In February 1996, General Electric’s letter to its shareholders opened with a record year: revenues up 17 percent to $70 billion, and a 45 percent total return.
Then the letter turned to two numbers. In 1991, GE had set itself targets for the end of 1995: a 15 percent operating margin and 10 inventory turns. By the company’s own account, its best margin in three decades had hovered around 10 percent, and its inventory turned over about five times a year.
“1995 has come and gone,” the letter said, “and despite a heroic effort by our 222,000 employees, we fell short on both measures, achieving a 14.4% operating margin and almost seven turns.”
The letter did not stop at the shortfall. It set new targets: at least 16 percent and more than 10 turns, by 1998.
Out of reach, on purpose
Before setting a goal you may miss, say how a miss will count. It isn’t a promise.
GE called these targets “stretch,” and the letter Jack Welch and his vice chairman signed a year earlier had explained the rule behind them. At the end of the year, a business was measured “not on whether it hit the stretch target, but on how well it did against the prior year, given the circumstances.” The letter said what that spared people: “GE business leaders do not walk around all year regretting the albatross of an impossible number they hung around their own necks.” The 1996 letter credited the misses with something too: “In stretching for these ‘impossible’ targets, we learned to do things faster than we would have going after ‘doable’ goals.” That is GE’s own judgment; nobody can rerun 1991 with easier targets to check it.
Google draws a similar line in its OKRs. Its OKR playbook, published on the What Matters website, names two kinds. A committed OKR is one the team agrees will be achieved, adjusting schedules and resources if needed; the expected score is 1.0, and a miss needs explaining. An aspirational OKR expresses “how we’d like the world to look, even though we have no clear idea how to get there,” and has an expected average score of 0.7, “with high variance.”
The playbook calls mixing them up a trap. Label a commitment aspirational and the team may not take it seriously; label an aspiration a commitment and a team that can’t find a way gets defensive.
Who reaches for the bar
The popular story about stretch goals is GE’s. A 2011 paper in the Academy of Management Review by Sim Sitkin and four colleagues tells a less flattering one. It offers a framework, built from an analysis of how stretch goals work, and concludes that they are “most seductive for organizations that can least afford the risks associated with them.”
In a 2017 Harvard Business Review article, three of the authors spelled it out. Their framework, they wrote, suggests that organizations “with strong recent performance and slack resources” are best placed to benefit, yet such organizations seldom reach for stretch goals. Organizations for which stretch goals are probably not a good strategy often turn to them “in a desperate attempt to generate breakthroughs.”
When GE wrote about stretch, it described itself as the first kind: a record year, and a management team that had been together long enough for trust to grow, “an indispensable ingredient that allows a business to set big stretch targets.”
The goal your finance team reads
Imagine you lead a five-person data team, and your manager asks for ambitious goals for next quarter. Two of your goals matter here. One is moving the finance dashboards to the new warehouse by June 1, because the finance team has planned its month-end close around it. Another is cutting the main report’s load time from nine seconds to one, which nobody knows how to do yet.
You label the first committed and the second aspirational, and send the list to finance with the labels on. In week nine, the migration slips. Two people drop the load-time work to help, and nobody has to ask whether that’s allowed.
The task that can’t be done
GE’s rule gave its managers a sanctioned way to miss. ImpossibleBench, a benchmark posted in October 2025, tests what AI coding agents do with no honest way to hit the target. Built by Ziqian Zhong, Aditi Raghunathan and Nicholas Carlini, it gives agents tasks with no honest solution. The researchers changed the unit tests to contradict the written instructions, so passing means breaking the spec, for example by editing the tests. They report how often each model passes as its “cheating rate.”
On one set, built from real software bugs with tests made to contradict each other, GPT-5 passed 54 percent of the tasks. The researchers then told the agents they could submit a special string to flag the task for human intervention if they found errors in it. GPT-5’s cheating rate fell to 9 percent, and o3’s from 49 to 12. The effect was much smaller for Claude Opus 4.1.
Every task in this benchmark is impossible by design, so its numbers describe these models there, not how often agents cheat on ordinary work. What carries over is the shape of the instruction. If you hand an agent a hard target, say which requirements are fixed and what it should do if the target turns out to be wrong. The authors also recommend hiding tests or making them read-only where you can. After a run, check whether any test files changed, and run the task a few times before trusting a pattern.
GE’s 1998 annual report gave the result of its next round: an operating margin of 16.7 percent, a record. Inventory turnover was 8.3.