The bar for introducing new tooling is high at Amiqus, and it should be. When you're handling sensitive data for clients every day — background checks, identity verification, onboarding workflows — "move fast and break things" isn't a strategy available to you.
So when we decided to evaluate Claude Code, we didn't treat it like a typical productivity tool. We treated it like any other change that could affect our clients, our engineers, or the integrity of our systems: carefully.
How we started
We started small. A handful of engineers, clear guardrails, nothing near production data. The goal wasn't speed. It was confidence. Could we adopt this responsibly, in a way that fit who we are?
The guardrails weren't complicated, but they were intentional. We defined what data could go into the model — code, internal documentation, general engineering questions, not customer data. We agreed that any AI-generated output touching security-relevant or customer-facing code needed a human review before it went anywhere near a pull request. And we gave the trial cohort explicit permission to say it wasn't helping them. No pressure to perform enthusiasm we hadn't earned yet.
Eight weeks to full rollout
Eight weeks later, the whole engineering function was onboarded and Claude Code was formally accepted into the business.
That progression didn't happen because the tool was obviously great from day one. It happened because each week, the small group of engineers using it built up a body of real experience — moments where it genuinely saved time, cases where it didn't, and a shared sense of what it was actually good at. That experience became the foundation for the broader rollout.
Formal acceptance into the business required more than engineering sign-off. It needed to satisfy the same bar as any other tool that touches our systems: a clear data processing position, a defined scope of use, and evidence that the people using it understood both the capability and the limits. That work wasn't glamorous, but it's what made the rollout feel solid rather than rushed.
What changed
Once the trial was complete and the tool was fully embedded, four things had shifted in ways we hadn't fully predicted.
Knowledge stopped living in people's heads.Amiqus has brilliant engineers with deep domain expertise. That's also a single point of failure. Claude can now traverse our repos, Notion, and Slack and surface answers that previously only existed inside specific people. For a team of our size, that's a meaningful shift in resilience.
Confidence went up. We didn't set out to measure this one, but engineers started voicing greater certainty in their work. Not overconfidence — something more like the ease that comes from having a capable collaborator to sense-check your thinking. It was hard to fake and easy to notice.
Spec-driven planning shifted. A strong, product-minded engineer can now plan features at a depth that used to need a much bigger room. The ceiling on what one person can contribute quietly moved. We're still early in understanding how far that goes, but the direction is clear.
The barrier to building dropped. Designers are building. PMs are shipping internal tooling. Prototypes are becoming real things. This is the one that surprised me most — the skill floor for going from idea to working implementation dropped enough that people who wouldn't have touched a codebase six months ago are now contributing meaningfully.
What it actually costs
Honestly, AI is the most exciting thing to happen to software engineering in my career.
It's also exhausting. And I don't think we talk about that enough.
In the last six months I've gone from nearly hands-off the code to properly back in it. That's been brilliant for me, personally. But it's also taken a toll — and I'm someone who wanted this.
For engineers who didn't ask for it, or who are trying to absorb it alongside normal delivery pressure, the cognitive load is real. Learning something this fundamental doesn't happen in the background. It competes for the same headspace as everything else — as the sprint work, the code reviews, the architectural decisions that still need to be made by humans.
The cautious rollout we ran at Amiqus wasn't just about compliance and data boundaries. It was about giving the humans doing the learning some space to breathe. An eight-month rollout across an engineering function of around twenty — starting with five engineers in the pilot — might look slow from the outside. From the inside, it was about right.
We owe it to the people doing this work to be honest about what we're asking of them.
Safety as foundation, not obstacle
None of what we found happened because we moved fast. It happened because we treated safety as the foundation.
If you're an engineering leader in a regulated industry weighing something similar, I'm happy to talk through how we structured the trial, what the guardrails looked like in practice, and what I'd do differently next time.