Kill It When the Last Good Judgment Gives Up
Set the experiment's stopping rule at judgment exhausted, not at first failed test.
- Difficulty
- Moderate
- Time to result
- ~months to results
- Steps
- 6
- Confidence
- 89%
Gurley's framing comes from a Bezos interview at the Code conference that he says he could watch over and over. Asked when an internal experiment gets killed, Bezos answered: when the last person with good judgment gives up. Gurley's point is structural rather than motivational. That is not how other big companies work — most run one test, and if it fails they quit. A startup cannot quit, because quitting means shutting down, so it runs experiment one and two and three and four and five, then pivots and runs six and seven and eight, staying up all night because it has to work. The result is that startups get way more shots on goal than big companies do, and Gurley names this as one of the core reasons startups can compete with big companies at all. The enabling condition is tolerance for hacky experiments. Gurley's own evidence is a conversation with an Uber driver about eight years earlier who had to be at an Amazon warehouse at 2:30, where they loaded his car with packages, handed him burner phones and a manifest, and booked the ride over Uber. Gurley called Uber immediately to brief them. His judgement: no other large company would run that project — none, zero — and most companies he has worked with past 20 or 30 million in revenue would refuse on accounting grounds alone.
Origin
Jeff Bezos articulated the stopping rule in a Code conference interview several years before this episode, answering a question about when Amazon kills an internal experiment. Gurley pairs it with his own first-hand sighting of Amazon's early same-day delivery test running on top of Uber.
Core principles
- 01Most large companies run one test, and if it fails they quit — that single rule explains most of why startups can beat them.
- 02A startup cannot quit, so it runs experiments one through five, pivots, and runs six through eight, accumulating far more shots on goal.
- 03The correct stopping rule is not a failed test but the exhaustion of informed belief: kill it when the last person with good judgment gives up.
- 04For this to work the organisation must tolerate hacky experiments that do not fit its accounting or process.
- 05Scale is not the constraint — Amazon at hundreds of billions in value ran a same-day delivery test using burner phones and rides booked over Uber.
- 06Gurley's counter-observation is that most companies past 20 or 30 million in revenue would refuse to run that experiment at all.
How to run it
- 1
Write the stopping rule before you start
Decide in advance that the project dies when informed belief is exhausted, not when a test fails. Adopting Bezos's formulation verbatim gives the rule a defensible shape.
Pro tip Documenting it upfront prevents the rule being invented retroactively to defend a pet project.
- 2
Name the people whose judgment counts
Identify who the people with good judgment on this initiative actually are, so the stopping condition is checkable rather than rhetorical.
Pro tip Keep the list small and senior enough that their concession is credible.
Watch out Without a named list, the rule becomes a licence for indefinite funding by whoever refuses to concede.
- 3
Remove the process veto on hacky tests
Explicitly pre-authorise experiments that do not fit normal accounting or systems. Amazon's same-day test used burner phones and rides booked on a competitor's platform.
Pro tip Give experiments a standing exemption from the finance and systems review that kills them at the idea stage.
Watch out Gurley notes most companies past 20 or 30 million in revenue block exactly this class of experiment because someone says we will not know how to do the accounting.
- 4
Run consecutive attempts on the same idea
Treat a failed test as one attempt of many. Run tests one through five, then pivot the approach and run six through eight, the way a startup that cannot quit would.
Pro tip Count and report shots on goal per idea as an explicit metric alongside outcomes.
- 5
Kill it when the last believer concedes
Stop when the final person with good judgment gives up, and stop decisively at that point rather than letting the project drift.
Pro tip Ask the last believer directly and on the record; ambiguity keeps zombie projects alive.
Watch out The rule cuts both ways — it is a kill rule, not just a persistence rule.
- 6
Institutionalise it beyond one leader
Build the rule into the organisational framework rather than relying on a founder's personal protection. Gurley's admiration for Bezos is that he built an organisational framework to run the whole company on what he believes.
Watch out A rule that depends on one executive's presence dies with their next role change.
In the wild
Gurley was in an Uber whose driver had to be at a San Jose Amazon warehouse at 2:30. There, Amazon loaded the driver's car with packages, gave him burner phones and a manifest, and booked the trip over Uber. The driver showed Gurley the manifest.
→ It was the early days of Amazon same-day delivery, run as a hack on top of another company's platform. Gurley called Uber to brief them and concluded no other large company would run that project — none, zero.
Asked at the Code conference when an internal experiment gets killed, Bezos answered that it gets killed when the last person with good judgment gives up.
→ Gurley treats the line as the clearest available statement of why Amazon accumulates far more shots on goal than peer companies of its scale.
Common mistakes
Using a single failed test as the kill signal
Running one test and quitting on failure is the default at big companies and, in Gurley's analysis, one of the main reasons startups can beat them. It converts a noisy signal into a permanent verdict.
Letting accounting or process veto the experiment
Companies past 20 or 30 million in revenue refuse hacky tests because they cannot classify them. The Amazon example ran on burner phones and a competitor's app precisely because it ignored that objection.
Treating the rule as permission never to kill anything
The rule specifies a genuine stopping condition. Without a named set of judges whose concession ends the project, it becomes an excuse for indefinite funding of a favoured initiative.
Is it for you?
Best for
Executives designing innovation processes at companies large enough that process and accounting friction can veto an experiment.
Not ideal for
Genuinely capital-intensive bets where each iteration is enormously expensive and open-ended persistence is ruinous.
From the transcript
“this is a company that's worth 100 billions of dollars that is running an experiment on top of uber”
“said when the last person with good judgment gives up and that's not how other big companies work they don't run experiments that way in…”
“he has built a organizational framework to take what Jeff Bezos believes and run the whole company that way”
From the episode
#651: Legendary Investor Bill Gurley on Investing Rules, Finding Outliers, Insights from Jeff Bezos and Howard Marks, Must-Read Books, Creating True Competitive Advantages, Open-Source Strategies, Adapting Mental Models to New Realities, and More