Why We Built ArrangementLabs.ai

A few years ago I sat at a desk with a piano arrangement of a Bach prelude and a string quartet I was supposed to hand to a youth ensemble by the end of the week. I’m a musician — I trained in piano performance — and I’ve spent enough hours arranging that I know how the process goes. You sit with the score. You decide who plays what. You write each part out. You play it back in your head. You catch the mistakes. You revise. By the time the quartet has something readable, three hours of your evening are gone.

That night I did what I’d done a dozen times before: I opened my notation software and started copying. And somewhere around the second page, I asked the question that doesn’t go away once you’ve asked it. Why am I doing this by hand in 2024? Why is there no tool that handles the mechanical 80% of this work — the obvious voicing decisions, the range checks, the bar-by-bar redistribution — and leaves me with the 20% that actually needs my judgment?

I looked. The tools I found either didn’t exist or didn’t work. Music technology had solved transcription. It had not solved the redistribution problem. AI demos that produced quartet-shaped output. None of them produced something I’d hand to a player without rewriting it from scratch. The gap between “AI music tool you can demo on a podcast” and “AI music tool a working arranger keeps open on a deadline” was enormous.

That gap is the reason ArrangementLabs.ai exists.

What I believed when I started

I had a few convictions going in that turned out to matter.

The first: voice redistribution is a real craft. Not a content-generation problem. Not a style-transfer problem. A craft, with centuries of accumulated practice and a body of conventions that work. The conventions are well documented in any orchestration textbook. If AI was going to help with arrangement, it had to learn the actual craft, not produce something that looked vaguely arrangement-shaped.

The second: arrangers are not going away. The interesting future for AI in music is not autonomous composition. It’s tools that working musicians use to compress the mechanical parts of their workflow so they can spend more of their attention on the parts that need it. The framing matters: AI as draft assistant, not AI as replacement. Every working arranger I’ve talked to wants the first kind of tool. None of them want the second.

The third: the tools that will survive this decade are the ones built by people who actually use them. Not by ML researchers who want a music application for their model. Not by VCs who want a music vertical. By musicians who needed the tool and built it because nobody else would. I wanted to build for myself first and trust that other musicians would recognize the difference.

The fourth: the bar for “good arrangement” is set by working professionals, not by the people who post AI demos on social media. If the tool didn’t produce something a real arranger would use, it didn’t matter how impressive the demo looked.

These convictions shaped everything that came next. They’re still the reason ArrangementLabs.ai is built the way it is.

What we built

ArrangementLabs.ai is a tool that takes a score in any of the formats you’d reasonably have (MIDI, ABC, MusicXML, MXL) and produces an arrangement for any of a growing set of target ensembles. The user provides the music. The tool helps redistribute it.

It’s not a transcription tool. It doesn’t listen to audio and produce notation; there are good tools for that and we don’t need to add another.

It’s not a generation tool. It doesn’t compose music from prompts; that’s a different problem and a different market and we don’t want to be in either.

It’s an arrangement tool. The user has the music. The user knows what ensemble they want. The tool’s job is to handle the redistribution craft — the voice decomposition, the range checks, the bar-by-bar voicing, the bass-line decisions — and give the user back something they can play, edit, or refine.

The output is what an arranger would actually use: standard MIDI, MusicXML, and a readable PDF. You can take it into your notation software, you can hand a part to a player, you can A/B it against the original in your head. It’s not a black-box “AI experience.” It’s a working file that fits into how arrangers actually work.

We’re in private beta with a small group of working arrangers, composers, and conductors. They give us feedback every week. The model gets better when they tell us where it’s wrong, which is most of what we care about right now.

Who’s on the team

We’re three people.

I’m Marcin Konopka — the founder, also a working musician. I trained in piano performance and computer science. I run the company side and most of the infrastructure. The reason I’m writing this post in first person is that ArrangementLabs.ai is not an institution. It’s a small team built around a specific musical conviction.

Our AI engineer handles the model training and the systems side. The person doing this work has been deep in ML for years and was the right collaborator for taking the redistribution problem from “interesting research” to “tool people use.” We’re deliberately not pretending the AI side is solved magic. It’s hard engineering and it’s getting better month over month because someone is doing the careful work.

Our arranger-in-residence is a composer with conservatory training and credits across European film and concert music. This is the person who tells us when the AI output is actually good and when it just looks plausible. The musical judgment lives here. Without that judgment, an AI arrangement tool ships demos. With it, the tool ships something a working musician will use.

We are small on purpose. Three people aligned around a specific craft conviction is a better team for building this tool than a larger team with mixed motivations would be.

What we believe

A few things we’ve come to believe in the process of building this:

AI in music shouldn’t be about generating anything. It should be about helping people do specific things they care about, better. The general “anything machine” framing produces tools that demo well and ship nothing. The specific tool framing produces things people use.

Arrangers are not going away — and they shouldn’t. The musical knowledge required to make a good arrangement is real, learned, and valuable. The interesting future is AI tools that compress the mechanical parts of an arranger’s workflow without trying to replace the judgment that makes an arrangement good. We want our tool to make arrangers faster and bolder, not redundant.

The right unit of feedback is a working musician on a deadline. Not a focus group. Not a survey. Not a Twitter poll. The feedback that improves the tool is “here is the score I was actually working on, here is what your tool gave me, here is what I had to change.” That’s the loop we’ve built and that’s the loop that keeps the tool honest.

Quality is set by use, not by demos. The metric we care about is whether someone in our beta cohort opens the tool again next week. Demo-driven AI work optimizes for impression. Use-driven work optimizes for retention. Different shape entirely.

What’s next

We’re targeting public launch later in 2026, after the beta cohort tells us we’re ready. The criterion for “ready” is not a feature list. It’s that the tool produces output a working arranger would use without rewriting, across a reasonable variety of source material and target ensembles. We will not ship the public launch until that’s true.

We’re also expanding to the German market in 2026. There’s a working community of arrangers, conductors, and music educators in the German-speaking countries that fits the tool’s profile well, and we’re building out localized copy and language support for that audience. (The reason this post is in English: our primary audience for the beta is global musicians and journalists who write about AI music tools. Localized posts will come.)

We’re also continuing to publish on this blog. The plan is craft-level writing about arrangement, voice redistribution, and the practical questions working musicians face when AI tools enter their workflow. Not marketing copy. Not thought leadership. Writing about the actual problems, by people working on them.

If you arrange professionally and want to try the tool with your repertoire, join the beta waitlist. We’re letting in a small cohort of testers each month and prioritizing people with a specific repertoire problem in mind — youth ensemble conductors, film composers, working arrangers with regular deadlines.

If you’re a journalist or researcher writing about AI music tools and want to talk to a team that takes the craft side seriously, we’re at hello@arrangementlabs.ai.

If you read all the way down here, thank you. We’re going to keep building.

What you can read next

• How AI Redistributes Musical Voices Across Ensembles — the craft primer that pairs with this post.
• Subscribe via RSS for future posts on AI arrangement and the craft we’re trying to build for.