Some choices are easy to change in a document and expensive to change in real life. You can withdraw a price announcement, but customers may remember it. You can stop a trial, but people may already have organized their work around it. To decide how much process a choice deserves, ask what undoing it would actually require.
What Amazon's one-way and two-way doors mean
In his 2015 shareholder letter, Jeff Bezos distinguished consequential choices that are difficult to reverse from changeable choices that can use a faster process. He warned that treating every decision like the first kind can slow experimentation.
The useful question is therefore specific: if this turns out poorly, can we return to an acceptable position? “We can change our minds” is insufficient. Identify what money, time, trust or options would remain lost after the change.
Test what returning would cost
The Chair's checklist treats reversibility as a property to investigate, not a label to attach after deciding to move quickly. A short experiment may be reversible for the company while creating a serious inconvenience for someone else. Include those people in the assessment.
| Dimension | Question | Possible evidence |
|---|---|---|
| Money and time | What resources are spent even if we stop? | Setup costs, staff effort and the delay before another option is available |
| Customer expectations | What might people reasonably rely on? | Promises already made, work customers must redo and access they may lose |
| Commitments | What remains binding after we change direction? | Agreed terms, partner dependencies and responsibilities requiring expert review |
| Detection and recovery | Will we notice a problem early enough to respond? | A visible signal, a named owner and a tested recovery procedure |
Compare three versions of the same pricing idea
First, a team edits an internal draft of a new offer. No customer has seen it and no commitments depend on it. The immediate cost of changing the draft is mainly the team's time. A lengthy approval process may add little at this stage.
Second, the team offers a clearly described trial to a limited group. This can still be bounded, but customers now exist on the other side of the decision. What price was promised? How long does it last? Who handles a mistake? The team needs a plan for those obligations before describing the experiment as easy to undo.
Third, the company moves all existing customers to the new offer and reorganizes support around it. Restoring a previous setting will not necessarily restore lost goodwill, staff knowledge or the old operating process. The technical change may be small while the commercial commitment is large. These are fictional examples designed to expose the difference, not recommendations for a particular pricing policy.
Ask which missing information could change the choice
Bezos's 2016 letter suggested making many decisions with roughly 70% of the information one would like, while remaining able to detect and correct errors. The percentage is a management heuristic, not a measurable universal readiness threshold. It does not tell you which missing fact matters most in your situation.
Instead of counting how much research you have done, name the evidence that could reverse your recommendation. If the question is whether a proposed process can be restored within an hour, a recovery test may be more valuable than another broad market report. If the missing information is unlikely to change the decision, waiting for it may provide reassurance rather than insight.
Make a smaller commitment informative
A useful trial does two jobs: it limits the downside and answers an important uncertainty. Making something small without deciding what it should teach can produce activity without a decision. Define the smallest exposure that gives you relevant evidence, then decide who will look at that evidence.
Write a stop condition before the trial starts. A team that only agrees to “watch closely” can disagree later about whether the warning was serious. The condition need not be a precise number if that would be artificial; it can be a concrete event, such as a customer losing access that the offer promised to preserve.
- Question: which assumption is this trial meant to examine?
- Boundary: who participates, for how long and with what maximum commitment?
- Owner: who can stop it, and who handles the consequences?
- Signal: what observation would support continuing, changing or stopping?
- Recovery: what will happen to affected people if the trial ends?
A reversal can preserve direction without erasing the cost
Netflix's 2011 shareholder letter discussed both its pricing change and its announced, then withdrawn, separate DVD brand. The company itself judged the pricing change to have caused more customer damage. That history makes a useful distinction visible: changing one part of a transition does not require treating every part as a single indivisible decision.
Our interpretation is to assess the long-term direction, the rollout and the promises separately. You may be able to revise the sequence while preserving the objective. But the fact that a company eventually changed course does not mean the original step was costless or fully reversible for its customers. Describe what the correction actually repairs.
Sources & further reading
Follow the original materials behind this guide. Our analysis and exercises are The Chair’s interpretation. Read our editorial approach.
- Amazon: 2015 letter to shareholders
Page 5 discusses different processes for difficult-to-reverse and changeable decisions. This is Bezos's stated management approach.
- Amazon: 2016 letter to shareholders
The high-velocity decision section links speed with correction and presents the approximate information heuristic.
- Netflix: Q3 2011 letter to shareholders
Contemporaneous management discussion of pricing, customer reaction and the withdrawn DVD brand.
USE THE FRAMEWORK ON A REAL CASE