We set a goal: 40% of code written by AI.
Six weeks later, we measured the results. The number wasn't 40%. It was 65%. And when I drilled down into who wasn't following our commit tracking practice, and asked them directly - they swore they do 100% of their work with Claude Code. They can't go back. The real number was closer to 80%.
This was a pilot group of 30-40 engineers. They'd already been using GitHub Copilot for months - a path that started after a lecture I gave at Nayax back in July, before I joined as Chief AI. They wanted to start with something. They weren't AI beginners. And still - Claude Code was a leap.
One Person, One Day, Production-Ready
The story that shook everyone involved a pricing feature: a weekly grid where operators could set different tariff levels for each hour of each day. Complex UI. The kind of thing that typically needs a designer, a developer, coordination, and weeks of back-and-forth.
Instead, one engineer sat down with Claude Code. No designer. No additional developer. No product manager. Just a Jira ticket and a general idea.
Claude analyzed the existing codebase, proposed the UI structure, planned the architecture, and wrote the implementation. Ten commits over roughly 12 hours. A couple of days later, it was in production.
Was it the best design they could have come up with? Probably not. But was it bad? Definitely not. It was good enough that the engineering lead was happy to sign it off. And that's the point.
Good Enough Isn't a Compromise
"Good enough" sounds like lowering the bar. It's not.
Every task has requirements - some explicit, some implicit. When we spend weeks polishing a feature that users will interact with for three seconds, we're not being thorough. We're misallocating effort. "Good enough" means understanding what the task actually demands and meeting that bar efficiently.
The pricing grid could have gone through rounds of design reviews. It could have been pixel-perfect. Instead, it shipped in days, users could configure their tariffs, and the business moved forward. If something needed refinement later, it could be refined. But the core value was delivered.
This is the shift AI enables: not worse outcomes, but faster paths to outcomes that were always acceptable. The features we used to over-engineer out of habit? We can now ship at the quality they actually required.
Sometimes we can't compromise - and that's fine. When the bar is high, we know it. We specify it clearly, we push the agent harder, and we get something that meets that higher bar. But for everything else, "good enough" is exactly what it sounds like: good. Enough.
Why They Can't Go Back
The engineers in this pilot had been on GitHub Copilot for months. They weren't new to AI coding agents. So what changed?
It's not one thing. It's success rate. With Claude Code, they accomplish more. Their tasks succeed more often. The speed is there. And the full 200K context window (compared to 128K on GHCP) means complex, cross-file work doesn't hit walls as often, and triggers less context compaction.
There was also an unexpected benefit: Claude Code works in the terminal. That meant engineers who hated leaving their preferred IDE - JetBrains, Visual Studio, whatever - finally got full AI power without switching to VSCode. It covered more of the "dev market" by treating everyone equally.
Out of the entire pilot, only one developer moved back to GitHub Copilot. The reason? Claude ran slow on her laptop. Everyone else? They're not going back.
Power Users Aren't the Benchmark
Here's the lesson I almost missed.
Those of us who've been deep in AI tooling for the past year - we're not the benchmark. We see the spectrum of skills in every pilot. There are engineers who design a PRD with Claude over voice chat during their commute, then implement a working simulator that same evening. And there are engineers who struggle to understand how to talk to the agent at all.
Focusing only on the top performers, thinking they represent the company, thinking we've "won" adoption - that's a mistake. A deep one.
We need to provide support for those having a harder time. Workshops. Pairing sessions. Showing them how we talk to Claude, not just what Claude can do. That's a big effort, but it's worth it. The gap between power users and everyone else is where real organizational value gets left on the table.
What Comes Next
Claude Code is now the default at Nayax. Not because of a mandate - because the results spoke. Engineers who tried it didn't want to go back. Features that would have taken weeks shipped in days. And the 40% goal we set? We blew past it without trying.
This pilot covered 30-40 engineers. Nayax has over 300 developers. Scaling Claude Code across the entire R&D organization is the next chapter - building the support infrastructure, running workshops, helping everyone cross the learning curve at their own pace. The pilot proved the model works. Now we execute.
