The honest guide to an invented customer ethics case study

This is an invented customer ethics case study. The company below does not exist and none of this happened. It is a mid-sized consumer subscription business, created so a decision can be followed from the idea to the aftermath without anyone's name attached.

The interesting part is not that someone proposed something questionable. It is that four people saw it, three had doubts, and the process gave none of them a place to put the doubt.

What to take away

What was proposed

The retention team had a problem: offers were expensive and were going to customers who would have stayed anyway. A data scientist proposed a model to predict who was likely to leave, so that offers could be targeted.

That much is ordinary. The specific proposal was narrower. The model would be trained on support call transcripts, which the company held to improve service quality. Among its strongest signals were phrases associated with financial difficulty: mentions of job loss, of struggling to pay, of asking whether the plan could be paused.

The team's plan was to use the prediction in both directions. Customers likely to leave and easy to keep would receive a discount. Customers likely to leave because they could not pay would receive nothing, because they were unlikely to convert and the offer would be wasted.

Why it was approved

The proposal went to a monthly meeting with four people in it: the retention lead, the data scientist, a product manager and someone from marketing.

Three of them had reservations. None of them had a way to express one that carried the same weight as the model's projected saving, which was specific, was on the slide, and had been checked.

The product manager said it felt uncomfortable, and the retention lead said the terms covered support data use for service improvement, which this was. They were not doing anything to those customers, just not spending money on them.

The marketing representative asked if it would look bad if it came out. The answer: nothing would come out, because the model was internal.

Read cold, each answer changed the subject. "It feels uncomfortable" met a legal point. "Would it look bad" met a claim about disclosure.

The one question that would have settled it, whether a customer would recognize their own words from a call, was never asked. Nobody in the room owned it.

The decision was recorded as approved with a note that legal had confirmed the terms. Legal had confirmed the terms. Nobody had asked legal the other question, and it was not legal's job to volunteer it.

What the room got wrong about the permission

The terms permitted use of support recordings to improve service. The reuse was to decide who received a commercial offer, and the strongest features were signals of financial distress.

A permission for one purpose does not cover another. Data protection law calls this purpose limitation, a core principle of the UK GDPR and the EU GDPR; California's CCPA, as amended by the CPRA, likewise limits use of personal information to the purposes people were told about. The further the second purpose sits from the first, the less the original agreement carries. The UK's Information Commissioner's Office, which supervises the UK GDPR, publishes guidance on what makes consent valid that turns on this question.

The invented company's jurisdiction is left open, so the rule depends on where it sells: in the UK, the UK GDPR and the Data Protection Act 2018, supervised by the Information Commissioner's Office; in California, the CCPA as amended by the CPRA, enforced by the California Privacy Protection Agency and the Attorney General; in the US more broadly, the Federal Trade Commission Act's ban on unfair or deceptive practices, enforced by the FTC. Under the UK GDPR, processing likely to result in a high risk to individuals requires a data protection impact assessment before it begins. Whatever the jurisdiction, the room treated a legal check as if it had answered an ethical question. It had not.

The US Federal Trade Commission publishes guidance for business on protecting personal information, and the UK's Information Commissioner's Office sets out the security obligations that come with holding records of this kind. Those duties are separate from the question of what the recordings may be used for.

The pattern is not hypothetical. The US Federal Trade Commission's action against Facebook over how data reached Cambridge Analytica, and the UK Information Commissioner's Office's investigation into the same episode, both turned on personal information being used in ways the people concerned had not been told about.

There is also a plainer test, and it takes ten seconds. Say the practice aloud in the words used internally: we use what customers tell us about losing their jobs to decide not to help them. Nobody in that meeting would have said that sentence to a customer, which is the whole finding.

How it was found

Six weeks later a support agent noticed something. A customer who had described a redundancy on a call, and asked about pausing the plan, had been given no retention offer, while a colleague's customer with similar tenure had received one. She asked her team leader why.

The team leader did not know, asked, and was told the model handled it. She escalated, not through a reporting channel, which she did not think applied, but by asking the product manager directly. The product manager had been the one who said it felt uncomfortable.

That is how most things of this kind are found: laterally, by someone with no authority, asking a question that has an awkward answer. It took six weeks because there was no route by which the discomfort in the room could become a question anyone had to answer.

What it cost

The company stopped the model within a week of the question reaching a director. The costs were smaller than they might have been and they were not zero. Because the case is invented, no real figure attaches to any of it; what travels to a real business is the shape of the cost, not an amount.

Every customer who had been scored on distress signals and withheld an offer had to be identified, which was possible, and decided about, which was harder. The company ended up offering the standard retention discount retrospectively to the affected group, which cost roughly what the model had saved, plus the work.

The larger cost was internal. Two of the three people with reservations had raised them and been answered, and they learned a reservation loses to a number, a lesson that generalises to every subsequent meeting.

The support agent learned that asking sideways works, and that the official channel had not occurred to her. Both lessons are the subject of employee ethics, and both are expensive in ways no ledger shows.

What changed afterwards

Four things, of which one mattered.

A written rule that a legal check is not an ethical clearance. Sensible and easy to write. It changed little on its own.

A requirement to state the use in the customer's words. Any proposal reusing data collected for another purpose had to include one sentence describing the use as a customer would hear it. This is cheap and it is the item most likely to have stopped the original proposal.

A distress route separated from the commercial one. Signals of financial difficulty were removed from the retention model entirely and routed instead to a support path with different handling and no offer attached. Where a product touches credit, insurance or essential services, duties toward customers in difficulty are usually spelled out by sector rules — in the UK, the Financial Conduct Authority's Consumer Duty and its forbearance requirements for consumer credit; in the US, the Consumer Financial Protection Bureau's rules against unfair, deceptive or abusive acts and practices. The first question a policy team should ask about a proposal of this kind is which regulator and which rule apply, and it should ask counsel to confirm that answer before the proposal moves.

The decision moved. Proposals that treat groups of customers differently stopped being approved in the retention meeting and went to a director whose target did not include retention spend. This is the change that did the work, and it is the same structural point made in leadership ethics: a person whose result depends on the answer cannot be the one to give it.

What to take from an invented case

The failure was not that a bad idea was proposed. Bad ideas are proposed constantly and most are stopped in the room.

This one was not stopped: the room could not turn discomfort into an argument that could beat a number, and had no route to send it anywhere that could.

If your organization would have approved this, the fix is not a values statement. It is the second and fourth items above; both can be in place next week. The asymmetry that makes these proposals attractive is in customer ethics; the general shape of the conflict is in ethical dilemmas.

Common questions

Was targeting retention offers the problem?

No. Targeting offers is ordinary. The problem was the feature set: using what people said about financial difficulty, gathered for another purpose, to decide who was worth keeping.

Would asking customers have been practical?

For this use, asking would have meant describing it, and describing it would have ended it. That is not an argument against asking. It is the answer arriving early and cheaply.

Does a smaller company need all four changes?

Which rules apply to reusing support call recordings? It depends on where the company sells. In the UK, the UK GDPR and the Data Protection Act 2018, supervised by the Information Commissioner's Office, with purpose limitation as a core principle and a data protection impact assessment required for processing likely to result in a high risk. In California, the CCPA as amended by the CPRA, enforced by the California Privacy Protection Agency. In the US more broadly, the Federal Trade Commission Act's ban on unfair or deceptive practices, enforced by the FTC.

The sentence in the customer's words, and moving the decision to someone without the target, are the two that carry the weight. The other two are useful and would not, on their own, have stopped anything.