# Speed Without Review *On Trust and AI — Applied* **A story about a marketing team that did everything the modern playbook told it to, and the one thing the playbook left out.** --- I've been watching the Starbucks situation in Korea unfold for the last few weeks, and I keep circling two questions. How did this actually happen? And what would have stopped it? I don't have inside knowledge of what went on in that room, and I'm not going to pretend I do. But the company's own investigation has been made public, and the mechanism it describes is one I've spent a lot of time thinking about. It's a useful case for anyone whose business depends on its brand, which is to say, for Starbucks, the only case that matters. ## What happened On May 18, 2026, Starbucks Korea launched a flash promotion called "Tank Day" for a new line of large tumblers called Tank. May 18 is the anniversary of the Gwangju Democratization Movement, the 1980 uprising where the military sent paratroopers and tanks against pro-democracy protesters and killed hundreds of civilians. It's one of the founding wounds of modern Korean democracy, commemorated every year, taught in every school. The campaign copy also echoed the 1987 cover-up of student activist Park Jong-cheol, whose interrogators claimed he died when they slammed a table during questioning. To the team, "Tank Day" was a pun on tumbler size. To the country, it read as a joke about the machines that crushed Gwangju, published on the day set aside to grieve it. The reaction was immediate. Videos of people smashing Starbucks tumblers spread within hours. President Lee Jae Myung, speaking from a Gwangju memorial that same day, [called the campaign "inhumane."](https://www.koreaherald.com/article/10740635) Starbucks Korea pulled the promotion before the day was out, and it didn't matter. Shinsegae, the conglomerate that runs Starbucks Korea under license, [fired the chain's CEO](https://www.bbc.com/news/articles/c0k2dz4y6gxo) that week. Sales dropped hard. Every store in the country closed early for the first time since Starbucks entered Korea in 1999, so staff could sit through history training. The chairman bowed three times on national television. He and the former CEO were [booked as criminal suspects](https://www.theguardian.com/world/2026/jun/06/starbucks-south-korea-tank-day-promotion-blunder). Starbucks headquarters in Seattle, which licenses the brand but owns no part of the Korean business, called it an ["unacceptable marketing incident."](https://www.koreaherald.com/article/10740635) A coffee promotion did all of that. The obvious explanation is that an AI wrote something monstrous. That explanation misses the point entirely. ## What the investigation found A week later, Shinsegae held a press conference and disclosed the findings of its own internal investigation. The chain of failures it described is almost line for line what I wrote about in *[On Trust and AI](/book)*. The team [asked an AI for campaign ideas](https://www.digitaljournal.com/article/starbucks-korea-reveals-series-of-mishaps-leading-to-tank-day-campaign/) and chose May 18 for commercial reasons: weekday online sales perform better. The anniversary, they told investigators, "had never even crossed their minds." Executive Jeon Sang-jin said priority "was given to the speed and immediacy" of the campaign, and that "not a single objection was raised." Five people built it. Seven approved it. Some of them, Jeon said, signed off "as a matter of routine, without even opening the attached design file." Imagine being the person on that team who had a flicker of something. The date felt off, but you couldn't place it, and there was no field in the tool for "this feels off," no place to route a hunch. The brief was due that day. You let it go. Whether that person existed at Starbucks Korea, I don't know. But there was no system that would have caught the doubt even if they had it. The investigation found no evidence that anyone set out to mock Gwangju, and a police inquiry is still open, so intent isn't mine to assign. The mechanism doesn't need intent. ## Where the model sits in this Could the AI have known that May 18 was sacred? Yes. A well-built system with the right context could learn that. But the team gave it a goal, maximize engagement on a weekday, without specifying what to avoid. That's how most AI deployments work in practice: the humans set the objective and assume the model shares their cultural background. It doesn't. And even if you built the context in correctly, you can't assume it holds. A model that was prompt-injected or had its training data poisoned could produce this result even with the right guardrails in place. That's why you model the blast radius: what's the worst thing this system could publish if it went wrong, and would anything in the pipeline stop it? At Starbucks Korea, the answer to the second question was no. Using AI to generate culturally timed, calendar-aware marketing is a genuinely good idea. Promotions tied to local dates outperform generic blasts, and a model that can draft dozens of concepts against a promotional calendar is a real advantage for a fast-moving team. The problem wasn't the tool. It was that nothing outside the tool was built to catch the cases where the tool gets it wrong, whether by ignorance, misalignment, or attack. ## What would have caught it In the book I call this failure mode "humans as a mop": AI generates the work and a person will check it, except checking never gets the time, tools, or authority the job requires. Stack that with "neglected human oversight," where rigorous review erodes into rubber-stamping, and you get what Starbucks Korea had: a full approval chain that caught nothing. The pieces that were missing are components of what I call an [AI control plane](/book): deterministic software wrapped around the non-deterministic model, sorting what can ship automatically from what needs a human to own the call. A date-and-lexicon check would have caught this particular campaign. But that's just the first gate. The deeper gap is that nothing in the process forced meaningful review before brand-facing content shipped. A good system surfaces the key concepts and taglines in the campaign, then routes specific questions to an accountable human who is a cultural expert on the market where the campaign will run. Not "does this look okay?" in a channel. Specific, structured questions: what does this name mean in the context of this date in this country? The system makes review easy by presenting the right context alongside the right question, so the reviewer can decide in seconds instead of hunting through a PDF they may or may not open. ## Proving that review is real Even with the right system, there's the rubber-stamp problem: people who click through. You can't fix that with a checklist alone, because a person who rubber-stamps a checklist is the same person who rubber-stamps an email. What you can do is test the reviewers. On brand-critical paths, the system periodically injects test fires: synthetic campaigns that contain known problems a competent reviewer would catch. An offensive tagline. A culturally loaded date. A logo misuse. If the reviewer approves the test fire, you know. You have a measurable score for each reviewer on each type of content, and KPIs attached to missing the bad events. When the score drops, you intervene before a real campaign slips through. I talked about this pattern in a [conversation with WB-40](https://wb40podcast.com/2026/02/11/343-trust-in-ai/), a UK technical podcast, earlier this year: the only way to trust that your human review layer is functional is to test it the way you'd test any other critical system. And as an added bonus, the same system is an excellent way to train new reviewers over time and keep experienced ones operating at peak performance. Gamify it. Give the humans a reason to care on every review. ## The brand is the product Tank Day filled in the cost of skipping all of this, and the numbers are public. In the first week after the promotion, [card payments dropped 26 percent](https://en.yna.co.kr/view/AEN20260527006700320), a loss of 8.47 billion won ($5.7 million) in seven days. By June, monthly card sales had fallen to [100.39 billion won, down 33.9 billion won (~$22.6 million) from April](https://www.koreajoongangdaily.com/business/starbucks-korea-number-of-app-users-sales-slide-in-june/12755033), the lowest in eight months. The Starbucks app lost 1.13 million monthly users. [E-Mart's stock dropped 25 percent in four trading sessions](https://finance.biggo.com/news/gEtmap4BLfE1EzqPndk-), and analysts flagged a contractual call option that could let Starbucks headquarters [buy back the Korean business at a 35 percent discount](https://www.koreatimes.co.kr/business/companies/20260528/starbucks-fiasco-puts-shinsegae-under-financial-strain), a potential loss of 640 billion won ($427 million) for Shinsegae. This wasn't an industry-wide slump: competitors like Mega MGC Coffee dropped only [6 percent in the same period, while Twosome Place gained 20.8 percent in new app installs](https://www.asiae.co.kr/en/article/2026052716380650493) as customers actively migrated. For Starbucks, the math is more lopsided than it would be for most companies, because the brand *is* the product. They don't win on coffee. They win on what the green logo means. ## What to build The fixes Starbucks Korea announced afterward, sensitivity checklists, outside historians, risk assessment moved into planning, are the right components. Whether they also build the test-fire layer that proves the system stays functional over time is what separates a one-time correction from an actual change. --- *Alexander Feick is the author of [On Trust and AI](/book), a field guide for leaders navigating AI adoption with accountability. Related: [The Verification Gap](/verification), [Making Verification Tangible](/verification-tangible), and [When You Stop Owning Your Words](/delegation).* [Read the book](/book) | [Back to blog](/blog)