We started this company a year ago this week. We had a hypothesis: AI was going to make most builders into people who could not, on their own, ship the thing they had built; a software engineer one click away was the missing piece; the press, as a category, would emerge whether we built it or somebody else did.
A year in, three things turned out as expected. Three turned out differently. One we got wrong about how the work would feel.
What turned out as expected
The press is the unit. Customers do not want a contract, or a retainer, or a freelancer rotation. They want the green dot. They want it now. Volume per customer is higher than we modeled; ARPU compounds with usage faster than we expected because every press is a positive proof point that triggers another. The product is the right shape.
The bench scales. We worried about hiring software engineers fast enough. The bench is bigger than we predicted at this milestone, and the hire-to-bar ratio has held. We are not bottlenecked on engineers. We are bottlenecked on integrations.
Customers stay. Our churn is lower than every comparable B2B services company we benchmarked against. The relationship the product creates, the same engineer, across phases, is a retention mechanism we under-weighted in the model.
What turned out differently
The customer isn’t who we thought. We expected the product to be discovered by individual builders inside small companies. The first six months were the opposite: enterprise teams with internal-tools functions, mid-market growth teams, clinical-operations groups inside health systems. The non-engineering builder inside a large company turned out to be the beachhead, not the indie hacker. We rebuilt our messaging around that and shipped For Enterprise earlier than planned.
Healthcare moved fastest. We did not plan for HIPAA in year one. The market told us otherwise. We trained a sub-bench, signed BAAs, wrote a white paper, and now have HIPAA-segmented coverage in production. The economics work. We will publish more on this.
The integration is the edge. When we started, we thought integration with AI tools was a feature. It turned out to be the product. The press is only useful if it lands an engineer with context. Cursor, Claude, Replit, and Lovable each took weeks of partnership work and unlocked step-changes in customer outcomes. We treat integration depth as the central engineering investment now, not a peripheral one.
What we got wrong about how the work feels
We thought the press, on the engineer’s end, would feel like on-call. Asynchronous. Punctuated. It does not. It feels like a conversation. Engineers join a session and stay for forty minutes, not five, because the customer wants company through the build. The engineering team named this internally and we adopted it: not the press, the relay. The customer is running a leg, the engineer is running the next leg, and the baton is the build. We had been modeling something more transactional and the product had to be re-shaped to match what was actually happening.
We thought we were building an interrupt. We turned out to be building a conversation that comes in second.
What we’re doing in year two
We are organizing year two around three things. One: the enterprise rails, SSO, SAML, audit trails, the full compliance envelope, so a large platform team can adopt Relay without a six-month security review. Two: integration depth on five more AI builders, picked by customer demand. Three: growing the bench so a press is always answered, whatever hour you press it. We will probably miss one of these three. The other two will land.
Year one was a hypothesis-test. Year two is operations. The work gets less interesting and the company gets more durable. That seems like the right trade.
The founders, May 2026, on Relay’s first birthday.