CASE 02
The Apple Card Credit Limit Controversy
When a decision is lawful, unexplainable, and indefensible at the same time
In November 2019, a software developer posted that the Apple Card had given him roughly twenty times the credit limit of his wife — although they filed joint tax returns and she had the higher credit score. Within days a co-founder of Apple said the same had happened to him and his wife. A state regulator opened an investigation. Sixteen months later that regulator concluded it had found no unlawful discrimination. The company had broken no law, used no prohibited data, and could still not tell a single customer why their limit was what it was.
Course anchor
Ch. 6 (required) + Ch. 7–18 frameworks
Case period
Aug 2019 – Mar 2021 (legacy to today)
Decision point
Nov 2019 — how to answer “why?”
Pairs with Case 01. In The DAO, the rules were fully visible and executed without exception, and the harm came from following them. Here the rules are invisible and the harm comes from not being able to see them. Read together, the two cases cover both ends of the same problem: what happens to accountability when decisions are made by a system rather than by a person. Teams may take either case; teams taking this one may cite Case 01 as a comparison.
What this case is designed to teach
After analyzing this case, you should be able to:
- Analyze an ethical failure without a verdict to lean on. The regulator found no unlawful discrimination. You cannot assert that the company discriminated; you must work out what, if anything, it did wrong.
- Distinguish three different ways an unequal outcome can arise — using a prohibited attribute, using a proxy for it, or applying an identical rule to unequal starting positions — and explain why only the first is obviously unlawful while all three raise ethical questions.
- Apply Exhibit 6.7 to an issue whose reaction was wildly out of proportion to its measured intensity, and identify the factor the framework does not contain.
- Evaluate responsibility across an organizational boundary (Ch. 11): one firm owned the brand and the marketing, another made every credit decision, and customers could not tell them apart.
- Diagnose a communication failure (Ch. 14) in which the most damaging sentence was spoken by front-line staff who had been given no way to say anything else.
- Design an appeal path for an automated decision — and argue for who pays for it.
Part A — The Case
Read the whole narrative before looking at the questions. The facts you need are in this section and in the exhibits that follow.
A.1 Background: what a credit limit actually is
When a bank issues a credit card it must decide two things: whether to approve the applicant at all, and how much credit to extend. That second number — the credit limit — is set by an underwriting model that estimates how much the applicant can be expected to repay. Typical inputs include income, existing debt, payment history, length of credit history, and the applicant’s credit score.
In the United States, the Equal Credit Opportunity Act and its implementing rules (Regulation B) prohibit discrimination in credit on the basis of sex, marital status, race, religion, national origin, and age. The prohibition is strong enough that lenders are, in most circumstances, not permitted even to ask an applicant’s sex for an ordinary credit card application. Keep this in mind: “we do not use gender” is not a policy the bank chose. It is what the law requires.
The same rules require a lender that takes an adverse action — declining an application, or granting materially less favourable terms than requested — to tell the applicant the specific principal reasons for it. The legal framework, in other words, already assumes that a credit decision is something a human being can be given a reason for. Hold that thought.
A.2 The product, and who actually decides
Apple Card was announced in March 2019 and launched to United States customers in August 2019. It was presented as a rethinking of the credit card: no fees, a physical titanium card, spending categories displayed in the iPhone Wallet app, and a marketing line that told customers it was “created by Apple, not a bank.”
It was, in the way that mattered most, created by a bank. Goldman Sachs Bank USA was the issuer. Goldman held the regulatory licence, carried the credit risk, built and operated the underwriting model, and made every approval and limit decision. Apple supplied the brand, the hardware, the software interface, and the customer relationship. A customer who felt wronged by a credit decision experienced it entirely inside an Apple product, made by a company whose name did not appear in the marketing.
One further design decision matters more than any other in this case. At launch, Apple Card had no joint accounts, no co-signers, and no way to combine two people’s finances. Every applicant was assessed strictly as an individual. A married couple who shared a home, a bank account, and a tax return would nonetheless be scored as two unrelated strangers — each on whatever credit history stood in their own name alone.
Hold this in mind
Nothing about individual-only assessment mentions gender. But in a society where, within many married couples, one partner has historically held more of the credit accounts, the mortgage, and the earnings record in their own name, individual-only assessment does not treat the couple identically. It treats them consistently — and consistency applied to unequal starting positions produces unequal results. Whether that is a defect or a virtue is one of the things this case asks you to decide.
A.3 November 2019: the thread
In early November 2019, David Heinemeier Hansson — a well-known software developer with a large following — posted publicly that Apple Card had given him approximately twenty times the credit limit granted to his wife, Jamie Hansson. The couple filed joint tax returns, had been married for years, shared assets, and — by his account — she had the higher credit score. He called the outcome a sexist program.
The thread spread quickly, and the detail that carried it was not the ratio. It was his account of what happened when he asked why. Customer service representatives, he reported, could not explain the decision, could not identify which factor had produced it, and could not point to anyone who could. What they could say was that the decision had been made by the algorithm.
Within days Steve Wozniak, a co-founder of Apple, replied that he and his wife had experienced the same pattern at a ratio of about ten to one, despite holding no separate accounts or assets. On 9 November 2019 the New York Department of Financial Services (DFS), which regulates Goldman Sachs Bank USA, announced an investigation.
A caution about the evidence
The ratios in this case come from the individuals themselves, on social media. They were not audited, the underlying applications were never published, and neither couple’s full financial position is public. Two vivid anecdotes are not a dataset. Your report may treat them as allegations that triggered a regulatory response — it may not treat them as measurements of how the model behaved across its customers.
A.4 The response
Goldman Sachs responded publicly within days. The substance of its position, stated by the chief executive of Goldman Sachs Bank USA, was that credit decisions were based on an individual’s own creditworthiness — income, debts, payment history and similar factors — and that gender was not an input to the model, was not collected, and could not have been used. The bank also said that it would re-evaluate the credit line of any customer who believed theirs did not reflect their circumstances, and it later moved to support shared accounts.
Every sentence of that response was, so far as the record shows, true. It also failed almost completely, for three reasons worth separating:
- It answered a question nobody had asked. The complaint was not “you typed my gender into a computer.” It was “the outcome tracks gender and you cannot tell me why.” A denial that the input was used does not address an allegation about the output.
- It offered a remedy available only to people who complained. Re-evaluation on request is a real remedy for the customer who asks. It does nothing for a customer who never knew a different answer was possible — and, in practice, the customers who were re-evaluated first were the ones with the largest audiences.
- It did not come from the brand the customer was dealing with. Apple, whose name was on the card, on the phone, and on the marketing, said comparatively little in public.
Meanwhile the single most damaging sentence of the episode was not uttered by any executive. It was said, repeatedly, by front-line staff: it’s just the algorithm. Those employees were not being evasive. They had been given a system whose outputs they were required to defend and whose reasoning they had never been given access to.
A.5 Three competing explanations
By late 2019 three incompatible accounts of what had happened were in circulation. Each has serious people behind it. Your analysis has to engage all three.
- Nothing went wrong. The model does not use a prohibited attribute. The two publicised cases are anecdotes, not evidence of a pattern. Married couples frequently have very different individual credit histories, and a lender that assesses individuals will reflect that. A lender that instead assigned limits by household would be making decisions about one person using another person’s finances — which creates its own serious problems, including for anyone who needs credit independent of a spouse.
- The model reproduced an existing inequality. Even with gender excluded, the inputs themselves carry the imprint of decades in which credit, property and earnings were recorded disproportionately in men’s names. A neutral rule applied to those inputs will produce outcomes that track gender. On this reading the model is working exactly as designed, and that is the problem.
- The model was never the issue. The failure was that a consequential decision about a customer’s financial life could not be explained, questioned, or appealed by an ordinary person through any channel except public embarrassment. On this reading, whether the limits were statistically fair is almost beside the point.
A.6 The regulator’s answer
In March 2021, after an investigation lasting roughly sixteen months, the New York Department of Financial Services published its findings. The conclusion was that the Department found no evidence of unlawful discrimination on the basis of sex under fair-lending law. The underwriting model did not use gender. The differences the Department examined were explained by the individual credit characteristics the model was entitled to consider.
That was not the whole of the report. The Department was critical of the consumer experience — the opacity of the process, and customers’ inability to obtain a meaningful explanation — and used the episode to argue a broader point: that credit decisions increasingly rest on models whose reasoning is not visible to the people they affect, and that a finding of legal compliance does not by itself make such a system trustworthy. The regulator, in effect, cleared the bank of the charge that had been made and then raised a different one.
Why this is the hard part of the case
Most business-ethics cases end with a finding of wrongdoing that the analysis can be built around. This one does not. If you want to argue that Goldman behaved badly, you must do it without claiming it broke the law, because on the available record it did not. If you want to argue that it behaved acceptably, you must explain why a lawful decision that cannot be explained to the person it affects is acceptable. Both positions are available. Neither is free.
A.7 Aftermath
- The structural fix arrived later. In 2021 Apple and Goldman introduced the ability for two people to share an Apple Card account and combine their credit lines. The feature that would have prevented the original complaint shipped roughly eighteen months after launch — and after the controversy, not before it.
- Regulatory attention did not end with the fair-lending finding. In 2024 the U.S. Consumer Financial Protection Bureau took enforcement action against both Apple and Goldman Sachs over Apple Card customer service and dispute handling, ordering penalties and customer redress. Note what this says about the 2021 report: the Department’s criticism of the consumer experience — the part that carried no legal finding at the time — is the part that later produced enforcement. Verify the details and figures of this action before citing it.
- The partnership itself came under strain. Goldman Sachs subsequently retreated from consumer lending, and the future of the Apple Card partnership was reported to be in question. Check the current status before writing about it; this is still moving.
- The question generalized. Explainability of automated decisions moved from a specialist concern to a mainstream regulatory expectation across credit, insurance, hiring and beyond. The practical management question this case raises — can you tell the affected person why? — is now asked of systems far removed from credit cards.
Part B — Exhibits
You may reproduce these exhibits in your report, but you must extend them with your own analysis. Copying them unchanged earns no analytical credit.
Exhibit 1 — Timeline
| Date | Event | Why it matters for the analysis |
|---|---|---|
| Mar 2019 | Apple Card announced; marketed as “created by Apple, not a bank.” | Establishes the brand/decision-maker gap that drives the accountability question. |
| Aug 2019 | Apple Card launches in the United States. No joint accounts, no co-signers; individual assessment only. | The design decision that produced the outcome — taken before any controversy, for reasons that were never publicly ethical or unethical. |
| Early Nov 2019 | Developer posts that his limit is ~20× his wife’s despite her higher credit score and joint filing. | The trigger. Note the complaint is about the output, not the input. |
| Nov 2019 | Front-line staff reportedly unable to explain decisions; “it’s just the algorithm.” | The communication failure (Ch. 14) and the job-design failure (Ch. 12) in one sentence. |
| Nov 2019 | Apple co-founder reports a ~10× ratio for his own household. | Escalates issue intensity through proximity and source credibility — not through new evidence. |
| 9 Nov 2019 | New York DFS announces an investigation. | The legal/political component of the environment activates (Ch. 3). |
| Nov 2019 | Goldman states gender is not an input; offers re-evaluation on request. | Classic social responsiveness — short-term, pragmatic, triggered by visibility (Exhibit 6.2). |
| Mar 2021 | DFS reports no evidence of unlawful sex discrimination; criticises transparency and the consumer experience. | Removes the easy verdict. The ethics analysis has to survive without it. |
| 2021 | Shared Apple Card accounts introduced. | The structural remedy — available, but only after the reputational damage. |
| 2024 | CFPB enforcement action against Apple and Goldman over Apple Card servicing and disputes. | The non-legal criticism of 2021 becomes a legal finding later. Verify before citing. |
Exhibit 2 — Stakeholder inventory (starting point only)
Deliberately incomplete and unranked. Question 3 requires you to add missing stakeholders, assess power and legitimacy, and rank them.
| Stakeholder | What was at stake | Power to affect the outcome | How the episode ended for them |
|---|---|---|---|
| Applicants who received low limits | Access to credit; the implication that their finances were judged and found wanting, with no reason given. | Individually almost none. | Re-evaluation available — if they knew to ask. |
| The two publicised couples | The same stake as everyone else, plus a public platform. | Very high — large audiences and media credibility. | Re-evaluated quickly; triggered a regulatory investigation. |
| Goldman Sachs Bank USA | Regulatory standing, the consumer-lending strategy, reputation as a new entrant in retail banking. | Total control over the decision system. | Cleared on fair lending; criticised on transparency; later enforcement on servicing. |
| Apple | Brand equity built on customer trust and privacy — its most valuable intangible asset. | High commercially; none over credit decisions. | Largely absent from the public response; brand carried the damage anyway. |
| Front-line customer service staff | Their ability to do their job, and to face customers with an answer. | None — required to defend decisions they could not see. | Became the public face of a failure they did not cause. |
| The underwriting / data science team | Professional judgment; whether anyone asked them what the model would do to different groups. | High at design time; invisible afterwards. | Not publicly identified or heard from. |
| New York DFS | Consumer protection; its own credibility as a regulator of algorithmic lending. | Formal investigative authority. | Found no violation; issued criticism it could not enforce at the time. |
| Customers who never complained | The same treatment, without the remedy. | None. | Unknown — and that is the point. |
| Future applicants in other sectors | Whether “the system decided” becomes an acceptable answer elsewhere. | None; not present. | Inherited the precedent. |
Exhibit 3 — Three competing diagnoses
| Diagnosis | Core claim | What follows if it is right | Strongest objection to it |
|---|---|---|---|
| Compliance is enough Nothing went wrong |
The law prohibits using sex; the model did not use it; the regulator confirmed this. Two anecdotes are not a pattern. | No remedy is owed. The correct response is better communication, not a different model. | It makes the ethical question identical to the legal one — which is precisely the move Ch. 6 exists to resist. |
| Neutral rule, unequal inputs Structural |
Applying one consistent rule to historically unequal credit records reproduces the inequality in the records. | The firm owes outcome testing, not just input exclusion — and possibly a redesigned product. | The alternative — assessing people on a spouse’s finances — has its own victims, including anyone who needs credit independent of a partner. |
| No path to an answer Procedural |
The defect is the absence of explanation, challenge, and appeal — not the distribution of limits. | The fix is governance: explainability requirements, an appeal channel, staff authority. It works regardless of whether the model was fair. | It may be too comfortable: it lets the firm off the substantive question of whether the outcomes were defensible at all. |
Exhibit 4 — How an unequal outcome arises without a prohibited input
Students routinely collapse these three into “the algorithm is biased.” They are different mechanisms, with different evidence requirements, different legal status, and different remedies. Your analysis must say which one it is alleging.
| Mechanism | What happens | Legal status (US credit) | What would fix it |
|---|---|---|---|
| 1. Direct use Disparate treatment |
The protected attribute is an input and moves the result. | Clearly prohibited. Not alleged here, and the regulator found no evidence of it. | Remove the input. (Already the case.) |
| 2. Proxy variables | A permitted input correlates strongly with the protected attribute and stands in for it. | Contested and fact-specific; may be unlawful where a neutral practice has a disproportionate effect that business necessity cannot justify. | Test outputs by group; identify and justify or remove the offending inputs. |
| 3. Equal rule, unequal history | The inputs are genuinely relevant and correctly measured — but the populations arrive with different histories, so identical treatment yields different results. | Generally lawful. This is the mechanism most consistent with the public record in this case. | Not fixable inside the model. Requires a product change (e.g. joint accounts) or a policy choice about who bears historical disadvantage. |
This exhibit describes mechanisms for the purpose of management analysis. It is not legal advice, and fair-lending law differs by jurisdiction. If your argument depends on a legal conclusion, cite a source for it.
Exhibit 5 — Key figures
~20× / ~10×
Credit-limit ratios reported by the two publicised couples — self-reported, never audited
~16 months
From the opening of the DFS investigation to its published findings
0
Findings of unlawful sex discrimination by the regulator
~18 months
From launch to the introduction of shared accounts — the structural remedy
Data caution. The ratios are allegations by the individuals concerned, not measurements. Do not build a quantitative argument on them. The strongest numbers in this case are the durations — how long it took to answer the question, and how long it took to ship the fix.
Part C — The Assignment
C.1 How this case satisfies the syllabus requirements
| Req. | Syllabus wording | How you meet it in this case |
|---|---|---|
| (1) | Select a real organization or industry case and analyze it using the frameworks covered in this course. | Goldman Sachs Bank USA and Apple are real, documented organizations with a public regulatory record. Apply Ch. 7–18 frameworks in Questions 6–9 — at minimum communication (Ch. 14) and control (Ch. 18). |
| (2) | Include an explicit business ethics and social responsibility analysis: identify the ethical issues, map the stakeholders, assess conduct against ethical criteria and CSR expectations (Ch. 6). | Questions 2, 3, 4 and 5 are mandatory and carry the largest share of the marks. Section 3 of your report must be titled “Ethics and Social Responsibility Analysis.” |
| (3) | Close with a recommended course of action and the reasoning that supports it. | Question 10 — two parts: the November 2019 response, and a forward-looking design for automated decisions. |
| (4) | Written report and oral presentation in English. | Part G provides the terminology and presentation language support. |
C.2 First decision: which organization are you analyzing?
As in Case 01, this is where weak reports fail. Two companies are involved and they had completely different roles, powers, and duties. State your unit of analysis in the first paragraph and hold it for the whole document. Choose one:
(a) Goldman Sachs Bank USA
The regulated issuer that built the model and made every decision. The most defensible choice, and the one with the clearest duties.
(b) Apple
Owned the brand, the marketing, the interface and the customer relationship — and none of the decisions. Best for responsibility without authority, and Ch. 9 brand strategy.
(c) The partnership as a unit
Treat the two firms as one system with a split down the middle. Best for Ch. 11 structure — but you must define the boundary precisely or the analysis blurs.
(d) Algorithmic consumer lending as an industry
The most ambitious. Requires evidence beyond this one episode; do not choose it unless you will actually gather that evidence.
Whichever you choose, the others appear as stakeholders, not as the subject.
C.3 Four traps to avoid
- Asserting that the company discriminated. The regulator examined precisely that allegation and did not sustain it. You may argue that the outcome was unjust, that the design was careless, or that compliance was an inadequate standard — all of those are open to you. Stating as fact that it discriminated unlawfully is a sourcing error, and it will cost marks.
- Treating two anecdotes as data. The reported ratios are allegations. Build your argument on the documented conduct — the product design, the response, the explanation failure — not on numbers nobody audited.
- Turning this into a technology report. You may spend at most one page on how credit models work. Nobody is grading your understanding of machine learning. Exhibit 4 gives you all the conceptual apparatus you need.
- Blaming the front-line staff. “It’s just the algorithm” was a symptom. Ask who designed a job in which that was the only available answer, and why nobody asked before launch what employees would say when a customer demanded a reason.
Part D — Analysis Questions
Questions marked REQUIRED must be answered by every team — they implement syllabus requirements (2) and (3). From the questions marked CHOOSE, answer at least three, and include at least one from the Ch. 14 / Ch. 18 pair.
Context and unit of analysis REQUIRED
Define the organization you are analyzing (see C.2) and describe its external environment in 2019. Which environmental components exerted the most pressure? Pay particular attention to two of them: the sociocultural component (a period of intense public attention to algorithmic fairness) and the legal/political component (an established fair-lending regime that was written for human underwriters). Then characterize the organizational culture of your chosen firm using the Ch. 3 dimensions, and explain how a culture strong in one area — engineering excellence, or design and secrecy — can be simultaneously weak in the capability this episode demanded.
Frameworks: Ch. 3 external environment components; environmental uncertainty; dimensions of organizational culture; strong vs. weak culture.
Identify and rank the ethical issues REQUIRED
Identify at least four distinct ethical issues and state each as a decision faced by a named actor. Candidates include, but are not limited to: launching a consumer credit product with no joint account capability; marketing a bank product as “not a bank”; deploying a decision system whose outputs staff cannot explain; offering a remedy only to customers who complain; answering an allegation about outputs with a statement about inputs; and one firm accepting the brand benefit while the other absorbs the regulatory duty.
Then rank them using the six determinants of issue intensity (Exhibit 6.7). Present the ranking as a table with a one-line justification per factor. This case contains a specific analytical prize, and strong reports find it: measure the issue intensity of this episode honestly and you will find it is low on several factors — the financial harm per person was modest, consequences were not immediate, and consensus that a wrong occurred was genuinely contested. Yet the reaction was enormous. Identify what drove that reaction, and state plainly whether it is one of Exhibit 6.7’s six factors or something the framework does not contain. If the framework is missing something, say what, and say what a manager should do about it.
Frameworks: Exhibit 6.7 issue intensity; definition of an ethical dilemma; the limits of a framework.
Map the stakeholders REQUIRED
Extend Exhibit 2. For each stakeholder record the stake, the power to affect the outcome, the legitimacy of the claim, whether their interest was represented, and how the episode ended for them. Then answer three harder questions:
- Who got a remedy, and why them? Both publicised couples were re-evaluated quickly. The remedy was available to anyone who asked — but asking required knowing that the decision was questionable. Assess a remedy system in which the effective trigger is an audience. Is “responsive to complaints” the same as “fair”?
- Are the front-line staff stakeholders or agents? They caused reputational damage with a sentence, and they had no other sentence available. Locate them in your map and assign responsibility accordingly.
- What did Apple owe? Apple made no credit decisions and held no banking licence. It also took the brand benefit, owned the customer relationship, and told customers the product was not a bank’s. Argue whether brand ownership creates ethical duties in the absence of decision authority — and commit to an answer.
Frameworks: Stakeholder identification and prioritization; power–legitimacy–urgency; Ch. 6 CSR expectations beyond shareholders.
Assess conduct against ethical criteria and CSR expectations REQUIRED
This question has three parts, all required.
(a) Apply the ethics tests. Take the five tests in Exhibit 6.4 and apply them to three decisions: (i) launching without joint accounts, (ii) telling customers that gender was not an input, and (iii) offering re-evaluation only on request. Present a 3×5 matrix with a verdict in each cell. Note that the Front Page test is unusual here: it was not a thought experiment, it actually ran. Use that. Did the decisions that failed the real front-page test also fail it in advance — and if so, why did nobody run it before launch?
(b) Classify the CSR posture. Classify your chosen organization’s conduct as social obligation, social responsiveness, or social responsibility, with evidence from the timeline. Apply Exhibit 6.2 to the November 2019 response: ethical or pragmatic? ends or means? long-term or short-term? Then apply it separately to the 2021 introduction of shared accounts, and explain why the same firm can sit in different categories eighteen months apart.
(c) Classical vs. socioeconomic view. “We complied with the law, the regulator agrees, and our obligation ends there” is a coherent position with a long intellectual history. State it in its strongest form using the classical view and Exhibit 6.1. Then state the socioeconomic reply. Finally, take a position: if legal compliance is not a sufficient ethical standard, what is the standard, and who sets it? An answer that criticises compliance without proposing a workable alternative is not finished.
Frameworks: Exhibit 6.4 ethics tests; Exhibit 6.2 responsibility vs. responsiveness; Exhibit 6.1; classical and socioeconomic views.
Explain the behavior: how does a careful organization do this? REQUIRED
Use Exhibit 6.5 to explain conduct rather than merely judge it. Neither firm is a careless organization; both are famous for rigour in their own domains. Work through the determinants:
- Structural variables. Ch. 6 identifies goals, performance appraisal, and reward allocation among the strongest influences on ethical behavior. A high-profile product launch has a date. Explain how launch-date pressure converts a feature omission (joint accounts) into an ethical exposure, without anyone intending harm or making a decision they would recognise as an ethical one.
- Issue intensity at design time. Before launch, the absence of joint accounts had low issue intensity for the people deciding it: diffuse, hypothetical, no identifiable victim. After launch it had a name and a face. Use this to explain why organizations systematically under-weight harms that have not happened yet — and connect it to Ch. 18 feedforward control.
- Moral development and individual characteristics. Locus of control is directly relevant: an employee who believes outcomes are determined by systems beyond their influence behaves differently from one who believes they can act. Apply this to the representative saying “it’s just the algorithm.”
- Where was the ethics function? Identify which Ch. 6 mechanism — code of ethics, ethics training, independent social audit, a protected channel for raising concerns — would have surfaced this before launch, and state honestly whether you think it existed and was ignored, or did not exist.
Frameworks: Exhibit 6.5 determinants; Exhibit 6.6 stages of moral development; values, ego strength, locus of control; codes of ethics and independent social audits.
Communication: the sentence that did the damage CHOOSE
Analyze this as a communication failure at three levels (Ch. 14). Downward: what did front-line staff need to know that they were not given, and what did the organization assume they would say? Upward: customer service heard this complaint before it became public — what would have had to be true for that signal to reach the people who could act on it? Identify the barriers to upward communication that were operating. External: evaluate the November 2019 public statement as crisis communication — it was accurate, prompt, and unsuccessful. Explain how all three can be true at once, and state what a better statement would have said without conceding a legal claim the firm believed to be false.
Frameworks: Ch. 14 communication process, barriers, direction of flow, informal networks; crisis communication; Ch. 12 job design and employee empowerment.
Control: what should have caught this, and when CHOOSE
Apply the three types of control (Ch. 18). Feedforward: what testing should have happened before launch — outcome testing across groups, an explainability requirement, a “can a representative answer this?” gate — and what would each have cost? Concurrent: the firm held the outcome data continuously from day one. What monitoring would have detected a pattern in the limit distribution months before a customer noticed, and why is it plausible that nobody was looking? Feedback: assess re-evaluation-on-request as corrective action against the standards of a well-designed control system: timeliness, coverage, criteria, and who is authorized to trigger it.
Then the harder question. Ch. 6 discusses independent social audits as a mechanism for checking conduct against the organization’s own standards. Design the algorithmic equivalent: who conducts it, what do they get access to, to whom do they report, and what happens if they find something? Be specific about the conflict of interest, since the firm being audited pays for the audit — the same problem that appears in financial auditing.
Frameworks: Ch. 18 control process and feedforward/concurrent/feedback control; corrective action; Ch. 6 independent social audits.
Structure: responsibility across an organizational boundary CHOOSE
One firm held the brand, the interface and the customer relationship; the other held the licence, the model, the risk and the legal duty. Describe this arrangement using Ch. 11 concepts — authority, responsibility, accountability, and the chain of command — and explain what happens when the chain of command ends at a company boundary that the customer cannot see.
Then evaluate the marketing line “created by Apple, not a bank” against your structural analysis. Was it an accurate description of the user experience, a misleading description of the decision-making, or both? Finally: design the governance the partnership should have had. Who has the authority to halt a launch? Who answers a customer? Who is accountable to the regulator for a decision made by a model one party built and the other party’s customers experienced?
Frameworks: Ch. 11 authority, responsibility, accountability, chain of command, boundaryless organizations; Ch. 9 brand as a strategic asset.
Decision making and the exception path CHOOSE
Credit limit assignment is the textbook example of a programmed decision: repetitive, high volume, resolvable by a rule. Automating it is obviously correct. Using Ch. 2, explain what happens when a programmed decision system receives a case that is genuinely non-programmed — a customer whose circumstances the rule does not capture — and no human anywhere in the organization has the authority to decide it.
Then identify the decision-making biases visible in the episode, with the behavior that evidences each. Consider at least: confirmation bias in the internal response to early complaints, availability bias in how seriously a complaint was taken once it had a famous name attached, and overconfidence in the assumption that excluding a variable guarantees a fair outcome. Close by comparing this to Case 01: both organizations automated a rule and left no legitimate route for an exception. Is that the same failure, or two different ones?
Frameworks: Ch. 2 programmed vs. non-programmed decisions, structured vs. unstructured problems, decision-making biases; Ch. 11 delegation of authority.
Your recommendation REQUIRED
(a) The November 2019 response. Place your team in the room on 10 November 2019. The allegation is public, the regulator has announced an investigation, and you believe — correctly, as it turns out — that the model does not use gender and that you have not broken the law. Recommend the response. Your recommendation must state: the decision criteria and their weights; who inside the partnership makes this call and who must sign it; what you say publicly, in words you are willing to have quoted; what you offer customers, and to which customers; what it costs and who pays; the strongest argument against your position, stated fairly and then answered; and the conditions under which you would change course. Note the real constraint — an apology that concedes discrimination would be both untrue and legally damaging, so “apologise” is not by itself an answer.
(b) The forward-looking design. Design the governance you would require before deploying any automated decision that materially affects a customer. Address at least: an explainability standard stated in terms of what a customer must be able to be told; outcome testing across groups, including who runs it and who sees the results; an appeal path that does not depend on the customer having an audience; the authority front-line staff need, and what they are permitted to say; a pre-launch gate with the power to stop a launch, and who holds it; and a protected channel for employees who believe a system is producing unjustifiable results. For each element, name the Ch. 6 mechanism it implements and the specific failure in this case it would have prevented. Then state its cost — explainability constrains model choice, and outcome testing finds problems you then have to fix.
(c) Transferability. In one paragraph: your recommendation in (b) applies to credit. State whether it should apply to automated decisions in hiring, insurance pricing, university admissions, or medical triage — and if your answer differs by domain, say what makes the difference. “It depends” is only acceptable if you say on what.
Frameworks: Ch. 2 decision-making process; Ch. 6 codes of ethics, ethics training, independent social audits, protection of employees who raise concerns; Ch. 18 control design; Ch. 14 crisis communication.
Optional extensions (for teams wanting a harder problem)
- International context (Ch. 4). Fair-lending law, data protection law, and expectations about automated decisions differ sharply across jurisdictions. Using Ch. 6’s treatment of ethics in an international context, ask whether a firm should apply one global standard or the local standard in each market — and what the UN Global Compact implies for the answer.
- Human resources (Ch. 12). Design the customer service role for a product whose decisions are automated: what authority, what information access, what training, and what escalation rights would the job need to be doable?
- Strategy (Ch. 9). Apple’s central strategic asset is customer trust. Evaluate the decision to put that asset behind a credit product whose decisions Apple did not control, using SWOT.
- Comparison with Case 01. Write the comparative analysis directly: two organizations, two automated rule systems, two absent exception paths, two opposite transparency regimes. What single principle survives both cases?
Part E — Deliverables
Written report
- Length: 3,000–4,000 words of body text (roughly 10–14 pages), excluding cover page, exhibits, and references.
- Language: English throughout.
- Format: A4, 12 pt, 1.5 line spacing, numbered pages and numbered section headings.
- Citations: Every factual claim and every number carries a source. Consistent style (APA 7th recommended). A reference list is mandatory.
- Contribution statement: a short table naming which member led which section, signed by all members.
Required structure
- Executive summary (max. 250 words, written last)
- Case background and unit of analysis (Q1)
- Ethics and social responsibility analysis (Q2–Q5) — the largest section
- Management analysis (your three chosen questions from Q6–Q9)
- Recommendation and reasoning (Q10)
- Limitations and what would change our answer
- References · Appendices · Contribution statement
Oral presentation
- Time: 15 minutes presentation + 5 minutes Q&A (confirm in class).
- Participation: every team member speaks for a meaningful segment.
- Slides: maximum 15. Any slide that is only a paragraph of text will cost you marks.
- Language: English throughout, including Q&A. Reading a script aloud is not a presentation — use notes.
Suggested timing
- 2 min — The dilemma in one sentence, and your recommendation stated up front
- 2 min — Case facts, including the regulator’s finding (do not bury it)
- 6 min — Ethics and stakeholder analysis (your core content)
- 3 min — Management analysis: communication and control
- 2 min — Recommendation, costs, and the strongest counter-argument
Prepare for these three questions — they will be asked: The regulator found no violation, so what exactly are you accusing them of? Who pays for your recommendation? Would you apply the same standard to a system that benefits you?
Sourcing and academic integrity
- Prefer primary sources: the New York DFS report, the companies’ own published statements, regulatory filings and enforcement documents. Contemporaneous journalism is acceptable when cited as such.
- This case is unusually easy to get wrong from secondary sources. A great deal of commentary written in November 2019 assumed discrimination had occurred; the regulator reported otherwise in March 2021. If you cite 2019 commentary, check whether the 2021 findings contradict it, and say so.
- The credit-limit ratios are self-reported allegations. Attribute them to the individuals who made them; do not present them as verified figures.
- Where a fact is contested or still developing — the later enforcement action, the current status of the partnership — say so. Correctly flagging uncertainty earns marks; asserting a disputed claim as fact loses them.
- If you use generative AI tools, disclose what you used them for in an appendix. You remain fully responsible for every claim, citation, and number in the report — including any the tool invented. This case is a particularly easy one for a tool to hallucinate confident numbers about.
Part F — Grading Rubric
Total 100 points, mapped to the four syllabus requirements. The final group report is worth 30% of the course grade.
| Pts | Criterion | Excellent | Weak |
|---|---|---|---|
| 15 | Case understanding & framing Req. (1) · Q1 |
Unit of analysis declared and held; the regulator’s finding stated accurately and early; the three mechanisms in Exhibit 4 kept distinct; technology explained briefly and only where it bears on the argument. | Drifts between Apple and Goldman; asserts discrimination as established fact; long tutorial on machine learning; treats anecdotes as data. |
| 30 | Ethics & CSR analysis Req. (2) · Q2, Q4, Q5 |
Four or more issues identified and ranked; the gap between measured issue intensity and actual reaction identified and explained; Exhibit 6.4 applied to multiple decisions with honest, conflicting verdicts; the classical position stated at full strength before being answered. | Frameworks named but not applied; moral verdict substituted for analysis; the compliance argument dismissed rather than engaged; every test conveniently agrees. |
| 15 | Stakeholder mapping Req. (2) · Q3 |
Goes beyond the given exhibit; power vs. legitimacy distinguished; the customers who never complained and the front-line staff both located and analyzed; Apple’s duty confronted directly. | Exhibit 2 reproduced with minor edits; only the two famous couples discussed; staff treated as culprits rather than stakeholders. |
| 15 | Management frameworks Req. (1) · Q6–Q9 |
At least three frameworks from Ch. 7–18 applied to generate insight the facts alone do not give; framework limits acknowledged where they strain. | Frameworks decorate the report rather than drive it; definitions copied from slides; conclusions would be identical without them. |
| 15 | Recommendation & reasoning Req. (3) · Q10 |
Specific and actionable; survives the constraint that the firm did nothing unlawful; costs named and allocated; the public wording actually drafted; strongest counter-argument answered; conditions for revision given. | Vague (“be more transparent”); costless; recommends an apology for something the firm did not do; forward-looking design not linked to any failure in the case. |
| 10 | English communication Req. (4) · report + presentation |
Clear, well-organized English; management and credit terminology used correctly; presentation delivered (not read) with all members contributing; Q&A handled with composure. | Meaning obscured by language errors; terminology misused; slides read aloud; one member carries the presentation; Q&A avoided. |
Note on fluency: this is an EMI management course, not a language examination. Grammatical errors that do not obstruct meaning are not penalized. Misusing a management concept is penalized, because that is a content error. The same applies to misusing a legal term: writing “discrimination” where you mean “unequal outcome” is a content error, not a language error.
Part G — EMI Toolkit
Terminology and language support for writing and presenting this case in English.
G.1 Key terms
| Term | Meaning in this case |
|---|---|
| Issuer | The bank that actually extends the credit, carries the risk, and holds the regulatory licence. Here: Goldman Sachs Bank USA — not Apple. |
| Underwriting | The process of assessing an applicant’s creditworthiness and deciding whether, and how much, to lend. |
| Credit limit | The maximum balance a cardholder may carry. The number at the centre of this case. |
| Protected attribute | A characteristic (sex, race, religion, age, marital status) that the law forbids using as a basis for a credit decision. |
| Disparate treatment | Treating someone less favourably because of a protected attribute. Requires the attribute to actually influence the decision. Not sustained here. |
| Disparate impact | A facially neutral practice that falls disproportionately on a protected group. May be challenged even without intent — which is why “we never used gender” does not end the argument. |
| Proxy variable | A permitted input that correlates so strongly with a protected attribute that it effectively stands in for it. |
| Adverse action notice | The notice a lender must give stating the specific principal reasons for an unfavourable credit decision. Note the assumption built into the law: that a reason exists and can be stated. |
| Explainability | The ability to state why a system produced a particular result, in terms the affected person can understand and challenge. Distinct from transparency, which is about disclosing the system’s existence or design. |
| Black box | A system whose inputs and outputs are observable but whose internal reasoning is not available for inspection — sometimes not even to its operator. |
| Appeal path / redress | The route by which an affected person can challenge a decision and obtain a remedy. Its absence is, on one reading, the entire case. |
| Joint account / co-signer | Arrangements letting two people be assessed and held liable together. Absent from this product at launch. |
| Compliance | Meeting the requirements of the law. Note carefully: a compliance finding answers a legal question, not an ethical one. Using the two words interchangeably is the single most common error in this case. |
| Enforcement action | A formal step by a regulator imposing penalties or required changes. Distinct from a report or finding, which may carry criticism without legal consequence. |
| Fiduciary duty | A legal obligation to act in another party’s interest. A credit card issuer generally does not owe one to an applicant — worth stating, because students often assume otherwise. |
| Remediation | Fixing the harm already caused, as distinct from fixing the process going forward. A complete recommendation addresses both. |
G.2 Useful language for the presentation
Being precise about the allegation
- “The regulator found no unlawful discrimination. Our argument is not that the law was broken.”
- “We are alleging an unequal outcome, not a prohibited input.”
- “This is a question of disparate impact, not disparate treatment.”
- “Compliance was achieved. Our claim is that compliance was not sufficient, and here is why.”
Conceding, then answering
- “The strongest version of the company’s position is… We accept that much. Where we differ is…”
- “It is true that two anecdotes are not evidence of a pattern. Our argument does not rest on them.”
- “We accept this imposes a real cost. We think it is justified because…”
Handling questions
- “Let me separate the legal question from the ethical one.”
- “Do you mean the November 2019 response, or the governance we propose?”
- “We don’t have evidence on the outcome distribution, so we would not want to speculate about it.”
Avoid these
- “The algorithm was sexist.” — states as fact what the regulator did not find
- “AI is biased.” — too vague to analyze; say which mechanism from Exhibit 4
- “They should be more transparent.” — not a recommendation; say what a customer must be able to be told, by whom, and when
Part H — Sources and Further Reading
Primary sources (start here)
- New York State Department of Financial Services, report on its investigation into Apple Card credit-limit practices, March 2021. Available from dfs.ny.gov. This is the single most important source in the case — read it before any commentary.
- Goldman Sachs’ public statements of November 2019 regarding Apple Card credit decisions.
- Apple’s Apple Card product and marketing materials from 2019, and the later announcement of shared Apple Card accounts.
- U.S. Consumer Financial Protection Bureau enforcement documents concerning Apple Card servicing (2024). Verify details before citing.
- The Equal Credit Opportunity Act and Regulation B, for the legal baseline on prohibited factors and adverse action notices.
Course materials
- Robbins, S. P. & Coulter, M., Management, 15th ed., Global Edition, Pearson, 2021 — Ch. 6 (Exhibits 6.1–6.8), Ch. 2, 3, 9, 11, 12, 14, 18.
Secondary sources
There is a large academic and policy literature on algorithmic fairness, explainable AI, and fair lending. Search your library databases for (“algorithmic fairness” OR “explainable AI”) AND (credit OR lending), and separately for “disparate impact” AND algorithm. Use this literature for the conceptual framework in Exhibit 4; return to primary sources for every fact about what these two companies actually did.
A note on the facts in this document. This case summarises the widely reported record. Several elements — the later enforcement action, the current status of the Apple–Goldman partnership, and the precise wording of the regulator’s conclusions — should be checked against primary sources before you rely on them, and some are still developing. Verify every figure and legal characterisation you use, and report disagreement where it exists. Do not cite this case document itself as the source of a fact in your report.