What Our Office AI Jukebox Taught Us
We built an AI-powered office jukebox just for fun. It turned into one of the better lessons we've had about working with AI — and about the kind of team we are.
Every office has its small rituals. Ours, for a while, was a quiet skirmish over the speaker in the corner. Somebody would queue up a playlist, somebody else would skip it, and a third person would wander over and quietly take the whole thing in a different direction. Music in a shared space is a tiny, surprisingly emotional negotiation, and we were not winning it.
So one afternoon, mostly for fun, a couple of us started building a better referee. We called it Jukebox — an AI-powered office jukebox, our own little in-house version of the “what should we play next” problem. The idea was simple: let the team make requests in plain language, let some AI sort out the vibe, and let the music keep rolling without anyone having to babysit it.
It started as a joke and turned into a habit
Here’s the thing about building something silly: you learn just as much as you do building something serious. Maybe more, because nobody’s stressed and everybody has opinions.
Jukebox got opinions immediately. People typed in moods instead of song titles — “something for a Monday,” “we just closed a big ticket, celebrate.” Watching the AI try to make sense of that was genuinely funny, and then genuinely useful. The first version got the energy wrong about a third of the time. The fixes weren’t magic; they were the boring, real work of paying attention to what people actually meant versus what they typed.
That’s lesson one, and it’s the one we keep relearning: AI is only as good as how carefully you frame the problem. A vague request gets a vague answer. The moment we got specific about what “good” looked like, the whole thing got better.
What a toy taught us about real work
A few things stuck with us long after the novelty wore off.
Start small and let people poke at it. We didn’t write a spec. We shipped something rough into the office and let the whole team break it. The feedback we got from five people using a thing for real beat any plan we could have drawn up on a whiteboard.
The unglamorous parts are the actual work. The fun idea took an afternoon. Making it reliable — handling weird inputs, not playing the same three songs on a loop, failing gracefully when something hiccuped — took the rest of the time. That ratio is true of basically every AI project worth doing.
Keep a human in the loop. Jukebox suggests; people still skip. Nobody wanted a robot dictating the office soundtrack, and honestly, that’s the right instinct for a lot more than music. The good version of AI hands you better options faster — it doesn’t take the decision away from you.
Trust comes from living with it. We didn’t believe Jukebox was any good because someone demoed it. We believed it because we used it every day for weeks and it kept being right more often than not. There’s no shortcut for that.
The kind of team we are
We’re not telling you this because the office jukebox is changing the world. It absolutely is not. We’re telling you because it says something about how we work: we’re the kind of people who, when we get curious about a tool, build a little thing on a Friday and learn out loud from it.
That curiosity is the whole point. The same instinct that makes you build a fun office gadget — what if we tried it, what does it actually do, where does it break — is exactly the instinct you want pointed at the problems that actually matter. Internally, that’s a jukebox. For the businesses we work with, it’s the repetitive, time-eating work that quietly slows a team down. Same habit of mind, much higher stakes.
If you’re curious how that builder mindset shows up in the real work we do, here’s a closer look at how we put AI to work for our clients. And if you’re the kind of person who’d have wandered over to mess with the jukebox on day one — we’d probably get along just fine.
What Our Office AI Jukebox Taught Us
We built an AI-powered office jukebox just for fun. It turned into one of the better lessons we've had about working with AI — and about the kind of team we are.
Every office has its small rituals. Ours, for a while, was a quiet skirmish over the speaker in the corner. Somebody would queue up a playlist, somebody else would skip it, and a third person would wander over and quietly take the whole thing in a different direction. Music in a shared space is a tiny, surprisingly emotional negotiation, and we were not winning it.
So one afternoon, mostly for fun, a couple of us started building a better referee. We called it Jukebox — an AI-powered office jukebox, our own little in-house version of the “what should we play next” problem. The idea was simple: let the team make requests in plain language, let some AI sort out the vibe, and let the music keep rolling without anyone having to babysit it.
It started as a joke and turned into a habit
Here’s the thing about building something silly: you learn just as much as you do building something serious. Maybe more, because nobody’s stressed and everybody has opinions.
Jukebox got opinions immediately. People typed in moods instead of song titles — “something for a Monday,” “we just closed a big ticket, celebrate.” Watching the AI try to make sense of that was genuinely funny, and then genuinely useful. The first version got the energy wrong about a third of the time. The fixes weren’t magic; they were the boring, real work of paying attention to what people actually meant versus what they typed.
That’s lesson one, and it’s the one we keep relearning: AI is only as good as how carefully you frame the problem. A vague request gets a vague answer. The moment we got specific about what “good” looked like, the whole thing got better.
What a toy taught us about real work
A few things stuck with us long after the novelty wore off.
Start small and let people poke at it. We didn’t write a spec. We shipped something rough into the office and let the whole team break it. The feedback we got from five people using a thing for real beat any plan we could have drawn up on a whiteboard.
The unglamorous parts are the actual work. The fun idea took an afternoon. Making it reliable — handling weird inputs, not playing the same three songs on a loop, failing gracefully when something hiccuped — took the rest of the time. That ratio is true of basically every AI project worth doing.
Keep a human in the loop. Jukebox suggests; people still skip. Nobody wanted a robot dictating the office soundtrack, and honestly, that’s the right instinct for a lot more than music. The good version of AI hands you better options faster — it doesn’t take the decision away from you.
Trust comes from living with it. We didn’t believe Jukebox was any good because someone demoed it. We believed it because we used it every day for weeks and it kept being right more often than not. There’s no shortcut for that.
The kind of team we are
We’re not telling you this because the office jukebox is changing the world. It absolutely is not. We’re telling you because it says something about how we work: we’re the kind of people who, when we get curious about a tool, build a little thing on a Friday and learn out loud from it.
That curiosity is the whole point. The same instinct that makes you build a fun office gadget — what if we tried it, what does it actually do, where does it break — is exactly the instinct you want pointed at the problems that actually matter. Internally, that’s a jukebox. For the businesses we work with, it’s the repetitive, time-eating work that quietly slows a team down. Same habit of mind, much higher stakes.
If you’re curious how that builder mindset shows up in the real work we do, here’s a closer look at how we put AI to work for our clients. And if you’re the kind of person who’d have wandered over to mess with the jukebox on day one — we’d probably get along just fine.