Two real decisions about whether to go ahead. In one, a small test saved a company from a mistake. In the other, there was no time to test at all, and waiting would have cost lives.
The whole pathway · Should we…? — the quick guide · The Price of the Pick →
Ten items. Say your answer out loud before you reveal it.
Read the sentence, guess the meaning of the word in bold, then reveal.
Three terms: a judgement call · the downside · sunk cost. Say the right one, then reveal.
One missing word fits every phrase in the group. Say it, then reveal.
Read each part. After it, answer the quick check from memory, then discuss one question. The figures are rounded, and the facts are as the cases were reported.
You may have met this case in The Hiring Tool That Learned the Wrong Lesson. This time, look at it through the lens.
In 2014, a team of machine-learning engineers at Amazon began building a program to read job applications. The company receives a very large number of CVs, and reading them takes a great many recruiters. The idea was simple: give the program a pile of CVs and let it pick out the best five, the way a shopper picks a five-star product.
To teach it what a good candidate looks like, the team fed it ten years of CVs sent to the company, together with what had happened to them. By 2015 the engineers saw that the scores were not gender-neutral for technical jobs. Most of the CVs came from men, so the program treated male CVs as the pattern of success. It marked down CVs containing the word “women’s”, and, according to people who spoke to the news agency Reuters, it downgraded graduates of two women’s colleges.
The engineers corrected the signals they found. The trouble was that they could not be sure what else the program had learned. According to the same sources, recruiters looked at its recommendations but never relied on its ranking alone. So the tool stayed small: one input among several, and easy to switch off. By the start of 2017, executives had lost hope in it and the team was disbanded.
1. What did Amazon want the program to do, and what data did it learn from?
2. What did the engineers discover in 2015, and why could they not simply fix it?
3. Why was it significant that recruiters never relied on the ranking alone?
1. Read a very large number of CVs and pick out the best ones. It learned from ten years of CVs sent to the company, and the outcomes.
2. The scores were not gender-neutral for technical jobs, because most of the past CVs came from men. They could remove the signals they found, but could not be sure the program had not learned other, hidden ones.
3. Because the tool was never the only way people were chosen. The decision to depend on it was never taken, so stopping cost little and the mistake was found before it did wide harm.
You may have met this case in Seven Deaths and a Hundred Million Dollars. This time, look at it through the lens.
At the end of September 1982, over roughly three days, seven people in the Chicago area died suddenly. Each had taken Extra-Strength Tylenol capsules shortly before collapsing, and tests showed that the capsules had been filled with cyanide. Tylenol was made by McNeil, a subsidiary of Johnson & Johnson, and it was the leading painkiller in the United States, with more than a third of the market.
Investigators soon established that the poison had not entered the bottles in a factory. Someone had taken packets from shop shelves, opened the capsules, added cyanide and put the bottles back. For several days nobody could say how many bottles were affected, or where. The question for the chairman, James Burke, and his team was not whose fault it was. It was what to do before anyone knew.
The company stopped all Tylenol advertising at once, set up telephone lines for worried customers and doctors, and spoke to the press openly. Then it went much further than anyone required. It withdrew Tylenol capsules from every shop in the country, around thirty-one million bottles, at a reported cost of roughly a hundred million dollars. The poisonings had happened in one city, and some argued that a national recall would destroy a product that could be saved. Within weeks Tylenol’s market share fell from over a third to well under a tenth. It came back about six weeks after the deaths, with tamper-evident packaging, and within about a year had won back most of its share.
1. Why did the company’s lack of blame not settle the question?
2. What did the company do beyond what anyone required, and what did it cost?
3. Why did some people argue against a national recall?
1. The poison had been added after the bottles left the factory, so the company was not at fault. But its name was on every bottle, nobody knew how many were affected, and the first calls came from journalists, not from its own systems.
2. It withdrew Tylenol capsules from every shop in the country, about thirty-one million bottles, at a reported cost of roughly a hundred million dollars.
3. The deaths were in one city, so a national recall could look excessive; officials reportedly worried it would show future criminals that such attacks worked; and it risked destroying a product that could otherwise be saved.
Both stories began with the same question: should we go ahead? At Amazon, the answer was to keep the tool small and keep testing it. At Johnson & Johnson, the answer was to act completely, immediately. Neither answer can be copied into the other case. The difference lies in what the facts allowed.
At Amazon, the decision was reversible: because the tool was never relied on alone, it could be stopped without harm. It could be tried small: the engineers could study its scores before anyone acted on them. And waiting cost little: nobody was harmed by not having an automatic CV reader. All three answers pointed to the same place. Keep it small, find out, and stop if it fails.
At Tylenol, none of that held. A death cannot be undone, there was nothing to test on a small scale because the risk was already in homes across a country, and every day of waiting added risk. The three answers pointed the other way: decide now. The lens does not say which case was easier. It says which questions to ask. And it leaves the real argument open: some people see the Tylenol recall as a triumph of values, while others point out that the company was clearly the victim and that doing right was affordable.
1. What were the three facts that made the Amazon decision a ‘keep it small’ case?
2. What changed in each of the three answers for Tylenol?
3. What does the lens not do?
1. It was reversible (never relied on alone), it could be tried small (scores studied before anyone acted on them), and waiting cost little.
2. Nothing could be undone, there was nothing to test on a small scale, and waiting added risk every day. So the answer was to decide now.
3. It does not say which case was easier or which decision was right. It tells you which questions to ask, and the real argument about values remains open.
A go / no-go question always sounds the same: should we do it? But it can be answered well only by asking three smaller questions first. They take about ten seconds, and between them they point to one of four recommendations. It is a thinking aid, not a formula: it helps you reach a provisional answer, and you should say what would change it.
Each situation answers the three questions. Which recommendation does the lens point to? Choose, then check.
1. Pilot — it is reversible, testable, and waiting is cheap, so a small trial costs little and teaches a lot.
2. Decide now — harm cannot be undone, there is no time for a trial, and waiting is the most expensive option.
3. Stage it — it is hard to undo and cannot be tried small, but it can be done in steps with a checkpoint between each.
4. Decline for now — it cannot be undone or tested, waiting costs little because nobody else is bidding, and the case is weak.
5. Pilot — it can be undone, it can be tested on one year group, and waiting for the results costs nothing.
6. Decide now — waiting increases the danger every hour, and the fix itself is reversible, so acting immediately is the lower risk.
7. Stage it — it is irreversible and too large to try small, but the half-at-a-time plan creates natural checkpoints.
A trial that cannot fail is a delay in disguise. Agree in advance what result would stop the rollout. If nothing could, you have already decided.
Sunk cost. Money already spent is not a reason to continue. Ask what you would decide if you were starting today with nothing spent.
You may be wrong about the waiting cost. People who want to act tend to overestimate it, and people who want to wait tend to underestimate it. Ask who is saying it, and what evidence they have.
A good recommendation names its conditions. This language lets you say when you would act, and what would change your mind, in a formal, compact way.
Rewrite each sentence without if. Write the missing first word.
1. If the pilot should fail, we would return to the old system. → the pilot fail, we would return to the old system.
2. If the change were reversible, I would approve it today. → the change reversible, I would approve it today.
3. If we had known about the risk, we would have waited. → we known about the risk, we would have waited.
4. If the board should refuse, we will need a second plan. → the board refuse, we will need a second plan.
1. Should — a possible future event, formal
2. Were — a hypothetical present or future situation (were for all subjects)
3. Had — a past situation, third-conditional meaning
4. Should — a possible future event
Write the verb in brackets in the correct form.
1. I would suggest that the board (wait) until the first results arrive.
2. Instead of (rush) the decision, we should set a deadline and a checkpoint.
3. The directors insisted that the trial (be) reviewed after six weeks.
4. If we change everything at once, we risk (lose) the customers we already have.
1. wait — base form after suggest that, even with a singular subject
2. rushing — -ing after the preposition instead of
3. be — base form after insist that
4. losing — risk is followed by -ing
Now give one recommendation about a real decision, in one breath, using either an inversion (Should it fail…) or a condition word, and finish with I’d change my view if…
Run the three questions in your head, then speak. Choose a task, start the timer, and talk.
Choose one: Should a company cut its prices to win a large customer? Should a hospital adopt a new treatment before the final trial results? Should a team change its working hours for a term? Say the kind of question it is, run the three questions aloud, then give your recommendation, one reason and the condition that would change your mind.
Describe two “Should we…?” questions from any field where the lens points in opposite directions. Say what is different about the three answers in each. Finish with the one fact that decided it.
Give a recommendation, then your partner says: “But what if you are wrong?” Answer in thirty seconds. Name the checkpoint or the result that would change your view, and say who would make that call.