Seven people signed off on a coffee cup and not one of them opened the file.
That's the detail that stuck with me out of the whole Starbucks Korea mess. Not the AI bit. Not even the promotion itself. Seven managers in the approval chain, and the internal investigation found some of them "signed off on it, as a matter of routine, without even opening the attached design file contained in the email," according to Bangkok Post's report on the company's own findings. The AI just made the mistake faster. The humans made it official.
What actually happened
In May 2026, Starbucks Korea ran a tumbler promotion branded "Tank Day," built around a "tank tumbler" cup design. It launched on 18 May, the anniversary of the 1980 Gwangju uprising, a pro-democracy revolt crushed by the South Korean military. Official figures record the crackdown killed 165 civilians, with 65 listed as missing and 376 others later dying of injuries, and many believe the real toll was higher. In South Korea, "Tank Day" on that date reads roughly the way "Bomb Day" would read in a lot of other places. It is not subtle once you know it.
But the team behind it didn't know it, apparently. The employees involved told investigators they'd "asked AI for suggestions and that the May 18 anniversary had never even crossed their minds," Shinsegae executive Jeon Sang-jin said, as reported by Bangkok Post. One employee later told Korean outlet MK they'd asked AI about the slogan and "didn't think of any link to May 18," and that the phrasing was chosen to match the rhythm of an earlier tumbler slogan, "Fits neatly in a bag." Nobody was trying to cause offence. They were trying to hit a deadline with a catchy line, and AI gave them one.
The fallout was not small. The Guardian reported the campaign spiralled into mass boycotts across South Korea. Shinsegae, which runs Starbucks under licence there, confirmed a sharp drop in sales and sacked the head of Starbucks Korea, Son Jung-hyun. The group's chairman held a press conference and apologised publicly. All of that from a tumbler.
Why this isn't really an AI story
Here's the bit that matters if you run a business and you're nowhere near Seoul. This wasn't an AI failure. AI did exactly what it's built to do: it pattern-matched a catchy phrase against a brief, at speed, with confidence. It has no institutional memory of a 1980 uprising unless someone tells it to check. It doesn't feel the weight of a date. It just writes.
So the failure was downstream of that. Seven approvals, zero pushback, and at least some of those approvals given without anyone opening the file. That's not an AI problem. That's a process with no friction left in it. Somewhere along the way, "AI wrote it, it sounds fine" became a substitute for "a human looked at it and thought about who might read it."
So I see a version of this every week with clients running their own campaigns through GHL, or agencies pushing content out for a dozen accounts at once. AI drafts the caption, the email subject line, the social post. Someone glances at it, it reads fine, it goes out. Most of the time that's genuinely OK, most captions aren't sitting on top of a national tragedy. But "most of the time" is exactly the trap. The one time it matters is the time nobody was checking, because the checking had gone the way of a lot of things that used to be someone's job, and slowly stopped being anyone's job at all.
And it's not only the copy. I've watched agencies hand AI the client's whole content calendar for a launch month, dates included, on the logic that a scheduling tool doesn't need a human to babysit it. Nobody checks whether a "Bank Holiday sale" post lands on the wrong bank holiday for that country, or whether a "big day" hook accidentally sits on a date that means something else entirely to a chunk of the audience. None of that is malicious. It's just speed with the friction taken out, the same speed that got seven Starbucks Korea managers to approve a file none of them opened.
The fix that would have actually worked
And no, the fix here isn't "don't use AI for marketing." Shinsegae didn't say that either, they said the problem was a lack of "social and historical sensitivity" in the review, per Bangkok Post. The fix is boringly simple and it's the same one I'd give any business using AI to write anything customer-facing:
One human, every time, actually reads it before it goes live. Not skims. Reads. For anything tied to a date, a place, a name, or a number, that person spends thirty seconds asking "does this date or phrase mean anything I don't know about." If you operate in a market you don't live in, day to day, that check matters more, not less, because you're the one least likely to catch a local reference an AI happily generated without knowing why it might be a problem.
If you're running campaigns for clients through an agency, this is worth turning into an actual checklist, not a vibe. Who reads the final copy before it ships. What do they check. What happens if they're not sure. Seven people in a chain isn't a safety net if none of them are actually reading the file. One person with a genuine checklist beats seven people rubber-stamping, every single time.
And the check doesn't need to be clever. It's one plain question, asked out loud, by an actual person, before the send button gets pressed: does anything in here, the date, the name, the phrase, mean something I don't know about. If the answer is "I'm not sure," that's the answer. Find out before it ships, not after a boycott tells you.
But there's also a second lesson here for anyone building AI into their own operations, which is most of the people reading this. AI is brilliant at volume. Seven captions, twelve email variants, a month of social posts, done before lunch. What it's rubbish at is knowing what it doesn't know. It will write "Tank Day" with exactly the same confidence it writes "Fits neatly in a bag," because to the model, both are just words that sound catchy together. It has no idea one of them sits on top of a massacre. That's not a bug you patch. That's a permanent limit, and the only fix is a human who actually knows the market putting eyes on it before it goes out.
The AI didn't ruin Starbucks Korea's month. A broken approval process did, and AI just made it faster to get there. That's the whole lesson, really. Speed without a check isn't automation, it's just a faster way to find out what you missed.
Want to see how this could work in your business? Book a call and let's talk about where you're at and what's possible.
Brewed by Steven, poured by Viktor
About Steven Tann: Steven helps business owners build systems that run themselves using AI. After 10+ years helping 7,000+ businesses and building his own autonomous operations, he's the bloke who actually does it, not just talks about it. Find out more at steventann.com.