Starbucks Korea & Wrap-Up (Optional Reading)

Note on the Starbucks Case We will not be discussing the Starbucks case in class. If you have a few minutes, please skim the case text and Exhibit 1. You do not need to prepare the case or discuss it in your living groups, though I would be delighted if you did.

Case Summary

According to the case, Starbucks Korea launched a promotion for large-capacity stainless steel tumblers on May 18, 2026, using the name "Tank Day" and an onomatopoeic Korean slogan. To the team, these words seemed to convey capacity, sturdiness and memorability; to many consumers, the date and the words connected to historical trauma in Korea.

May 18 is linked to the commemoration of the Gwangju Democratization Movement; the slogan was in turn associated with official cover-up rhetoric from another historical incident. The problem came from the combination of words, date, region and public memory — it can't be caught just by checking whether a single line of copy is grammatical or obviously offensive on its own.

The case cites reports and company statements saying the team consulted AI tools while drafting the copy, that the approval chain involved seven managers, and that some approvers reportedly did not open the attachment. The materials do not provide the full prompts, model versions or each person's actions, so we cannot conclude that AI independently created all the content, nor precisely allocate each individual's responsibility.

The campaign was pulled within hours, but screenshots had already spread. The case describes apologies, management changes, boycotts and investigations, followed by nationwide early store closures for training on historical and social sensitivity. The fallout from a headquarters marketing mistake also landed on store employees. Training can fill knowledge gaps, but it does not automatically change who needs to see what, who can stop something, and who bears final responsibility.

Exhibit 1 places AI uses across twelve marketing-mix elements, covering activities such as demand insight, product, pricing, brand, channels, sales, advertising, promotion, packaging, display, service and fulfillment. So the case is not only about writing copy, but about how AI takes part in the entire market-facing decision chain. As generation gets faster, human judgment and organizational design matter no less.

Key Data

Fact	Meaning	Source and limits
Campaign date 2026-05-18	The date and local history together form the context	pp.1–2; as narrated in the case
Scale of the Korean business	The case cites over 2,000 stores and 23,000 employees in 2026	p.2; not operating data verified in real time for these notes
Ownership/operating model	The case says Starbucks sold its original 50% stake in 2021; since then E-Mart holds control, Shinsegae runs day-to-day operations, under a licensing model	p.2; the licensing model is not taken to mean the global brand bears no governance responsibility
Approval chain	Reports say seven executives and managers were involved, some of whom reportedly did not open the attachment	p.1; not seven independent, effective reviews, nor can we assert that none of the seven read it
Speed of takedown	Stopped within hours; screenshots had already spread	p.1; digital spread is not fully reversible
Follow-up training	The case describes over 2,000 stores nationwide closing early for training on 2026-06-22	p.3; no data provided on post-training error rates or trust recovery
Exhibit 1	12 marketing-mix elements; the "Personal Settings" row in the table actually discusses sales work	pp.4–5; these notes translate it by content as "personal selling," noting the original row title appears to be a typo
Survey percentages	Exhibit 1 cites surveys from different years and different vendors/consultancies	Not merged into one overall adoption rate, and not used directly as proof of ROI
Exhibit 2	Different failures: voice ordering, customer-service promises, factual errors in demos, promotional promises, visual expectations	pp.6–7; failure mechanisms differ, so not all incidents can be attributed to the same model flaw

Key People

Actor	Role	Relationship to resolve
Marketing creative team	Used AI, selected candidate content	Responsibility between generated suggestions and the decision to publish
Approvers/management	Review and launch authorization	Whether a sign-off corresponds to clearly defined review content and real time spent
Local operating company/Shinsegae	Operations and public response	Local historical knowledge, release process, store support
Global brand owner	Licensed brand and shared reputation	How global minimum standards and local judgment fit together
Store employees	Face customers and public emotion	Should not take the blame for headquarters' decisions; should get facts and escalation support
Affected groups and consumers	Experience and judge the brand's behavior	Repair should listen to their experience, not just explain that offense was unintended
AI tools	Provide candidates and assist with analysis	Hold no organizational authority and do not bear the brand's final commitments

Core Tensions

1. The tension between more creativity and faster launches, and enough time for contextual understanding and checking.
2. More layers of approval do not necessarily mean more effective judgment; responsibility can also get diluted layer by layer.
3. The relationship between a global brand's consistent baseline and local history, language and social memory.
4. The visibility of after-the-fact training versus the durability of changes to everyday processes.
5. Short-term customer-experience metrics versus trust and respect, which are hard to quantify quickly.

My Answers

Q1. Was this mainly an AI failure or a management failure?

It was a failure of an AI-assisted organizational workflow. The tool may have generated or helped refine inappropriate material, but people selected it, paired it with a date, approved it, and released it. The case does not provide enough records to isolate each contribution. Calling it only an AI failure would hide the decisions that allowed the output to become a public campaign.

The failure was contextual. A product name, sound, date, and local history jointly changed the meaning. A generic grammar check or a second model asked whether the copy was “good” would not reliably test that combination. A reviewer needed the complete campaign, the intended market, and the release calendar.

Seven names in a chain do not prove seven meaningful reviews. If everyone assumes someone else checked the issue, another approval can add delay without protection. I would ask what each approver had to inspect, whether the exact final asset was visible, and whether anyone could stop the launch without penalty.

Training can improve knowledge, but it should support a redesigned workflow. I would assign a campaign owner, specify cultural review triggers, require local review for sensitive material, and record approval of the final version. I would also protect staff who raise uncertainty. These are proposals; the case does not show that such changes have already been implemented or proved effective.

Q2. Where can AI help across the customer journey, and where is judgment needed?

Exhibit 1 is a map of possible applications, not a claim that every application should be automated. I would group the twelve elements by the decisions they support and the cost of error. One model can contribute to several stages, but responsibility stays with the relevant business owner.

At discovery, AI can organize reviews, service conversations, and demand signals. Humans must check whether the sample represents customers and whether a pattern supports a causal claim. At design, it can propose products, positioning, packaging, and prices. Managers must test feasibility, actual benefits, and customer understanding.

At acquisition and purchase, it can draft campaigns, prioritize leads, personalize offers, and improve product discovery. Review must cover context, factual claims, authorized prices and terms, and whether the firm can fulfill its promise. A high click rate does not establish a good customer outcome.

At delivery and service, it can help forecast stock, retrieve policy, draft replies, and route cases. Decisions involving exceptions, conflicting policies, or distress need a clear human path. Finally, analytics should connect conversion to returns, complaints, retention, and delivery quality. Begin with a narrow use case and a measured baseline; expand when quality and net benefit are demonstrated.

Q3. What would you change immediately, and how would you know it works?

Immediately, stop the offending assets across channels, preserve records for review, explain the error accurately, and support front-line staff. Engage affected groups with listening and a concrete repair process. Do not imply that store employees created the headquarters campaign. Do not claim that lack of intent cancels impact.

Before the next campaign, require the owner to review a one-page release packet: final wording and visuals, date, audience, geography, evidence for claims, and any sensitive references. Add an independent local reviewer when a trigger is met. Give that reviewer stop authority and a rapid escalation path. Avoid adding seven generic signatures to every routine item.

Over the next month, pilot this process on a limited campaign set. Keep useful AI support for drafting and organizing information. Log model-assisted work and material edits enough to reconstruct the decision, without creating unnecessary data collection. Maintain a versioned calendar and reference set, but do not assume a sensitive-date list can replace local judgment.

Measure serious errors escaping review, meaningful issues caught before launch, review time, rework, complaints, and recovery outcomes. A rise in near-miss reports may initially reflect better reporting rather than worse behavior. Review samples independently and adjust thresholds. The cost is reviewer capacity and slower approval of some campaigns; prioritize it where harm is hard to reverse.