Sunk Cost Fallacy

Why does this matter to founders?#

Startup investments are identity-heavy. A product, market thesis, key hire, or architecture choice is usually attached to a founder's public credibility. Updating then feels like conceding, and the reputational cost of appearing to have been wrong gets silently added to the cost of stopping. Some of that cost is real and belongs in the model. Most of it is a story about the past.

Runway converts delay into a hard cost. Continuing a weak initiative burns cash, but the larger loss is usually the milestone another team could have hit. When runway is finite, "one more quarter" is a company-level allocation decision, not a team-level one.

Fundraising narratives create commitment. Once a thesis is in the deck, the organisation starts selecting for confirming evidence. Investors are generally better served by a disciplined update than by a company defending an obsolete slide — and they price the credibility of your stop rules, not just your conviction.

Roadmaps accumulate invisible inertia. Completed code, migration work, promises made to named customers, and hard-won internal expertise make a feature feel valuable because it was difficult. Customers pay for outcomes, not for how hard the outcome was to produce.

How do you run a continuation decision?#

1. State the decision forward. "From today, should we invest the next $X and Y weeks here rather than in the best alternative?" Never "should we waste what we already spent?" — the second framing embeds the bias in the question.
2. Classify every cost using the table above. Separate sunk from avoidable, recoverable, and shutdown. This step alone resolves a surprising number of arguments, because the two sides are usually referring to different rows.
3. Rebuild the forecast from current evidence. Estimate future cash flows, probability of technical success, adoption, time-to-evidence, and required capital in base, upside, and downside cases. Past spending may inform these probabilities if it produced knowledge, assets, or commitments. It does not enter simply because it was large.
4. Price the next-best use of the same resources. Name the specific project that will not happen. See Opportunity Cost.
5. Precommit the stop rule. Before the next tranche, record the milestone, evidence threshold, budget, deadline, and decision owner — paid conversion, gross-margin target, reliability threshold, qualified pipeline. Moving the criteria after bad news is the operational signature of escalation.
6. Ask whether a stage gate beats both options. If a cheap experiment would resolve the main uncertainty before the large commitment, that third path usually dominates.
7. Add process protection. Assign a reviewer who did not sponsor the original decision, run a pre-mortem, and compare against a clean-sheet alternative: "if we were starting today with no history, would we choose this?"

Worked example: continuing a stalled build#

A team has spent $600,000 over nine months on a vertical integration. Finishing it requires another $250,000 and one squad-quarter. Current evidence puts the probability of a working, sellable result at 45%, worth $700,000 of first-year gross profit if it lands.

Stopping is not free, and it is not worthless:

• $30,000 recovered from a cancelable cloud commitment
• $80,000 of components reusable on other work
• $60,000 of shutdown cost — customer migration and wind-down
• the squad is freed for the next-best project, worth $90,000 of expected gross profit this year

Continue:

(0.45 × $700,000) − $250,000 = $315,000 − $250,000 = $65,000

Stop:

$30,000 + $80,000 − $60,000 + $90,000 = $140,000

Stopping wins by $75,000. Note what does not appear anywhere in either line: the $600,000. It is identical under both choices, so it cannot discriminate between them. If it is in your spreadsheet, it is in the wrong spreadsheet.

Now the counterpoint — when persistence is not a fallacy. Suppose two design partners signed contracts contingent on delivery, so stopping triggers $120,000 in refunds and penalties, and forfeits reference accounts the go-to-market plan depends on. That is a future cash consequence of a current choice, so it belongs in the model:

stop = $140,000 − $120,000 = $20,000

Continuing now wins by $45,000 — and the reasoning is entirely legitimate. This is the distinction the fallacy label routinely obliterates: reputation, contractual commitment, and relationship consequences are forward-looking costs, not sunk ones. "You're falling for sunk cost" is not a rebuttal to them.

And the option that usually beats both. Suppose $60,000 and four weeks would resolve the core technical uncertainty before committing the remaining $190,000:

−$60,000 + (0.45 × ($700,000 − $190,000)) = −$60,000 + $229,500 = $169,500

The staged path beats continuing ($65,000) and beats stopping in either variant. Caveat: this figure assumes the gate is perfectly informative and that four weeks of delay costs nothing — neither is true. A partially informative gate and a real delay cost will shrink the advantage, sometimes to zero. Treat the staged number as an upper bound and the ranking as the finding, not the amount.

What are the common mistakes?#

• Declaring the past irrelevant. Prior work creates reusable assets, obligations, learning, and evidence. Those change the forward forecast legitimately. Only the unrecoverable balance is off-limits.
• Ignoring exit costs. Severance, customer migration, warranty obligations, contractual penalties, and reputational damage are future consequences of stopping — the cost of shutting down is not sunk.
• Using "don't throw good money after bad" without a counterfactual. The question is always relative to the next-best use of the same resources, not relative to zero.
• Moving stop criteria after bad news. Redefining success repeatedly converts a staged experiment into open-ended escalation while preserving the appearance of discipline.
• Weaponising the term. "Sunk cost fallacy" has become a rhetorical device for shutting down colleagues who cite real commitments. Ask which row of the cost table the objection lives in before accepting the label.

When does this framework break?#

When the future is treated as known. Early-stage uncertainty is irreducible. Point estimates of a 45% probability convey confidence the evidence does not support; carry ranges and value the next experiment rather than the next forecast.

When option value is not modelled. A project with poor expected value can be worth continuing if it holds a cheap, high-information next step, or preserves access to a market that will be expensive to re-enter. That is real value, but it must be stated as a mechanism with a price, not invoked as a mood.

When capabilities are shared. Code, data, brand, certifications, and trained teams often survive the product they were built for. Model reuse explicitly, at a specific value, rather than treating a programme as wholly sunk or wholly salvageable.

When the metric is noisy. Persistence is genuinely necessary in startups, and false negatives are common at low volume. The antidote to sunk-cost bias is not constant pivoting — it is a precommitted learning cadence that can tell slow evidence apart from repeated failure of the same assumption.

When incentives, not reasoning, are the problem. If the sponsor's compensation, title, or standing depends on the project surviving, no framework will fix the decision. Change who decides.