TL;DR. Kill criteria are conditions you set before starting something that tell you when to stop. The idea has direct experimental support. In a 1992 study in the Journal of Applied Psychology, Itamar Simonson and Barry Staw compared techniques for reducing commitment to losing courses of action, and one of the three most effective was setting minimum target levels that, if not achieved, would lead to a change in policy. The research also contains a trap that most quitting advice misses. Chip Heath showed in 1995 that people who set budgets for their investments also de-escalate incorrectly, quitting in response to sunk costs, and that the budgets they set are based on the breakeven of total costs and total benefits, a total that includes what has already been spent. A kill criterion built on past spend is the sunk-cost fallacy running in reverse. The fix is to make every criterion forward-looking, attach a date, pre-commit the action, and give the trigger to someone who is not the most invested person in the room. For the date, a base-rate table we derived from 32 cohorts of US business openings shows the annual exit rate falls from 20.9% in year one to 9.0% by year five: most of the information arrives early, so your checks should too.
This piece is part of our personal decision stack. It is the exit-side companion to the pre-mortem protocol: the pre-mortem stops bad plans from starting, kill criteria stop running plans from outliving their evidence.
Why the quit decision belongs at the start
The moment you most need to decide whether to quit is the moment you are worst placed to decide it. You have invested time, money and reputation. You have told people. The information that the plan is failing arrives gradually, and each piece can be explained away.
Research on escalation of commitment has documented this for decades, and its most convincing evidence comes from outside the lab. Barry Staw and Ha Hoang studied professional basketball in a 1995 paper in Administrative Science Quarterly. They used the order in which players were picked in the draft as a measure of how much teams had invested in them. Teams gave more playing time to their most highly drafted players and kept them longer, even after controlling for on-court performance, injuries, trade status and position. The investment, not the performance, was driving the decision.
Hal Arkes and Catherine Blumer reported a similar pattern in 1985 with theatre season tickets: people who paid full price attended more performances than people randomly given reduced-price tickets.
A 2025 caution belongs here. Michał Białek and Emilia Biesiada tested the short hypothetical scenarios that much sunk-cost research uses, including classic ones from Arkes and Blumer, across two experiments with 395 participants. Internal consistency was poor, with omega values between 0.14 and 0.57, and scenarios that looked alike correlated only moderately. The lab vignettes are a weaker measuring instrument than textbooks imply. The field evidence, like the basketball data, is where confidence should rest, and it points the same way.
The practical conclusion is simple. If commitment grows with investment, the cheapest time to decide your exit conditions is before the investment exists.
What reduces escalation, according to the evidence
Table 1. What the research says about quitting well (verified data from the primary sources)
| Source | Design | Finding | What it means for kill criteria |
|---|---|---|---|
| Simonson and Staw, 1992 | Experiment comparing de-escalation techniques | Three most effective: making negative outcomes less threatening; setting minimum target levels that trigger a policy change if missed; evaluating people on their decision process rather than the outcome | Write the threshold first, make stopping safe, judge the process |
| Heath, 1995 | Lab studies, including one with real monetary incentives | People also de-escalate incorrectly in response to sunk costs; they set mental budgets based on the breakeven of total costs and total benefits; escalation is likely mainly when no budget is set or expenses are hard to track | Track resources, but do not let past spend set the stopping point |
| Heath, 1995 | Same paper | People are more willing to invest time than money to salvage a money sunk cost, and more willing to invest money than time to salvage a time sunk cost | Cap time and money separately |
| Levitt, 2021 | Randomized field experiment: people stuck on a decision flipped a coin | For important decisions such as quitting a job, those told by the coin to make a change were more likely to change, more satisfied and happier six months later | For big life decisions, the default error is excessive caution |
| Wrosch and colleagues, 2003 | Three studies: 115 undergraduates, 120 younger and older adults, 45 parents | The capacity to disengage from unattainable goals and to re-engage in new ones was associated with higher well-being, with interactive effects | A kill criterion needs a named next goal |
| Staw and Hoang, 1995 | Field data from professional basketball | Sunk costs predicted playing time, trades and survival after controlling for performance | The most invested person should not own the trigger |
Two numbers from the Levitt study help calibrate it. In the 2016 working-paper version, the experiment produced 22,511 usable coin tosses. For the decisions classed as important, a heads result raised the share of people who made a change by about 11 percentage points at two months. The coin did not decide for people. It nudged the undecided, which is precisely the population kill criteria are for.
The budget trap
Most advice on quitting says: set a limit. Heath’s paper shows why that advice is incomplete.
When people in his studies set a budget, they tracked investment against it and treated an exhausted budget as a reason to stop. That prevented runaway escalation. But the rule people used to set the budget was based on the breakeven of total costs and total benefits, and total costs include the money already sunk. The result was a second error: stopping a project whose future value justified continuing, because the past had used up the budget.
Put the two findings side by side and the design rule for kill criteria becomes clear:
- A limit on what you spend is a tracking device. It keeps expenses visible, which Heath identifies as the condition that prevents escalation.
- A signal about what happens next is a decision device. As long as the signal genuinely measures the future, sunk costs cannot distort it in either direction.
A good kill criterion therefore has both, and they do different jobs. “I will stop if I have spent six months” is a tracking device masquerading as a decision. “I will stop if, at six months, fewer than a stated number of paying customers have renewed” is a decision.
This is where the CEO and the student each earn their place. The CEO sets the limit and owns the call. The student insists the call rests on what the evidence says about the future, not on the size of the bill so far.
Setting the date: what base rates say about timing
A deadline is a bet about when useful information will arrive. For new ventures, public data answers that question.
The US Bureau of Labor Statistics publishes the survival of private-sector establishments by the year they opened, from openings in the year to March 1994 through openings in the year to March 2025. We recomputed the year-by-year exit rates from the raw surviving counts for all 32 cohorts and took the median at each age.
Table 2. Annual exit rate of new US private-sector establishments by age (CEOtudent editorial framework, derived from BLS Business Employment Dynamics survival data, cohorts opened 1994 to 2025)
| Age | Median share of the previous year’s survivors that exited | Range across cohorts | Median share of the original cohort still operating | Cohorts observed |
|---|---|---|---|---|
| Year 1 | 20.9% | 19.1% to 24.8% | 79.1% | 31 |
| Year 2 | 13.8% | 10.7% to 17.2% | 68.3% | 30 |
| Year 3 | 11.0% | 8.8% to 14.5% | 60.2% | 29 |
| Year 4 | 9.8% | 7.7% to 12.6% | 54.6% | 28 |
| Year 5 | 9.0% | 7.1% to 11.2% | 49.9% | 27 |
| Year 7 | 7.4% | 6.1% to 9.1% | 41.5% | 25 |
| Year 10 | 6.1% | 5.1% to 7.9% | 33.9% | 22 |
| Year 15 | 5.0% | 3.9% to 6.7% | 25.5% | 17 |
| Year 20 | 4.6% | 4.1% to 6.3% | 20.4% | 12 |
The most recent complete data point matches the long-run pattern. Of the 988,310 establishments that opened in the year to March 2024, 769,449 were still operating a year later, a first-year exit rate of 22.1%.
Three implications for the date field:
Front-load your checks. The first-year exit rate is 2.3 times the fifth-year rate and 3.4 times the tenth-year rate. If most of the sorting happens early, a single review at year three wastes the period when the signal is strongest. Put the first check within months, not years.
After the early years, survival alone is weak evidence. By year five, annual exit has settled near 9% and keeps drifting down. A venture that is still alive at that point has passed the steepest filter, but being alive tells you less each year. From here, the state part of the criterion has to carry the weight.
Use the curve as a prior, not a verdict. These are establishments, not founders, and no longer operating is not the same as failing: the table does not say why an establishment stopped. The table tells you how uncertainty is distributed over time. It does not tell you whether your project is good.
The same logic applies to jobs and strategies, even without a comparable dataset: ask when the first reliable signal will exist, and schedule the check for that moment rather than for a round number.
The CEOtudent Kill Criteria Protocol
The template below is our construction. Each field maps to a finding in Table 1 or Table 2.
Table 3. The eight fields of a kill criterion (CEOtudent editorial framework)
| # | Field | What to write | Rule | Evidence basis |
|---|---|---|---|---|
| 1 | Signal | An observable indicator of future value | Must describe what happens next, not what you have spent | Heath, 1995 |
| 2 | Threshold | The minimum level that justifies continuing | Written before you start; a number where possible | Simonson and Staw, 1992 |
| 3 | Date | When you check | Placed where the first reliable signal will exist; front-loaded for new ventures | BLS base rates |
| 4 | Action | Stop, shrink or change direction | Chosen before you see the result | Simonson and Staw, 1992 |
| 5 | Trigger owner | The person who calls it | Not the person most invested in continuing | Staw and Hoang, 1995 |
| 6 | Safe exit | What makes stopping survivable | Reputation, role and money protected in advance | Simonson and Staw, 1992 |
| 7 | Next goal | Where the freed time and money go | Named now, so disengaging has somewhere to go | Wrosch and colleagues, 2003 |
| 8 | Resource caps | Separate limits for hours and money | A tracking device, reviewed alongside the signal, never instead of it | Heath, 1995 |
Two operating rules make the template hold under pressure.
You may change a criterion only before its date. If new information makes a threshold look wrong, change it in writing, with the reason, while the outcome is still unknown. Changing it on the day you miss it is escalation with paperwork. A decision journal is the natural place to record this.
Judge the decision, not the result. Simonson and Staw found that evaluating people on their decision process rather than the outcome reduced escalation. If you pull a kill switch on schedule and the project would, in hindsight, have worked, you still followed a good process. Treat the result as data for setting the next threshold.
Three worked templates
The thresholds below are placeholders. The right numbers depend on your situation, and inventing a universal threshold would be exactly the kind of false precision this piece argues against.
Table 4. Kill criteria for a project, a job and a strategy (CEOtudent editorial framework, illustrative)
| Field | Side project or product | Job | Strategy, for example a new AI workflow |
|---|---|---|---|
| Signal | People who pay, or commit to pay | Growth in responsibility, skill or pay against the path you expected | Rework needed and time to a finished deliverable, compared with your old workflow |
| Threshold | At least [N] paying customers | [A named responsibility] or [a named skill] gained | Clearly better than the old workflow on [one named measure] |
| Date | [8 to 12] weeks after launch | Your next formal review, or [12] months | After [N] real deliverables, not practice runs |
| Action | Stop and publish what you learned | Start an external search while employed | Revert and document why |
| Trigger owner | An accountability partner | A mentor outside your reporting line | The person who receives the output |
| Safe exit | Framed publicly as an experiment from day one | Searching is not resigning | The old workflow stays documented until the check |
| Next goal | The second idea on your list | The role you would target | The next candidate tool or process |
| Resource caps | [X] hours a week and [Y] in spend | [X] hours a month on development outside work | [X] hours of setup time |
The strategy column matters more every year. New tools make it easy to start an experiment and awkward to admit it has not paid off, because the setup time feels like an investment. That is a sunk cost like any other. Our piece on when to let AI make the decision covers the other side of that boundary.
Where good thresholds come from
Kill criteria are only as good as the thresholds in them. Three sources are better than intuition.
Your pre-mortem. Every failure cause a pre-mortem surfaces comes with an earliest visible signal. That signal is a ready-made row for field 1.
A probability, stated in advance. Before starting, ask what chance of success would still justify the effort, and what observable result would push your estimate below it. Our guide to thinking in bets covers how to make that estimate honestly.
The assumptions underneath the plan. If the plan depends on a belief about customers, an employer or a market, the criterion should test that belief directly. The assumption audit is built for surfacing them.
When kill criteria are the wrong tool
When the signal arrives too late to use. Some worthwhile pursuits, like building deep expertise or a long research effort, produce outcome evidence only after years. Outcome thresholds will either fire too early or be set so far out that they never bind. Use learning milestones and resource caps instead, and review them on schedule.
When the decision is cheap to reverse. A kill criterion costs attention to design. For decisions you can undo tomorrow, decide and correct.
When you are looking for permission to quit. Heath’s premature de-escalation is real. If you find yourself tightening a threshold because continuing feels hard rather than because the evidence changed, the criterion is being used as an exit excuse. The test is the same as for escalation: was the change written down before the date?
Frequently asked questions
What are kill criteria?
Conditions, set in advance, that tell you when to stop or change a project, job or strategy. The term was popularised for individual decisions by Annie Duke’s 2022 book Quit. The supporting mechanism, setting minimum target levels that trigger a change if missed, was tested by Simonson and Staw in 1992.
Isn’t this just a way to give up sooner?
Not if the criteria are forward-looking. Heath’s research shows that limits anchored on past spending cause premature quitting. Criteria anchored on future value protect you from quitting too early as much as from staying too long.
Do people quit too much or too little?
It depends on the setting. In Levitt’s randomized study of important life decisions, people nudged toward change were happier six months later, which suggests excessive caution. In Heath’s budgeted lab tasks, people sometimes stopped too early. The design of the criterion decides which error you are exposed to.
How many kill criteria should one decision have?
One to three signals is a practical ceiling. With more, one of them will almost always look acceptable, and a mixed picture is the easiest thing to rationalise.
What if I hit the criterion but still believe in the project?
Then take the pre-committed action and, separately, write a new proposal as if starting fresh, with new criteria. If the case is strong without counting what you have already put in, it will survive that test.
Who should hold the trigger if I work alone?
Someone who gains nothing from you continuing: a peer, a mentor or an accountability partner. Tell them the criterion and the date in advance, and ask them to raise it on that date whether or not you do.
Sources
Simonson, I., and Staw, B. M. Deescalation strategies: a comparison of techniques for reducing commitment to losing courses of action. Journal of Applied Psychology, 1992, volume 77, issue 4, pages 419 to 426.
Heath, C. Escalation and de-escalation of commitment in response to sunk costs: the role of budgeting in mental accounting. Organizational Behavior and Human Decision Processes, 1995, volume 62, issue 1, pages 38 to 54.
Levitt, S. D. Heads or tails: the impact of a coin toss on major life decisions and subsequent happiness. The Review of Economic Studies, 2021, volume 88, issue 1, pages 378 to 405. Sample and first-stage figures from the NBER Working Paper 22487 version, 2016.
Wrosch, C., Scheier, M. F., Miller, G. E., Schulz, R., and Carver, C. S. Adaptive self-regulation of unattainable goals: goal disengagement, goal reengagement, and subjective well-being. Personality and Social Psychology Bulletin, 2003, volume 29, issue 12, pages 1494 to 1508.
Staw, B. M., and Hoang, H. Sunk costs in the NBA: why draft order affects playing time and survival in professional basketball. Administrative Science Quarterly, 1995, volume 40, issue 3, beginning on page 474.
Arkes, H. R., and Blumer, C. The psychology of sunk cost. Organizational Behavior and Human Decision Processes, 1985, volume 35, issue 1, pages 124 to 140.
Białek, M., and Biesiada, E. On the low reliability of sunk cost vignettes. Brain Sciences, 2025, volume 15, issue 8, article 808.
US Bureau of Labor Statistics. Business Employment Dynamics, Table 7: survival of private sector establishments by opening year, total private sector, data through March 2025.
Duke, A. Quit: The Power of Knowing When to Walk Away. Portfolio, 2022.
The exit rates, ranges and ratios in Table 2 were computed by us from the surviving-establishment counts in the BLS table and checked against the survival rates BLS publishes; every recomputed rate matched the published figure within rounding.
This content was compiled with the support of AI following in-depth research, then written and prepared for publication by the CEOtudent editorial team.
This post is also available in:
