On June 3, 1989, a 17-year-old in Indiana died while speeding to deliver a pizza for a Domino’s store. By mid-June his family and friends were circulating petitions against the company’s best-known promise: your pizza in 30 minutes, or $3 off. His mother called the guarantee “an invitation to break the law, to speed.”
Domino’s answered with arithmetic. That July it announced that a letter from its president, Dave Black, would go out with some 20 million pizza deliveries. A pizza was usually ready in 12 minutes, he wrote, and delivery areas averaged two miles or less, which “leaves our drivers nearly 18 minutes to make it to your home. After millions of safe deliveries, we know that’s plenty of time.”
Then the letter added something new. If you saw a Domino’s driver driving carelessly, here was a toll-free number to call.
The company now had two gauges: the clock on every order, and customers watching the road. The guarantee stayed, and the money stayed on the clock.
The clock and the road
Every number you push on needs a partner that shows what the pushing costs.
Intel’s Andy Grove made the case in his 1983 book High Output Management. As Robert Heller summarizes the book, indicators steer attention toward whatever they track, so a manager has to guard against overreacting to any one of them. His answer was to pair them, so that both an effect and its counter-effect get measured. His example was inventory. Cut stock to save money and nobody may notice the shortages that start holding up production, unless you track inventory levels and shortages side by side. He gave a second rule for output: put a measure of quality next to every measure of quantity.
A restaurant kitchen shows why. Time every plate from order to pass and the cooks will get faster, some of them by sending food out before it’s done. Count the plates sent back next to the times, and the two numbers start to argue with each other. That argument is the point. Neither number alone can tell you whether the kitchen got better or just got quicker.
At Domino’s, the second gauge had a weakness built in. It depended on strangers noticing a car, finding a number and calling it. The first gauge was printed on every promise, and customers could cash it in for $3. One number came with money attached. The other came with a phone number.
What a pair can’t do
Grove wrote this as practical advice for managers, not as a tested finding, and it has limits. Two numbers can both be gamed. A partner that nobody owns, or that moves too slowly to act on, is decoration. And a dashboard of twenty pairs is forty numbers, which is close to having none. A pair works best when you can name the trade-off in advance: this push will tempt people to cut that corner, so that corner gets its own count.
The hiring sprint
Imagine you lead recruiting at a 200-person company, and the chief executive wants time to hire cut from 45 days to 30. You can see how to get there: fewer interview rounds and faster offers.
You can also see what that push might cost, so you put a second number on the same slide: of the people hired this quarter, how many their manager would hire again after 90 days. You agree with the chief executive, before the quarter starts, that if the second number drops by more than a set amount, the 30 days gets reviewed with it.
Ninety days in, time to hire is down to 31 days. The second number has dropped too, past the line you agreed on, so the meeting opens with the review instead of the congratulations.
When the scoreboard counts only hits
AI models get scored on leaderboards, and most leaderboards, according to OpenAI, rank models on accuracy: the share of questions they get exactly right. In September 2025 the company argued that this is one reason models make things up. A model that guesses when it doesn’t know will sometimes guess right. A model that says “I don’t know” scores zero on that question.
It showed the gap with one of its own comparisons, on a test of short factual questions called SimpleQA. An older model, o4-mini, answered 24 percent of questions correctly. A newer one, gpt-5-thinking-mini, got 22 percent. On accuracy alone, the older model wins.
Now add the partner. The older model gave a wrong answer to 75 percent of the questions and declined to answer 1 percent. The newer one was wrong on 26 percent and declined on 52 percent. OpenAI’s proposed fix is to penalize confident errors more than abstentions, and give partial credit for saying you’re not sure.
That is one test, an example the company picked to make its point, and it says nothing about how often either model is wrong in your work. The habit carries over, though. When you try a model or a prompt on your own questions, record three counts, not one: right, wrong and “I don’t know.” Run each version several times, because answers vary from run to run, and check the wrong ones against what the task actually requires.
Domino’s kept its guarantee for more than four years after the letter. In December 1993, a St. Louis jury awarded nearly $79 million to a woman whose car had been hit by a Domino’s delivery car. Four days later, the 30 minutes were gone. The verdict, said the owner, Tom Monaghan, “was certainly the thing that put us over the edge.”