Between February and April 2026, beta customers pressed for an engineer 4,217 times. Each press is a moment a person building software with an AI decided that the next step was easier with a human in the room. We labeled every session by what happened in the seconds before the press, and again by what happened in the thirty minutes after. Four categories cover 92% of the sample.
One. The confidence wall (38%)
The build technically works. The builder has clicked through it ten times. They cannot bring themselves to ship it because they do not know if it will hold. They press for a person who will read it once and answer one question: can I send this to a customer? Average session length: 14 minutes. Resolution rate: 96%.
Two. The integration cliff (29%)
The AI produced a complete, runnable artifact. Connecting it to anything outside the sandbox, a database, an auth provider, a payment gateway, an email service, fails repeatedly in non-obvious ways. The builder presses. Average session length: 41 minutes. Resolution rate: 89%.
Three. The deploy moment (17%)
The build runs locally. The builder has a domain. They have not deployed software before. They press because the gap between “works on my machine” and “works in production” turns out to be most of the work. Average session length: 1h 12min. Resolution rate: 94%.
Four. The ownership question (8%)
The build is shipped. Something has changed in the world, a payment failed, a user reported a bug, a third-party API silently changed shape. The builder presses because the build has stopped being a build and started being a system, and systems want a person to own them. Resolution rate: 99%; mean-time-to-resolve: 23 minutes; relationships continued: 71%.
The 8% that doesn’t fit
About one in twelve presses we couldn’t cleanly classify. They’re mostly conversations. The build is fine; the builder wants a second opinion on a decision that isn’t really a code decision (a vendor, an architecture choice, whether to rewrite). We think this is its own category and will keep watching it.
The takeaway, for us, is that the press isn’t random. It clusters around four legible moments, and the fastest path to making AI builds reliable is to make the press cheap, instant, and obvious at exactly those four moments. That’s the design of the desktop, and increasingly the design of every integration.
Relay Research, with help from the engineering team. Anonymized, aggregated. Methodology note available on request.