I've been running automated AI code review on my personal projects since around October last year. Eight months. Claude Code's /review command, wired into GitHub Actions, triggering on every PR and every iteration. At some point I stopped noticing it — which is how you know something actually works. The bugs it caught became bugs I didn't ship.
When I joined Nayax, I knew code review was going to become a bottleneck. The volume of PRs from a large developer organization, the pace AI coding was setting — human reviewers were going to get buried. I already had a solution that had been running on my own projects for months. I trusted its quality. So I brought it.
The tools I almost bought
Before deploying anything, I evaluated the market. Tools like CodeRabbit and Baz have built real businesses here — solid products, per-seat pricing in the $25 to $30 per developer per month range. For a small team, that cost barely registers. For hundreds of developers, it becomes a real budget line — the kind that ends up in a spreadsheet and gets scrutinized.
Then in March, Anthropic shipped their own official Code Review feature: a multi-agent system that runs parallel specialized agents on every PR, cross-references findings, and posts ranked comments. Genuinely more thorough than anything else I've seen. The price is $15 to $25 per review.
Now look at how that compounds. The reviewer triggers on every iteration, not just when a PR opens — a single active PR can accumulate several reviews before it merges. At $15 to $25 each, any organization shipping at real volume is looking at a serious six-figure annual line. Anthropic's official reviewer is genuinely impressive. It's also built for a different use case — one where those economics work. At our volume, they don't. That's not a critique. It's a different tool for a different question.
I went a different direction.
What we actually built
The setup at Nayax mirrors what I'd been running at home: GitHub Actions, triggering on every PR. Same command, same automation, just at a different scale.
When I say deployment was easy, I don't mean the YAML was trivial. I mean there was no procurement cycle. No vendor security review. No CISO sign-off. No legal approval. Claude was already in the house — already approved, already governed, already trusted. All we had to do was configure a CI job.
That distinction matters more than it sounds. Every new vendor is months of process at a company our size. This was an afternoon.
What it actually does
The reviewer classifies comments across four severity levels. Critical comments block the PR.
That blocking rule wasn't a default we started with. It's where the data led us. Developers don't have to act on every comment — but they have to engage with each one: fix it, or close it with a "won't fix" and an explanation. The same discipline most of us applied throughout our careers when receiving code review. Which means every critical comment gets adjudicated by a developer with full context and every incentive to push back. A quarter in, none has been overturned. That's why critical comments block the PR.
It's also connected to Jira. The integration works as well as the requirements it checks against — and at Nayax, the product team has been writing thorough specifications with AI for a while now: detailed epics, fully specified user stories, clear acceptance criteria. Claude doesn't just review the code — it verifies the requirements were actually met. That only works because the requirements are worth verifying.
The result: PRs arrive to human reviewers in a different state. Styling violations, coding standard mismatches, missing requirement coverage — handled before a human touches it. Reviewers focus on what actually needs human judgment.
The honest part
There's a complaint I hear consistently: when developers iterate on a PR, new review comments appear that could have surfaced in the first pass. That's real. It's frustrating, and I don't have a fix for it yet.
What I do have is perspective. I've spent a career doing code review. When there's a lot of noise in a diff — many changes, unclear intent — the important issues tend to surface later. That's not an AI problem. That's what review looks like.
The numbers
We're running this on over 100 repositories — the highest-traffic ones, covering roughly 80% of all code shipped at Nayax. We're approaching the end of a full quarter of deployment. The rest of the repositories are on their way.
And here's my side of the math. Our reviews run as API calls, billed in tokens, on the Claude account Nayax already pays for. Last month, the entire thing — 100+ repositories, every PR, every iteration — cost less than $1K. The same volume of reviews at $15 to $25 each would have cost at least an order of magnitude more. Not $30 per seat. Not $25 per review. Less than $1K, total.
The loop is closing
AI is writing the requirements. AI is writing the code. Now AI is reviewing it.
Human reviewers still matter — more than before, actually. They're just no longer the first line of defense against things a machine can catch. They're the final judgment on things a machine can't.
That's the shift. Not a tool decision. A rethinking of where human attention belongs in the development process.
Default on.
