Decision-Making Under Uncertainty: Mental Models for Founders

I made the worst decision of my founder life on a Tuesday. I remember it clearly because I remember how fast it felt. I had a customer on a Zoom call telling me a story about a workflow, and by the time the call ended I had already decided to rebuild our core product to fit that workflow. No second opinion. No probability check. No thought about whether I could come back if I was wrong. I told the team that afternoon. We spent four months on it. The customer did not buy. Neither did anyone else. We shipped a feature to an audience of one and he ghosted.

The post-mortem was embarrassing because the mistake was not the decision. It was the process. I had treated a one-way door like a two-way door. I had confused conviction with evidence. I had bet four months of the runway on a single N-of-1 signal and never even noticed I was making a bet.

That is what most founder decision-making looks like. Fast, emotional, single-input, and invisible to the person making it. And it is a quiet killer. 42% of startups fail because they chose the wrong problem to solve, according to CB Insights research on 483 startup post-mortems. Another 29% fail because the timing was wrong. Both are decision failures. Not execution failures. Not talent failures. The founder made a call under uncertainty and got it wrong, and the company died of it.

Good founders are not born with better judgment. They just have a different kit of mental models for deciding when to think fast, when to think slow, what to solve for, and what to ignore. This post is that kit.

What you will get from this post

• A framework that separates decision quality from decision outcome (they are not the same)
• 8 mental models that cover 90% of founder decisions, with a routing diagram for when to reach for which
• A reversibility matrix, an expected value worksheet, and an inversion checklist you can run in under 10 minutes
• The contrarian take on “data-driven decisions” that most founders get wrong
• A Monday-morning action plan to install this as a weekly habit

Why decision quality is the only founder skill that compounds

Here is an uncomfortable stat. Cognitive biases affect 91% of startup decision-making, and confirmation bias alone is implicated in 32% of product failures. 68% of founders who fail use phantom traction metrics. Poor financial modeling leads to unexpected cash crunches in 76% of shutdowns. The data is clear: startups do not die from bad luck. They die from bad decisions that looked reasonable in the moment.

Think about what that means. The average founder will make between 50 and 100 decisions per week. Hiring, pricing, positioning, product scope, channel selection, partnership terms, when to fundraise, when to shut down a line. Over a 3-year company lifetime, that is roughly 10,000 decisions. If your decision quality is 1% better per decision, you compound. If it is 1% worse, you compound the other way.

This is the thing no accelerator curriculum teaches well. Execution speed is overrated relative to decision quality. You can execute fast in the wrong direction for a very long time before the market notices. By the time the cash runs out, the decisions that killed the company were made months earlier, usually while everyone was feeling productive.

The first move: separate decision quality from decision outcome

The single most useful mental move I have made as a founder is learning to hold these two things apart.

Decision quality = the process you used to decide, given the information and time you had.

Decision outcome = what actually happened after.

You can make a high-quality decision and get a bad outcome. You can make a terrible decision and get lucky. If you judge yourself only by outcomes, you will learn the wrong lessons in both cases. Most founders are stuck in a feedback loop where they are either beating themselves up for bad luck or taking credit for good luck, and either way they are not getting better.

Annie Duke, the poker player turned decision coach, calls this “resulting.” It is the cognitive mistake of grading a decision by the outcome alone. In poker, you can play a hand perfectly and lose. In startups, you can pick the right market and still be early. In hiring, you can run a great process and get a bad fit because the candidate was going through a divorce you did not know about.

The fix is to build a decision journal. Not a fancy one. I use a Notion page with five columns: the decision, the date, what I knew at the time, what I chose, and my confidence level. Then I come back every 90 days and grade the process, not the outcome. I have been doing this for 3 years. It is the single highest-impact habit I have installed as a founder. It has rewired how I think about almost every call.

The Decision Quality Framework

Here is the framework I use before any decision that feels heavy. It has two axes and four quadrants, and it takes 30 seconds to place a decision on it.

The frame is simple. If your inputs are bad (you are working off rumors, sample sizes of one, or your own wishful thinking) and your process is bad (you decided in the moment, alone, without a model), you are gambling. This is where most founder decisions live. Not because founders are dumb. Because the default is gambling.

The target is the top right. Good inputs means you talked to the right number of customers, checked the data, and validated your assumption instead of your hope. Good process means you picked a mental model that fits the decision, ran it, and wrote down why.

When I look back at the Tuesday rebuild disaster, I was in the bottom left. A customer told me a story (bad input, N of 1), and I decided on the call (bad process). Had I been in the top right, I would have logged it, waited 48 hours, called 5 more customers to see if the pattern repeated, and asked my cofounder to argue the opposite. The decision might still have been the same. But the process would have given it a 30% chance of working instead of a 3% one.

The 8 mental models that cover 90% of founder decisions

You do not need 100 mental models. You need 8, and you need to know which one to reach for when. Here they are, in the order I reach for them.

1. Reversibility (two-way doors vs one-way doors)

Jeff Bezos laid this out in his 2015 Amazon shareholder letter. Two-way door decisions let you walk back if you are wrong. Launch a feature, try a pricing page, hire a contractor, run an experiment. If it does not work, you can return. These should be made fast by one person with maybe 70% of the information they wish they had.

One-way door decisions do not let you walk back cheaply. Sell the company, fire a cofounder, sign a 3-year lease, raise a priced round, ship an irreversible data migration. These deserve a full process, a calendar delay, and at least two opposing viewpoints.

The failure mode most founders hit is the inversion of this: they treat two-way doors like one-way doors (analysis paralysis on a landing page copy change), and one-way doors like two-way doors (signing a term sheet after one meeting). Both are expensive. The second is lethal.

Practical rule: if the decision is a two-way door, give it 70% of the info and ship. If it is a one-way door, give it 24 hours, write it down, argue the opposite, and then decide.

2. Expected Value (bet sizing, not certainty chasing)

Ray Dalio puts this best. Every decision is a bet. It has a probability of being right and a reward if it is. It has a probability of being wrong and a penalty if it is. A good decision has positive expected value. It does not have to be the most probable outcome. It has to have the right math.

Here is the move that trips up most founders. You do not always bet on what is most likely. You bet on what has the best reward-weighted probability. A 20% chance of a 10x return beats a 90% chance of a 1.1x return. The expected value is 2.0 vs 0.99.

This is why every smart VC bets on a portfolio of long shots. They are not trying to be right. They are trying to be mathematically positive across many bets. Solo founders have one portfolio of one company, so they should be even more careful about expected value. You do not get a second at-bat.

The simplest worksheet:

The trick is calibrating P(win). Most founders overestimate by 2x to 3x. The fix is brutal: look at your last 10 bets, grade them honestly, and apply your historical hit rate, not your optimistic one.

3. Inversion (what would make this fail?)

Charlie Munger got this from the mathematician Carl Jacobi. “Invert, always invert.” Instead of asking how to make the decision succeed, ask how to make it fail. Then avoid those things.

This sounds simple. It is shockingly powerful. I used it on a pricing change last year. The forward question was, “How do I justify raising prices 40%?” I was going to come up with 10 reasons. Classic confirmation bias. The inverted question was, “What would cause a 40% price hike to destroy the business?” That list surprised me. Churn spike, competitor undercut, positioning whiplash, support load from angry customers, trust damage with existing users. Five of those were serious. Two were fatal if they happened together.

I still raised prices. But I did it by 18%, grandfathered existing users, and pre-announced. Zero churn spike. If I had only asked the forward question, I would have raised 40%, lost 25% of the base, and spent a month on damage control.

Inversion works because it turns off confirmation bias at the cost of feeling uncomfortable for five minutes. That trade is always worth it.

4. Regret Minimization (the Bezos 80-year-old test)

Bezos used this in 1994 to decide whether to leave his Wall Street job and start Amazon. Project yourself to age 80. Look back at this decision. Will you regret having done it, or regret not having done it?

The framework is useful for one specific class of decision: the one where the downside is recoverable but the upside is a life arc. Starting a company. Moving countries. Leaving a stable job to go solo. Asking someone to work with you. Quitting.

It does not work for operational decisions. Do not run regret minimization on which payment processor to use. Do run it when the decision is between a safe path and a path that might change the shape of your life. The question cuts through noise because future-you is calmer and more honest than present-you.

I have used this exactly four times in my career. Every time it pointed to the risky path. Every time I was right.

5. Pre-Mortem (how will this die?)

Gary Klein’s technique. Before committing to a plan, assume it has already failed one year from now. Write the obituary. What killed it?

This works because teams will not tell you a plan is weak when you are asking them to support it. But they will tell you why it failed once you give them permission to imagine it has. Pre-mortems are also the single best tool against the planning fallacy, which is the reliable human habit of underestimating how long things will take by 40% to 60%.

I run a pre-mortem for every plan that commits more than one month of team time. 45 minutes. Everyone writes silently for 10 minutes, then reads out. Top three risks almost always show up in at least half the submissions. Those three become explicit mitigation plans in week one. If we cannot mitigate them, we do not start the plan.

6. OODA Loop (speed of iteration under pressure)

John Boyd, the fighter pilot. Observe, Orient, Decide, Act. Then loop. The edge is not in making the perfect decision. It is in cycling through the loop faster than the environment changes.

This is the model for startup environments where the ground keeps shifting. AI model releases, competitor moves, pricing wars, regulatory updates. You cannot pause and do a 6-month strategy offsite. You have to observe (what is actually happening in the market this week), orient (what does this mean for our position), decide (pick one move), act (ship it). Then watch the signal and loop.

OODA is not about being right every cycle. It is about reducing the lag between “reality changed” and “we moved.” Teams that cycle weekly beat teams that cycle quarterly. Teams that cycle daily beat teams that cycle weekly. This is why scrappy AI startups have been eating horizontal SaaS incumbents for two years. They cycle faster.

7. Second-Order Thinking (and then what?)

Howard Marks’ mental model. First-order thinking stops at “if I do X, then Y happens.” Second-order thinking keeps going: “and if Y happens, then Z happens, and if Z happens, then this other thing I did not expect happens.”

Most failed founder decisions are first-order right and second-order catastrophic. Hire a sales team to scale revenue? First-order: more revenue. Second-order: your gross margin drops 40%, your culture shifts to quota-driven, your product roadmap starts bending around short-cycle enterprise asks, and suddenly you are a different company. Cut your lowest-paying customers to focus upmarket? First-order: higher ACV. Second-order: you lost your word-of-mouth engine, the long tail that built your SEO.

The move is to always ask “and then what?” at least three times before committing. If you cannot answer the third “and then what,” you are making a first-order decision on a second-order problem. Stop.

8. System 1 vs System 2 (know which brain is deciding)

Daniel Kahneman’s two-system model. System 1 is fast, automatic, pattern-matching. System 2 is slow, effortful, analytical. Both are necessary. The mistake is letting System 1 decide things that need System 2, or slowing down System 2 for things System 1 handles just fine.

System 1 is right for decisions where you have deep pattern-match reps. An experienced founder knows what a bad hire smells like in the first 15 minutes of an interview. That gut is not magic. It is 500 interviews of compressed data. Trust it for what it is trained on.

System 1 is dangerous for decisions in new domains. A first-time founder’s gut on fundraising terms is worthless. No reps. No pattern. All noise. That is where System 2 has to take over, and where “just trust your gut” is terrible advice dressed up as wisdom.

The practical question: have I made this kind of decision 50+ times before? If yes, run fast, trust the gut, and check your journal later. If no, slow down, pull a framework, get a second opinion, and force System 2 to do the work.

Which model should you use when? A routing map

Here is the flow I walk through before any non-trivial decision. It takes about 90 seconds.

Notice the pattern. The first question is always reversibility, because that sets your clock. Two-way door? Move. One-way door? Stop. Everything else is downstream of that.

The second question is either domain familiarity (for reversible decisions) or magnitude of regret (for irreversible decisions). These two questions route you to the right model in under a minute, and they filter out 80% of the cases where founders overthink or underthink.

When to use which model: the cheat sheet

The Reversibility × Consequence matrix

Reversibility alone is not enough. You also need to weight by how bad the consequences are if you get it wrong. Here is the 2×2 I use as a gut check before any one-way door decision.

The bottom-right cell is where most company-ending mistakes happen. Signing a punitive investor term sheet to close a round you could have walked away from. Hiring a senior exec without backchannel references. Agreeing to an enterprise deal with exclusivity clauses. These are permanent on the company’s trajectory, and they deserve all five models running in sequence before you sign.

The contrarian take: “data-driven” is a trap

Every accelerator, every VC deck, every founder Twitter thread tells you to be “data-driven.” I am going to tell you the opposite. Data-driven is a trap for founders making decisions under uncertainty, and it is probably why you are making worse decisions right now than you would if you did not have a dashboard.

Here is why. Data is good for decisions where the past predicts the future. Optimizing a checkout flow on a mature product with 10,000 users per day? Data all the way. Deciding whether to enter a new market that does not exist yet? Data is useless, because there is no data about a future that has not happened.

What you actually need for most founder decisions is not data. It is a probabilistic belief. A belief with a number attached. Not “I think this will work.” Not even “I am pretty confident this will work.” Instead: “I think this has a 35% chance of working, and here is what would update me up or down.”

That is a different cognitive discipline than being data-driven. It is being calibrated. Superforecasters (the people studied in Philip Tetlock’s research who consistently beat intelligence agencies at predicting world events) are not the ones with the most data. They are the ones with the most honest probability estimates, updated most often.

The practical shift is this. Stop asking “what does the data say.” Start asking “what is my probability estimate, what would change it, and what evidence would update it fastest.” Then go hunt that evidence. That is a 10x better question for most founder decisions, because it turns you into a Bayesian instead of a data tourist.

The secondary problem with data-driven is that it is often a political shield. Founders who cannot defend a decision retreat to “the data says” to avoid owning it. This lets them off the hook for having a point of view, which is exactly what founders are paid to have. If your decision is just “what the data says,” your customers do not need you. They can hire a decent analyst.

Three decisions I got wrong and what model would have caught them

Case 1: The Tuesday rebuild

The one I opened with. Customer call, instant decision, four months wasted. What would have caught it: inversion (“what would make this rebuild fail?”) and expected value (the probability that one customer’s workflow generalized was maybe 15%, and the cost of being wrong was 4 months of runway). Either model, run for 10 minutes, would have killed the decision on the spot.

Case 2: The senior hire without backchannel

Hired a VP of Engineering from a good company. The process felt good: multiple interviews, reference checks from names they provided. Three months in, the team was in revolt. I had not done the backchannel. If I had called two engineers who had worked under them (not with them), I would have heard the same story I heard after firing them. What would have caught it: pre-mortem (“imagine this hire fails in 6 months, what killed it?”) plus reversibility weighting (senior hires are close to one-way doors because the cost to unwind is 3 to 6 months of team trauma).

Case 3: The SEO bet

I went all-in on SEO for 9 months in a product category where Google was about to release AI overviews. The overviews cratered traffic for long-tail queries by 40% to 70% across the category. My expected value math was right. I just had not done second-order thinking on “what if the platform itself changes.” I was solving for the game board I saw, not the one that was being drawn. OODA would have helped here too, if I had been cycling monthly instead of quarterly I would have caught the shift sooner.

The pattern is clear. In each case, the missing model was not exotic. It was one of the eight. And the time to run it was under 30 minutes. Three decisions, maybe 90 minutes of thinking total. Those 90 minutes would have saved about 18 months of wasted effort.

What to do Monday morning

Here is the install plan. This is what to do this week. Not next quarter. This week.

Monday: Open a Notion page. Call it “Decision Journal.” Five columns: date, decision, context, choice, confidence (1 to 10). Log every non-trivial decision you make this week. Aim for 10 to 20 entries.

Tuesday to Thursday: Before every decision that feels heavy, pause for 90 seconds and run the routing map. Ask: reversible? Stakes? Domain familiar? Then pick one model from the 8 and run it on paper. For two-way door decisions, do not overdo this. Use System 1 + the journal. For one-way door decisions, run at least two models in sequence.

Friday: Review the week’s journal. Grade each decision’s process, not its outcome. You probably do not know the outcomes yet. That is fine. You are grading whether you used a model, whether you stated a probability, whether you wrote down what would change your mind. 3 out of 10 is normal for week one. The goal is to get to 7 out of 10 by week four.

Week four: Run a quarterly review. Look at the 50 to 80 decisions you logged. Find the ones where outcomes have landed. Grade outcome independently from process. You will find a few high-process / bad-outcome (accept, it is variance) and a few low-process / good-outcome (scary, because you got lucky). Those are your most valuable data points. The first teaches patience. The second teaches humility.

Ongoing: Once a quarter, do a pre-mortem on your biggest active bet. 45 minutes. Assume it has failed one year out. Write the obituary. Top three risks become explicit mitigation plans. If you cannot mitigate them, change the bet.

That is the whole program. It is not complicated. What makes it work is doing it every week for 6 months. By month 6, you will notice that you are quietly making different choices than you would have made a year ago. Not bigger ones. Better calibrated ones. Smaller downside, similar upside, more honest about what you do not know. That is what decision quality compounding feels like from the inside.

The founders I watch who have been building for 10+ years and are still operating at peak performance do not have more raw intelligence than the ones who burned out at year 3. They have better decision hygiene. They journal. They run premortems. They ask “and then what” three times. They know when to trust gut and when not to. They accept that a bad outcome from a good process is not their fault, and a good outcome from a bad process is not their win. That posture, held for 10 years, is the compound interest.

Frequently Asked Questions

What is the best decision-making framework for founders?

There is no single best framework. Different decisions need different models. For reversible, low-stakes decisions, trust your gut (System 1) and log it in a journal. For irreversible, high-stakes decisions, run at least three models in sequence: Expected Value, Inversion, and Pre-Mortem. The meta-framework is reversibility first: Jeff Bezos’s two-way door vs one-way door test. That single filter tells you how much time to spend on the decision.

How do founders make decisions with incomplete information?

The goal is not to gather perfect information. It is to calibrate a probability estimate and identify what evidence would update it fastest. Ray Dalio’s rule is to treat every decision as a bet with a probability and reward. Jeff Bezos says 70% of the information you wish you had is usually enough for reversible decisions. The actual founder skill is knowing what threshold applies: 70% for two-way doors, 90%+ for one-way doors where the consequences are permanent.

What are the biggest decision-making mistakes founders make?

Three patterns kill more startups than any other. First, treating reversible decisions as irreversible (analysis paralysis on small bets). Second, treating irreversible decisions as reversible (signing a punitive term sheet in a hurry). Third, judging decisions by outcomes instead of by process (“resulting”), which means you learn the wrong lessons from lucky wins and unlucky losses. Research on 483 startup post-mortems shows 42% die from wrong-problem choices, which are almost always bad process on a one-way door decision.

How do you avoid analysis paralysis as a founder?

Use the reversibility test. If the decision is a two-way door (you can walk back), Jeff Bezos’s rule applies: decide with 70% of the information you wish you had. Most founder analysis paralysis is spent on reversible decisions that should take 30 minutes. The rest of the energy saved goes into the small number of one-way doors that actually deserve deep thought. If you cannot name the one-way door cost of the decision you are stuck on, you are probably overthinking a two-way door.

What is the difference between a good decision and a good outcome?

A good decision uses a sound process: you had good inputs, picked a fitting mental model, ran the math, wrote down your reasoning, and accepted the probability of being wrong. A good outcome is just what happened. They are independent. You can make a great decision and lose. You can make a terrible decision and win. The single highest-impact habit for founders is grading decisions by process, not outcome, through a decision journal reviewed every 90 days. Annie Duke calls the opposite error “resulting,” and it is the most common cognitive mistake founders make.

When should founders trust their gut vs run the numbers?

Trust your gut (System 1) when you have 50+ reps in the specific decision type. Trust the numbers (System 2) when the decision is new to you. Most first-time founders should not trust their gut on fundraising terms, senior hires, or pricing, because they have no pattern-match library. Second-time founders often over-trust gut in new domains where their old patterns do not apply. The test question: have I made this exact kind of call 50+ times before? If no, slow down, pull a framework, get a second opinion.

What is a pre-mortem and how do I run one?

A pre-mortem is a structured exercise where the team assumes a plan has already failed one year from now, then writes why it failed. Invented by Gary Klein. Run it as a 45-minute session: 10 minutes of silent writing (“imagine this project failed by April 2027, here are 5 reasons why”), then read-outs, then cluster the top 3 risks. Those three get explicit mitigation plans. If you cannot mitigate them, do not start the project. Pre-mortems work because they give teams permission to surface doubts they would not raise during a planning session, and they counter the planning fallacy (40 to 60% underestimates on timelines).

How often should I review my decisions as a founder?

Weekly for the journal entry itself (log each decision with confidence level). Every 90 days for grading the process. Every quarter, do a pre-mortem on your biggest active bet. Annually, look at outcomes and recalibrate your probability estimates. This is the routine of superforecasters studied in Philip Tetlock’s Good Judgment Project: not more data, just more honest recalibration. The founders who do this for 6+ months report noticeably different decision quality, though the effect is hard to see in any single week.

Further reading on this blog

• The pillar page for this cluster: The Founder Operating System. where decisions fit into the broader mental model for running yourself like you run a startup.
• Cluster sibling: The Solo Founder Mental Health Guide. decision fatigue is a real thing, and burnout crushes decision quality first.
• Cluster sibling: Validate a Startup Idea in 48 Hours. how to get good inputs for the biggest one-way door most founders ever walk through.
• Cross-pillar: The AI-Native Founder Playbook. how AI tools change the cycle time on OODA loops and what that means for decision cadence.
• Cross-pillar: The Solo Founder Scaling Playbook. hire-vs-automate is a classic reversibility matrix call.
• Cross-pillar: The AI Opportunity Map 2026. second-order thinking on where to place bets when the game board keeps changing.

If this helped, the single best thing you can do is start the decision journal today. Not Monday. Today. The compounding only starts when the logging starts.