claude Code shipped to every engineering laptop through the managed device fleet.

The team built on top of the default tooling rather than just deploying it. Standard Java scaffolding starts from encoded patterns rather than blank files. Modernization approaches get packaged into subagents so the next team inherits the work. The most concrete example is data transformation: AWS Lambda functions that used to take three to five days to build by hand are now generated from a specification by an internal subagent connected to an MCP server in under an hour. Engineers use full access to the 1M-token context window, reaching for Opus or Sonnet depending on the task, with spend managed at the budget level so no one pauses mid-task to justify a more capable model.

The clearest single proof point came from the core transaction service — the system that runs every payment on the network. The team upgraded it from Java 8 to Java 21 and through a major Spring version in under four days, against an original estimate of four weeks. At the broader scale, goals set for an 18-month horizon were complete in seven months, fast enough that an external consulting engagement brought in for the work was no longer needed. Data transformation Lambdas compressed from three to five days to under an hour.

Adoption reached above 90% of engineers and then spread beyond engineering: Claude Cowork requests came in at roughly 20 per day from finance, legal, marketing, and operations — including the CFO asking for access to financial analysis skills he had spotted himself. Satispay is now planning to ship roughly 50% more features in the first half of the year and is running two agentic pilots: issue-to-PR automation where Claude drafts pull requests from tickets filed by other teams, and agentic fraud investigation where a workflow tied directly to financial loss moves from human-in-every-step to agent-led.