Is There an AI Playlist That's Actually Okay to Play in a Cafe?

A little while ago I wrote a piece about AI background music playlists in cafes. The short version: a four-hour AI playlist scraped off YouTube gets caught by your customers within days, and instead of complaining, they stop coming.

After I published it, a question came back at me from my own side.

So is AI music just never okay?

Honestly, saying that would be convenient. It also isn't true. In Japan, AI-made store BGM is already selling as a monthly subscription, and the number of shops using it is growing. Nobody's reporting customers fleeing.

Same AI. One version wrecks a shop, the other one works. Let's look at what's different.

In Japan it's already a 980-yen-a-month product.

There's a service called BGM Pro, run by Easy Inc. It supplies roughly 5,000 original AI-produced tracks for store use, starting at 980 yen a month. The company states that every track is cleared for commercial use and that you don't need to file separately with a rights body like JASRAC. As of June 2026 it passed 100 paying shop contracts.

A hundred isn't a big number. But it's a number that confirms the thing works. Shops are paying real money for it, and the count keeps climbing.

Which flips the question around. An AI playlist grabbed off YouTube and BGM Pro's AI tracks are, tool-wise, not that different. Both came out of a generative model. And yet the outcomes diverge.

What separates them isn't the tool. It's the editing.

The difference is exactly one thing. Someone curated one of them. Nobody curated the other.

Assembling 5,000 tracks for store use means far more tracks got thrown out upstream. It means someone did the work of filtering out the ones with strange lyrics, the ones whose choruses never close, the ones that keep circling the same subject. A four-hour YouTube video has none of that process. Generate, stitch, upload. There's no reason to filter — view counts come from the thumbnail and the title, not the quality of the track.

The industry drew its line in the same place. Spotify removed 75 million tracks it judged to be spam over the past year. But look at the policy it published alongside that, and using AI is not itself treated as a demerit. It requires AI usage to be disclosed in the credits via the DDEX standard, while explicitly stating that the disclosure won't be held against you. What gets filtered is mass uploading, duplicates, metadata manipulation, and abnormally short tracks.

So here's the summary. The criterion isn't whether AI made it, but whether a human touched it. The platforms set it that way first, and customers' ears turn out to catch the same thing.

Which means "is AI music okay to use" is a badly formed question. Formed properly it becomes: has this playlist cleared two gates?

Gate one: are you allowed to play this?

Before we get to taste, let's talk rights. This is more tangled than it looks.

The public performance royalty may not apply. Purely AI-generated tracks not registered with KOMCA (Korea Music Copyright Association) aren't works under the association's administration, so they fall outside royalty collection. That's the basis for AI-only services advertising "free to play in your store."

But not owing a performance royalty doesn't mean you're allowed to play it. Two more things snag here.

The first is platform terms of service. The terms for YouTube and general streaming services permit personal, non-commercial use only. It makes no difference whether the audio is AI or human. Playing it in a shop is a terms violation on its own. This is a separate matter from copyright, and quite a few people don't seem to know it.

The second is the terms of the generation tool itself. That's changed a lot recently.

In November 2025, Warner Music and Suno settled their litigation and signed a licensing deal. The terms included this: all models trained without authorization get destroyed, downloads get restricted to paid accounts, and tracks made on free accounts are barred from commercial use. Universal and Udio reached a similar agreement around the same time, and the Sony suit is still ongoing.

What that means is that the logic of "I generated it for free so I'll use it for free" is narrowing. A shop owner has no way of knowing which account and which model produced a free AI playlist on YouTube. And they're playing it in a commercial space without knowing.

Korean copyright law still has no explicit provision for AI-generated works. Gray area is the accurate term. But there's only one way to stand on the safe side of a gray area.

Can you point to who granted commercial use, in writing? If that question has an answer, you pass. If it doesn't, you fail. It doesn't matter whether it's a store music service, a licensed music library, or an AI-only BGM service. The criterion isn't the tool, it's the document.

Gate two: is this music right for your shop?

Clearing the rights isn't the end of it. Now it's the ear's turn.

Let me clear away a common misconception first. Store music is not the job of playing good music, it's the job of playing fitting music. And that isn't a matter of taste; it's been confirmed experimentally.

There's a famous 1999 experiment run by Adrian C. North, David Hargreaves and colleagues in a supermarket wine aisle. Play French music and French wine sold better; play German music and German wine did. On the French music days, French wine outsold German 5 to 1. The shoppers hadn't even registered that the music was French.

There's one more. In 1993, Charles Areni and David Kim alternated classical and Top 40 pop in a wine store, and on classical days shoppers picked more expensive wine. Not more bottles — a higher price per bottle.

The most interesting part of that study isn't the result, it's the interviews. Out of 44 people, one brought up the music unprompted as a reason for their purchase, and when asked directly whether the music had influenced their choice, 86% said no.

Customers don't evaluate music. They react to it, and they don't even know they reacted. Which makes "our customers don't care about music" less an observation than a misconception. It isn't that they don't care — it's that they aren't conscious of it, and those two things are completely different.

So gate two's question isn't "is this track good?" It's is this sound telling the same story as what we sell? Brewing specialty beans while playing lo-fi of unknown origin is the same species of mismatch as playing German music in the French wine aisle.

So how do you actually pick?

Translate the two gates into practice and it comes out like this.

One, listen in sets. Judge on one track and everything passes. AI audio is perfectly fine at the single-track level. Listen to at least thirty minutes — eight to ten tracks in a row. The problem only surfaces in repetition.

Two, count the repeated subjects. Listen to thirty tracks and write down the subjects that show up in the lyrics. Coffee, rain, seasons, days of the week. If the same word comes back more than five times, that playlist gets caught within days.

Three, split it by time of day. Don't run the day as one block. Instrumentals during the morning work hours, vocals in the afternoon conversation hours as long as the lyrics don't jump out, tempo nudged up in the last hour before close. Three files is enough, and just doing this much starts you designing for dwell time.

Four, ask your staff. This is the surest method and the one nobody uses. The owner isn't in the shop all day; the staff are. Run a playlist for a week, and instead of "how's this music?" ask "is there anything in here you're sick of?" You have to change the question to get an answer.

Five, keep proof of commercial rights on file. Screenshot the terms page, save the contract email, whatever — you need something to produce when somebody asks later. If you have nothing, that playlist already failed gate one.

What does it cost?

Let's look at realistic numbers.

For a domestic store music service, Bizmelon's coffee shop rate for 50–99㎡ of floor space is 23,000 KRW (about $17) a month. The public performance royalty is billed separately at 4,000 KRW (about $3) a month for the same bracket. Call it 27,000 KRW (about $20) a month all in. Depending on the service, some products bundle the royalty in.

The Japanese AI BGM above started at 980 yen a month. Even at a rough exchange rate that's under 10,000 KRW.

Nearly three times the difference, and the reason is simple. You're paying for different things. A store music service buys you the right to play tracks that real singers sang. An AI BGM service buys you the right to play generated output. The difference is whether a customer hears a song they know.

Which one is right is for the shop to decide.

• If customers humming along and pulling out Shazam is an asset for your space, songs they know have to be playing. That's a store music service
• If the sound needs to recede into the background, curated AI audio is a valid answer too

There's only one wrong answer: playing something nobody chose.

Something still nags at me.

There's something I should be honest about here. Wrap this up on cost and taste alone and a piece of it goes missing.

The more AI music fills up shops, the smaller the share left for the musicians and composers who would have been in that slot. This isn't an emotional claim; it's already in dispute. The American Federation of Musicians (AFM) filed suit against Universal and Warner, on the grounds that their members' recordings were licensed to AI companies without compensation or credit. The labels and the AI companies reached agreements, and the people who actually made the sound were left out of them.

I can't settle this question on a shop owner's behalf. But I can hand over the material for the judgment.

Using curated AI BGM is legal, and at this moment it's a practical choice. At the same time, it's right to make that choice knowing which side of the industry it votes for. Doing it unknowingly and doing it knowingly are different acts, even when the result is identical.

At minimum I can say this much: taking an unattributed four-hour video off YouTube and pressing play belongs to neither camp. It's just not choosing anything.

Good shops don't pick their music. They decide it.

"Picking a playlist" can sound like a light thing to do. It's background noise anyway.

But what those experiments say is the opposite. Customers don't listen to music and then judge. They judge inside the air the music made. How expensive a wine to buy, how long to sit here, whether to come back next week. And then they say it wasn't because of the music. The way 43 out of 44 did.

So store music is less something you pick than something you decide. You're determining what kind of sound your space makes, and if you don't make that decision, the YouTube algorithm makes it for you.

You can use AI-made tracks. There are just two things to clear. Whether you're allowed to play it, and whether it's right for your shop.

These days when I walk into a cafe I sit still for about three minutes and listen to the speakers first. Before the coffee arrives, I have a rough idea of what kind of place it is. Before anyone mentions the beans.