How does a company turn AI buzz into real impact? By letting everyone build, learn, and share in the journey.
- Checkout.com is integrating AI in all aspects of their operations, not just product delivery.
- Engineering velocity increased, with a 56% rise in Pull Requests throughput from late 2025 to early 2026.
- 95% of pull requests now involve AI, significantly enhancing coding efficiency through their platform, Agent HAL.
- An internal deployment environment, Arrakis, has enabled employees to create over 160 unique AI tools in just three weeks.
- Emphasis on building a culture of experimentation and learning around AI adoption across the organization.
Over the past decade, we have built payment infrastructure that some of the world’s largest companies depend on. Machine Learning, and more recently AI, have always been at the heart of what we ship to our customers.
But like many others right now, we are going all-in on AI, not just in the products we ship, but in every corner of how we work.
What you tend to hear from companies on this topic is often declarative: we have embraced AI. What you hear less of is the honest, in-progress version. What it actually looks like to build an AI strategy in real time, while the tools are still evolving, best practices are still forming, and the results are still emerging.
We will be sharing what we are learning: what is working, what we are struggling with, and what we are genuinely uncertain about. We hope it starts a useful conversation.
This is our first installment. Here is where we are today.
Table of Contents
Setting the scene
For over a decade, our ML and AI has optimized payments for the world’s largest merchants. Our Intelligent Acceptance platform has unlocked over $20.7 billion in merchant revenue – running 26,000 optimizations per minute, each one a signal feeding the next. Every decision compounds, and the right system finds the next improvement faster than any human team working alone.
We are now applying that same logic to ourselves. The same instinct, redirected inward.
Today, nearly all of our employees are actively using AI tools on a regular basis. We have trialled different models and tools, learning about their strengths and weaknesses, and landing on our current preferred toolset for developers. Models are always changing in capability, so we adapt accordingly and select those that are best-suited for the task, be it writing, coding, or data analysis. We are still early in this journey, but the trajectory is clear. AI is speeding up production, review, and workflows to realize benefits over weeks, rather than months.
Building faster
The clearest evidence so far is in our engineering velocity. Between October 2025 and April 2026, our Merged Pull Request (PRs) throughput grew by 56%. More telling is that over the last three months per-engineer output grew by 32%. This is compounding in real time. Needless to say, our top engineers are showing significantly bigger velocity gains.
The driver is visible in the underlying data. In September 2025, roughly 10% of our pull requests (PRs) had any AI involvement. By March 2026, that figure was 95%. The inflection point arrived with the advancement in LLM coding capabilities (Opus 4.6) and Agent HAL (Yes, that HAL), our in-house agentic platform.
HAL monitors engineering backlogs, writes code, and opens pull requests for human review. It does not merge its own code; human judgment stays in the loop. We believe this is important given where we are in the journey, and given the nature of what we do: a highly reliable, critical system that our merchants around the world depend on. As we build confidence, we will evolve the guardrails accordingly. We strongly believe that as more engineers adopt HAL, the shape of the velocity curve will continue to bend.
We are also seeing a meaningful shift in how design and engineering interact. One of our front-end-heavy teams connected Figma directly to their React Native codebase, enabling designers to generate production-ready code without the usual engineering handovers. Engineers shifted entirely into a review role. Features that once required multiple back-and-forth cycles between design and engineering now move from mockup to production in a fraction of the time, with live device previews replacing lengthy design QA cycles.
Beyond throughput, we are starting to see the signals that really matter. A small team delivered a customer-facing feature end-to-end in two months instead of four. Integrations that used to take three to four weeks of engineering effort are now built and tested in a matter of hours from a single prompt. These are early signals, but they are the ones we care about most. PR throughput is a good leading indicator, but customer impact is the goal.
And since customer impact is the goal, engineering velocity alone is not enough. If engineers move ten times faster and the processes around them do not, product managers and cross-functional workflows become the bottleneck. That is the next unlock we are actively building for: AI-assisted requirements definition that feeds our coding agents and AI-embedded legal, compliance, and finance review that reduces the organizational back-and-forth that slows everything down.
Working smarter
The second dimension sits beyond the software development lifecycle.
We want every team, not just those whose day-to-day job is to produce code, to go beyond question-and-answer mode and build with AI. To that end, we have built Arrakis: an internal deployment environment (think of it as our own version of Lovable) where any employee can ship an AI-powered productivity tool that meets our security guardrails, without a lengthy review process. The results surprised even the most optimistic among us. In just 3 short weeks, our early adopters built over 160 unique apps that ranged from personal productivity to workflow efficiency. A centralized skills repository holds hundreds of reusable artefacts: a skill built by one team becomes available to everyone. The loop is starting to form: build, share, inspire, build more.
The outcomes are already spreading across functions. A Finance tool handling a payment query whose volume is projected to grow 300%, saving thousands of hours of manual processing annually, without adding headcount. Account Management teams with instant merchant performance visibility before every client call, recovering hours of preparation per interaction. Commercial outreach that previously required significant manual research, now takes seconds. These are just a handful of examples.
Running in parallel is Project Horus, our enterprise knowledge layer. By indexing and connecting data across Confluence, Jira, BigQuery, and Salesforce, we are building an AI that understands our codebase, our merchant relationships, and our operational context. The goal is to shorten the iteration cycle for both our people and our AI agents.
Performing better
Intelligent Acceptance ran 10.5 billion optimizations in 2025 and recovered $11.2 billion in merchant revenue that year alone, more than eight times what we had accumulated in all prior years combined. That is not just a better model. It is a compounding system: more transactions generate cleaner signals, cleaner signals produce smarter decisions, smarter decisions attract more volume. The loop has been running for years, and it accelerates.
What operating at this scale teaches you is that the optimal payment strategy is deeply local. An authentication approach that adds two percentage points for Visa in the US will destroy performance for Mastercard in Germany. A network token that lifts recurring transactions behaves differently for a streaming service than for a SaaS platform. Payment optimization is won in the intersections, not the averages.
Across our portfolio, Intelligent Acceptance delivers an average uplift of 3.8 percentage points — roughly $38 million in recovered revenue every year for a merchant processing $1 billion annually.
Expertise, though, has a ceiling. Even our best payment specialists can actively manage a finite number of optimizations at any one time. We regularly run over 200 simultaneous experiments at any given moment in time, each targeting a specific issuer-market-scheme combination. The long tail of possible intersections is far larger than any human team can cover alone. That is where our Performance and Fraud agent layer becomes a force multiplier. These agents do not replace our payment experts; they extend their reach. Every proposal an agent surfaces comes with a full picture: current performance, expected impact, fraud implications, and the gate criteria for going live. A payment expert reviews and approves. Experiments graduate through statistical gates. Nothing ships without human judgment. The result: the same expertise we have built across issuers, schemes, and markets for the world’s largest merchants can now generalize further and faster than any organization of people ever could. The merchant that once needed a dedicated payment or fraud specialist now gets access to the same playbook that the largest platforms in the world run on. That is not a distant roadmap, it is being tested now, and early results are encouraging.
Building a culture of builders
None of the above happens without the culture to sustain it. Getting a team of thousands genuinely fluent with these tools is a deliberate effort. We are investing in it across several dimensions, such as:
- dedicated AI training programmes,
- Slack channels where teams share what they have built,
- and AI-focused hackathons and product jams.
We have developed an AI Centre of Excellence that accelerates adoption and fields questions from across departments, and we showcase demos at our regular All Hands to spotlight what people have shipped. We have also raised the bar in hiring, with AI proficiency interviews now being piloted across several functions.
About 15% of our team members building with AI today are non-engineers. But, when we encouraged teams to build tools that improved their own productivity, we saw a real pattern: some people hit a wall. Not everyone across the company knew where to start, let alone how to build a working vibe-coded solution. Training helped, but the barrier of entry was still too high for some. That is where our AI Centre of Excellence stepped in with something more direct: a rapid-build programme where any team with a use case gets paired with a dedicated developer for a focused one-week co-build sprint. The hope is that tandem-vibe-coding helps teams pick up the skills to keep evolving the solution themselves – and tackle new ones. Sixteen use cases are active today, with dozens of more projects in the pipeline.
The most important lesson from all of this: doing beats reading. The leaders across our organisation who have made the most progress are the ones who stopped delegating their AI curiosity and started building things themselves, however rough, however imperfect. There is no substitute for the first time you experience a result that would have taken hours, delivered in seconds. That moment changes the frame entirely. No training programme or strategy document produces it. Getting people to their own version of that moment, quickly, is one of the most valuable things we can do as a leadership team.
What we are taking away from this
If there is a common thread running through everything above, it is that AI does not transform an organization on its own – the decisions around it do. No amount of reading or training substitutes for the moment you experience it yourself. The leaders who have moved furthest here are the ones who picked something up and built with it.
In regulated fintech, shipping faster means moving faster across every function — legal, compliance, scheme partnerships, operations, product — so that the work feeding engineering keeps pace with what engineering can build. Therefore, when we think about product velocity, we have to measure it on an organizational level.
These are the questions we are sitting with, and we will keep sharing them as we go. In the weeks ahead, we will be exploring how AI is changing the way we define and ship products, how our payment experts are working alongside agents, and what it actually takes to build a culture where everyone — not just engineers — becomes a builder.
We have spent years building systems that compound for our merchants. We are now building the same for ourselves.
At Checkout.com, there is no finish line. No such thing as done, or perfect — we are constantly getting better. With AI, that is more true than ever. We are just getting started.