One-Way Doors and Two-Way Doors: The Reversibility Framework for Everyday Decisions

TL;DR. The one-way and two-way door distinction comes from Jeff Bezos’s 2015 letter to Amazon shareholders; his 2016 letter added the rule of deciding with “somewhere around 70%” of the information you wish you had. Research adds two warnings. People are too slow to walk through: in Steven Levitt’s coin-toss experiment (22,511 usable tosses), people told by the coin to make a change were 24.9 percentage points more likely to have changed two months later, and on important decisions those who changed were no worse off at two months and much better off at six. And open doors have a hidden cost: in Gilbert and Ebert’s studies, 66.3% of participants preferred a changeable choice, yet people who could change their minds liked their outcome less. A CEOtudent analysis of primary legal texts finds built-in undo windows for purchases, credit and insurance of 3 business days to 30 days. CEO move: score each decision on five factors and match decision speed to the score. Student move: treat each two-way door as an experiment with a learning goal and a closing date, then close the door yourself.

This guide belongs to the Decision Systems cluster of the personal decision stack. For rare, life-shaping one-way doors, pair it with the regret minimization framework and a pre-mortem. For two-way doors, set kill criteria before you walk through.

Where does the one-way door idea come from, and what did it claim?

The terms are best known from Jeff Bezos’s 2015 letter to Amazon shareholders, filed with the US Securities and Exchange Commission on 5 April 2016 as Exhibit 99.1 to a Form 8-K. The primary text is more precise than most summaries of it.

In a section on “one-size-fits-all” decision making, Bezos split decisions into two groups. Type 1 decisions are consequential and irreversible or nearly irreversible: one-way doors. They deserve a slow, careful, consultative process, because once you walk through you cannot get back to where you were. Type 2 decisions are changeable and reversible: two-way doors. If one turns out badly, you reopen the door and go back, so these should be made quickly by individuals or small groups with good judgment.

The failure he named: growing organisations apply the heavy Type 1 process to Type 2 decisions, producing slowness, unthoughtful risk aversion, too little experimentation and less invention. A footnote adds the opposite error, harder to see because companies that use the light process for Type 1 decisions tend to disappear before they get large.

The 2016 letter, filed on 12 April 2017, turned the idea into rules for “high-velocity decision making”: never use one process for every decision; make most decisions with about 70% of the information you wish you had (“If you wait for 90%, in most cases, you’re probably being slow.”); use “disagree and commit” to move without consensus; and escalate real misalignment early instead of letting exhaustion decide. It also makes the point that matters most for individuals: if you are good at correcting course, being wrong may cost less than you think, while being slow is expensive for sure.

Three cautions. This is a management heuristic from one executive, not a research finding; the 70% figure is a rule of thumb, not a measured optimum; and the frame was built for organisations. An individual is both organisation and decision-maker, which makes both errors easy: heavy process on small choices (procrastination that looks like diligence) and light process on large ones (impulse that looks like decisiveness).

What does the research say about reversible choices?

Four lines of evidence each test a piece of the same logic.

People overweight the door they are already standing behind. Samuelson and Zeckhauser’s 1988 paper on status quo bias (abstract only consulted) reports that most real decisions include a status quo option, that experimental subjects disproportionately stuck with it, and that faculty choices of health plans and retirement programs showed the bias in consequential settings. A 2021 preregistered replication by Xiao, Lam, Piara and Feldman (n = 311 and n = 316) found strong support in three of the four original scenarios, but not in the fourth.

When a coin pushes people through a door, they tend not to regret it. Levitt’s experiment (Review of Economic Studies, 2021) invited people facing a hard decision to flip a coin on a study website. Roughly 23,500 tosses took place; 22,511 were usable. The most popular important question was “Should I quit my job?” with 2,186 tosses, followed by “Should I break up” with 1,686. People who got heads were 24.9 percentage points more likely to report a change at two months; the six-month estimate was slightly smaller (0.211, or 21.1 points). For important decisions, the paper concludes that those who made a change were no worse off after two months and much better off six months later, and that people may be “excessively cautious” with life-changing choices. Levitt calls the results merely suggestive: the sample was self-selected, happiness was self-reported, and completed surveys fell from 13,935 at two months to 8,159 at six.

The ability to change your mind can lower satisfaction. Gilbert and Ebert (2002) ran three studies with Harvard undergraduates. In Study 1, 64 photography students each kept one of two prints they had made; some could swap later, until Day 5. Those who could change their minds liked their prints less, and when phoned on the fourth day, only 1 participant asked to exchange. In Study 2a (60 students, art posters), people with a final choice grew to like their poster more; people free to swap during the next month did not. In Study 2b, 89 students read the procedure: 66.3% preferred the changeable version, and 84.3% thought the average student would too. People want the open door and do not foresee that it stops them from settling. The authors add the exception: changeability is worth having when consequences are unclear and the odds of a dreadful mistake are high. Samples were small and stakes low, and CEOtudent did not locate a direct replication, so treat this as a mechanism rather than an effect size.

Walking through a one-way door destroys an option, and the option has a price. In a paper later published in the Journal of Economic Literature (consulted as NBER Working Paper 3307), Robert Pindyck compares an irreversible investment to a financial call option: investing “kills” the option to wait for new information, and the lost option value is part of the cost. In his example, a factory costs $800, the product sells for $100 today, and next year the price moves to $150 or $50 with equal probability, discounted at 10 percent. Investing now has a net present value of $300; waiting a year and investing only if the price rises is worth $386, so the flexibility is worth $86, about 29% of the naive figure. Pindyck credits Arrow and Fisher (1974) and Henry (1974) with first making the point for environmental decisions, and states the limit: the less time there is to delay, and the greater the cost of delaying, the less irreversibility should matter.

Study	Sample	Key figure as printed	What it means for door decisions
Levitt (2021), Review of Economic Studies	22,511 usable coin tosses, online experiment	Heads 24.9 percentage points more likely to report a change at two months; 0.211 at six months	People on the margin of a big change stay put too often
Gilbert and Ebert (2002), JPSP, Study 2b	89 Harvard undergraduates	66.3% preferred the changeable version; 84.3% expected others to	People choose open doors without pricing their psychological cost
Gilbert and Ebert (2002), Study 1	64 photography students	Swap window closed on Day 5; only 1 participant asked to exchange	In this study, almost nobody used the undo
Xiao et al. (2021), Meta-Psychology	n = 311 and n = 316	Status quo bias supported in three of four replicated scenarios	The pull of the current door is robust, not universal
Pindyck (1990/1991), NBER / JEL	Worked numerical example	Invest now: $300; wait and decide: $386; flexibility: $86	Waiting pays when the door is truly one-way and waiting reveals something

Sources: the papers named in each row; figures as printed. The Study 2b total of 89 is the sum of the 55 women and 34 men the paper reports; the 29% ratio is a CEOtudent calculation.

How often do people actually walk back through the door?

Many decisions that feel final are reversed routinely.

Decision	How often it is reversed	Population and period	Source
College major	About 30% of students with a declared major changed it at least once within 3 years; about one-third in bachelor’s programs, 28% in associate’s; about 1 in 10 changed more than once	First-time students who entered in 2011-12, measured in 2014 (about 25,000 in the study)	NCES Data Point 2018-434
Job	12.9 jobs on average from ages 18 to 58; over 40% of them held from ages 18 to 24	9,964 people born 1957-1964, interviewed 1979 to 2022-23	BLS, NLSY79 release (August 2025, reissued July 2026)
Job started at 45-54	21% of these jobs ended in less than a year	Same cohort	BLS, NLSY79 release
Quitting, this month	3.1 million quits in August 2026; quits rate 1.9%	US total nonfarm, seasonally adjusted	BLS JOLTS, release of 29 September 2026

Source: figures as printed by NCES and BLS. The NCES cohort entered college in 2011-12; later cohorts may differ.

A major or a job often feels like a one-way door. The data says these are reversed often enough to be better described as doors with a heavy hinge: reversible, at a real cost in time, credits or income.

How long do institutions give you to undo a decision?

Legislators have already done a version of door classification: where people decide under pressure or with poor information, the law builds in an undo window. Read side by side, those windows show how short reversal periods are and what closes them. This is general information, not legal advice; national implementations differ, and every rule has exceptions not shown here.

Decision	Built-in undo window	Jurisdiction	Legal source	What it implies for your own door classification
Buying online or away from business premises	14 days; extended by 12 months if the trader did not tell you about the right; member states may allow 30 days after unsolicited home visits	EU	Directive 2011/83/EU, Arts. 9, 9(1a), 10	Cheap door, short window: decide fast, but note the closing date
The same purchase in the UK	Ends 14 days after the contract (services, digital content) or after the goods arrive	UK	Consumer Contracts Regulations 2013, reg. 30	Same logic outside the EU
Digital content you have started	Right lost once supply begins with your prior express consent and acknowledgement	EU	Directive 2011/83/EU, Art. 16(m)	Pressing “start now” can close the door
Door-to-door sale ($25 or more at home, $130 or more elsewhere)	Until midnight of the third business day	US	16 CFR 429.0 and 429.1	Pressure sales get a short timeout, not a long one
Consumer credit agreement	14 calendar days; under the new directive applied from 20 November 2026, generally ending 12 months and 14 days after signing if the terms and information were never received	EU	Directive 2008/48/EC, Art. 14; Directive (EU) 2023/2225, Arts. 26, 47, 48	Borrowing is a two-way door for two weeks only
Consumer credit agreement in the UK	14 days	UK	Consumer Credit Act 1974, s. 66A	Same two-week logic
Credit secured on your principal dwelling	Until midnight of the third business day; up to 3 years if the notice or key disclosures were not delivered	US	12 CFR 1026.23 (Regulation Z)	Home-backed debt gets a three-day pause before it locks
Individual life insurance	Between 14 and 30 days, set by each member state; member states may exempt contracts of six months or less	EU	Directive 2009/138/EC (Solvency II), Art. 186	Long commitments get slightly longer windows
Timeshare or long-term holiday product	14 calendar days; up to 1 year and 14 days, or 3 months and 14 days, if forms or information were missing	EU	Directive 2008/122/EC, Art. 6	High-pressure, long-tail products get the strongest extensions
Starting a new job (probation)	A probationary period may not exceed six months, with limited exceptions	EU	Directive (EU) 2019/1152, Art. 8	Employment has the longest trial window in the table
Leaving or ending a job in probation, Germany	Two weeks’ notice during an agreed probation of up to six months, versus a four-week basic notice period	Germany	BGB, section 622(1) and (3)	The earliest stage of a commitment is the cheapest exit
Consent to data processing	Withdraw at any time, but withdrawal does not affect the lawfulness of processing before it	EU	Regulation (EU) 2016/679 (GDPR), Art. 7(3)	Some doors only reverse going forward

CEOtudent analysis of primary legal texts (EUR-Lex, eCFR, legislation.gov.uk, gesetze-im-internet.de), retrieved 18 October 2026. General information only; not legal advice.

Four patterns stand out.

Windows are short. For purchases, credit and insurance, built-in undo periods run from 3 business days to 30 days; only employment probation runs to months. “I can always undo this later” is rarely backed by any rule.

A door you do not know about stays open longer. When the trader fails to tell you about the right of withdrawal, the EU purchase window runs 12 months beyond the original 14 days, about 27 times longer (a CEOtudent calculation). Timeshare windows stretch to a year and 14 days, US home-secured credit to 3 years. The personal lesson: the first step in classifying any decision is to find the undo mechanism and its deadline.

Acting can close the door. Starting a download, or having a service fully performed, can end the right of withdrawal.

Some reversals work only going forward. Withdrawing consent stops future processing but does not undo what was already done lawfully. Many “reversible” decisions are really stoppable: you can halt them, not unmake them.

How do you tell a one-way door from a two-way door?

Bezos’s letters give two categories but no sorting test. The test below is a CEOtudent editorial framework built from the evidence above, not a validated instrument.

The CEOtudent Door Test. Score each factor 0, 1 or 2 and add them up (maximum 10).

Factor	Question to ask	0	1	2
Cost to undo	What would it cost in money, time, reputation or relationships to get back to where you are now?	Trivial	Noticeable, recoverable within months	Large or impossible
Lock time	How long does the door stay open?	Indefinitely	A known window (days to months)	It closes the moment you act
Option value lost	Which future choices disappear if you go through?	None	Some, with substitutes	A unique option
Learning speed	How soon will trying it tell you whether it was right?	Days or weeks	Months	Years, or never clearly
Others bound	Who else does the decision commit?	Only you	A few people who agreed	Dependants, a team, customers, or third parties whose data or trust is involved

Veto rule: a 2 on cost to undo or on others bound means one-way handling, whatever the total.

Then match the process to the score.

Score	Door type	Process	Time budget	CEO move	Student move
0-3	Two-way door	Decide alone or with one trusted person, at roughly 70% of the information you wish you had	Same day or same week	Decide, commit, set a review date	Name what you expect to learn; check it at the review date
4-6	Hinged door	Look for a cheaper version first: a trial, a pilot, renting before buying, a probation period, a staged commitment	Inside the undo window, with one structured review	Negotiate an exit clause or staged commitment before you sign	Log the decision and its lock date in a decision journal
7-10, or any veto	One-way door	Slow, written, consultative; run a pre-mortem and a regret test	As long as new information still arrives cheaply	Own the call; once decided, “disagree and commit”	Study base rates and people who already went through this door

CEOtudent editorial framework. The time budgets are judgment aids, not research findings.

Three questions from Levitt’s experiment show the spread. “Should I try online dating?” scores near zero: cheap to undo, no lock, fast learning. “Should I quit my job?” usually lands in the hinged middle: the cost to undo is real but recoverable, learning takes months, and a few people are bound, so make it more reversible (line up the next role, leave on good terms) and decide within a set window. “Should I have a child?” triggers both veto factors. These are illustrations for a typical case; your scores depend on your circumstances.

Which doors lock over time, and which only look locked?

Doors that lock. Legal windows expire on a date, and options expire too: Pindyck’s examples include a patent or mineral lease about to run out. In ordinary life the equivalent is an offer with a deadline or an opportunity tied to a life stage. Inaction compounds as well: in Levitt’s data, the coin influenced only decisions made within the first two months after the toss, and later changes were unrelated to it. The nudge worked early or not at all. Practical rule: beside every hinged decision, write the date on which it locks.

Doors that only look locked. If many people like you made a decision and a large share reversed it, as with majors and jobs, it is probably not a one-way door.

The opposite error. Bezos’s footnote names it at the company level. For individuals it shows up as “I can always undo it” applied to things that bind other people: co-signing a loan, a public statement, sharing someone else’s information. The veto rule exists for this case.

Close the door behind you. Gilbert and Ebert’s finding suggests one more habit: once the trial behind a hinged door is over, stop reopening it in your head. Keep a door open only while you are learning from it; after that, it mostly costs satisfaction.

What changes when AI is part of the decision?

AI makes many steps cheap to try and cheap to throw away: a draft, a prototype, an analysis or a first pass at code. In door-test terms, AI lowers the cost to undo and raises the learning speed at the drafting stage, so more early decisions become two-way doors. That favours the Student side of the lens: run more small experiments and learn from each.

The irreversibility moves to the release step. A sent message, a published article, code shipped to users, an AI agent allowed to pay or delete, and personal data uploaded to a service are all harder to reverse than the drafts behind them. The legal texts show the shape of the problem: EU law lets people withdraw consent at any time, but withdrawal does not undo processing that already happened lawfully.

The CEOtudent rule is draft freely, release deliberately. Treat generation as a two-way door and apply the door test separately to the release. When an AI tool or agent would take an action that scores 7 or more, or 2 on others bound, keep a human checkpoint before that step. The boundaries for that handoff are covered in should you let AI make the decision?.

What this does not mean

• Not “always move fast.” Levitt’s participants were close to indifferent and willing to let a coin decide. The evidence says marginal deciders tend to be too cautious, not that anyone should quit a job.
• Not legal advice. The table lists headline rules; exceptions and national implementations apply, and rules change (the EU consumer credit directive changes on 20 November 2026).
• A heuristic, not a law of nature. The Amazon frame comes from one company’s leadership, and its own footnote flags survivorship.
• Small studies, real mechanism. The Gilbert and Ebert samples were 64, 60 and 89 students choosing prints and posters; the effect size in high-stakes decisions is unknown.
• Scores are aids to judgment. The door test structures a conversation with yourself; it does not measure anything.

Frequently asked questions

What is a one-way door decision?
A decision that is consequential and irreversible or nearly so; Bezos’s 2015 Amazon letter called it Type 1 and said it deserves a slow, careful, consultative process.

Where do the terms one-way door and two-way door come from?
They are best known from Jeff Bezos’s 2015 letter to Amazon shareholders, filed with the SEC on 5 April 2016. The 2016 letter extended the idea into rules for fast decision making.

Is the 70% information rule backed by research?
No. It is a rule of thumb from the 2016 letter: a reminder that waiting has a cost, not a measured threshold.

Should I keep my options open as long as possible?
Only while waiting reveals useful information cheaply. Flexibility has value for irreversible decisions (Pindyck), but keeping a choice changeable can lower satisfaction (Gilbert and Ebert). Keep a door open for a defined learning period, then close it.

Is quitting a job a one-way door?
Usually not; it is closer to a hinged door, reversible at a cost. In the US, people born 1957-1964 held 12.9 jobs on average from ages 18 to 58, and in Levitt’s experiment people nudged into important changes such as quitting reported being better off six months later.

How long do I have to cancel an online purchase in the EU or UK?
Generally 14 days under Directive 2011/83/EU and the UK Consumer Contracts Regulations 2013. The EU directive lists exceptions, including personalised goods, unsealed hygiene products and digital content already started with your consent; the UK regulations have their own detailed rules. This is general information, not legal advice.

How does AI change reversibility?
Drafts and experiments become cheap to reverse; sending, publishing, shipping and sharing data do not. Apply the door test to the release step.

Sources

1. Amazon.com, Inc. (2016). 2015 Letter to Shareholders. Exhibit 99.1 to Form 8-K, US Securities and Exchange Commission, 5 April 2016, accession no. 0001193125-16-530910. Company-authored document.
2. Amazon.com, Inc. (2017). 2016 Letter to Shareholders. Exhibit 99.1 to Form 8-K, US Securities and Exchange Commission, 12 April 2017, accession no. 0001193125-17-120198. Company-authored document.
3. Gilbert, D. T., and Ebert, J. E. J. (2002). Decisions and revisions: The affective forecasting of changeable outcomes. Journal of Personality and Social Psychology, 82(4), 503-514.
4. Levitt, S. D. (2021). Heads or tails: The impact of a coin toss on major life decisions and subsequent happiness. Review of Economic Studies, 88(1), 378-405. Advance-access version consulted; earlier version NBER Working Paper 22487 (2016).
5. Samuelson, W., and Zeckhauser, R. (1988). Status quo bias in decision making. Journal of Risk and Uncertainty, 1(1), 7-59 (abstract only).
6. Xiao, Q., Lam, C. S., Piara, M., and Feldman, G. (2021). Revisiting the status quo bias: Replication of Samuelson and Zeckhauser (1988). Meta-Psychology, 5.
7. Pindyck, R. S. (1991). Irreversibility, uncertainty, and investment. Journal of Economic Literature, 29, 1110-1148. Consulted as NBER Working Paper 3307 (1990).
8. National Center for Education Statistics (2017). Beginning College Students Who Change Their Majors Within 3 Years of Enrollment. Data Point NCES 2018-434. US Department of Education.
9. US Bureau of Labor Statistics (2025, reissued 2026). Number of Jobs, Labor Market Experience, Marital Status, and Health for those Born 1957-1964. News release USDL-25-1322.
10. US Bureau of Labor Statistics (2026). Job Openings and Labor Turnover, August 2026. News release USDL-26-1547; series JTS000000000000000QUR.
11. Directive 2011/83/EU on consumer rights, consolidated text of 28 May 2022, Articles 9, 10 and 16. EUR-Lex.
12. The Consumer Contracts (Information, Cancellation and Additional Charges) Regulations 2013, SI 2013/3134, regulation 30. legislation.gov.uk.
13. US Federal Trade Commission. Rule Concerning Cooling-Off Period for Sales Made at Homes or at Certain Other Locations, 16 CFR Part 429, sections 429.0 and 429.1. eCFR.
14. Directive 2008/48/EC on credit agreements for consumers, Article 14; Directive (EU) 2023/2225 on credit agreements for consumers, Articles 26, 47 and 48. EUR-Lex.
15. Consumer Credit Act 1974, section 66A. legislation.gov.uk.
16. Regulation Z (Truth in Lending), 12 CFR 1026.23, Right of rescission. eCFR.
17. Directive 2009/138/EC (Solvency II), Article 186. EUR-Lex.
18. Directive 2008/122/EC on timeshare and long-term holiday products, Article 6. EUR-Lex.
19. Directive (EU) 2019/1152 on transparent and predictable working conditions, Article 8. EUR-Lex.
20. Bürgerliches Gesetzbuch (BGB), section 622. gesetze-im-internet.de.
21. Regulation (EU) 2016/679 (General Data Protection Regulation), Article 7. EUR-Lex.

CEOtudent analyses: the legal undo-window table (sources 11-21), the CEOtudent Door Test and its speed bands, and two derived figures (the 29% value of flexibility in Pindyck’s example; the roughly 27-fold extension of the EU purchase window). Deliberately not used: any share of decisions said to be two-way doors, which no source consulted measures; Levitt’s 0.5-point correlational happiness gap, because the paper describes the scale both as seven-point and as 10-point; and Dixit and Pindyck’s 1994 book, not consulted. Arrow and Fisher (1974) and Henry (1974) are cited only as credited by Pindyck.

---

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: