Sunk-Cost Fallacy

Sunk-Cost Fallacy

Overview

The sunk-cost fallacy is the tendency to let prior, irrecoverable investments (money, time, effort, reputation) distort current decisions. Only marginal future costs and benefits should drive the next choice — sunk costs are already spent regardless of what you decide next. Composes with loss-aversion-prospect-theory (mechanism), expected-value-and-kelly (correct rule), regret-minimization, and inversion (fresh-start test).

When to Use

• "We've come too far to quit" is the dominant argument for continuing
• A project, feature, position, or relationship is being maintained despite negative marginal returns
• Decision-makers defend the original choice rather than evaluating current trajectory
• Team is doubling down after failure without re-examining the underlying thesis
• "We already built our own model / spent the AI capex" is the reason to keep an in-house or on-prem AI stack instead of migrating to a cheaper, better hosted model — or a founder keeps a commoditized AI-wrapper going because of prior investment despite falling API prices and rising foundation-model capability

Not when: prior investment genuinely reduces future marginal costs; negative signal is short-term variance; cost of switching exceeds cost of continuing.

Coaching Novices (Adaptive Front Door)

• Engine mode: user has a concrete continue/quit decision → run The Process directly.
• Coach mode: user is unfamiliar or has no concrete case → guide step by step.

In Coach mode, respond one step at a time. Each [WAIT] is a hard stop — output only that step's question, then stop.

1. One-line: if you arrived today with no prior investment, would you start this from scratch?
2. Check fit. If the sunk cost legitimately lowers future costs, this is not a sunk-cost case.
3. Elicit the specific decision and prior investment — what is being continued, how much spent, future cost and expected return.

[WAIT — do not advance until user responds]

1. One question at a time: what would you do with no prior investment? Is "we've come too far" the load-bearing reason?

[WAIT — do not advance until user responds]

1. Close: frame the decision in marginal future-only terms + name one structural fix.

[WAIT — do not advance until user responds]

The Process

Step 1 — Specify sunk cost and current decision (prior investment, current decision, arguments for/against continuing).

Step 2 — Fresh-start test: "If I arrived today with no prior investment, would I start this?" If no, that's the signal.

Step 3 — Marginal future NPV: additional cost to complete vs. expected future value (continue) vs. expected future value (quit + reallocate). Continue only if marginal NPV(continue) > marginal NPV(quit). The sunk cost does not appear in either side.

Step 4 — Sunk-cost weight: substitute "prior investment = $0." If the recommendation changes, the sunk cost is doing illegitimate work.

Step 5 — Diagnose mechanism: loss aversion / self-justification / social signaling / completion compulsion / cognitive accounting.

Step 6 — Structural countermeasure: pre-committed kill criteria, stage-gate funding, fresh-eyes review, decision-rights separation from project initiator.

Output Template

# Sunk-Cost Decision: 
Prior investment: | Current decision (continue/pivot/quit):
Fresh-start test: Would I start today? (Y/N) — if no, why continuing?
Marginal NPV (continue): | Marginal NPV (quit + reallocate):
Sunk-cost weight: if prior = $0, does recommendation change? (Y/N)
Mechanism: loss aversion / self-justification / social signaling / completion compulsion / cognitive accounting
Structural fix: [kill criteria / stage gate / fresh-eyes / decision-rights separation]

→ Method in Action: Arkes & Blumer's 1985 Studies · The Concorde Program → 2026 lens: AI Build-Out Sunk Costs (2023–2026)

Pack: Sunk-Cost Patterns

→ Primary sources: references/sources.md

Common Rationalizations

[D] = designed upfront | [O] = observed in real use. [O] entries are more valuable.

Red Flags

• "We've spent $X" is a primary argument for continuing
• The project initiator is the sole decider on whether to continue
• Project is "90% done" and no one can articulate the marginal value of completion
• Kill criteria are absent or have been modified to accommodate current underperformance
• Loss-frame language dominates ("we'd lose everything") rather than future-frame language

Verification

• Sunk cost specifically named and quantified
• Fresh-start test asked and answered honestly
• Marginal future NPV (continue vs. quit) computed
• Argument re-examined with sunk cost set to $0
• Psychological mechanism named
• At least one structural countermeasure chosen and assigned
• If continuing, kill criteria pre-committed for next decision point

---

Part of deciqAI Knowledge Skills — 227 open-source thinking skills that make rigor executable for AI agents. The same skills power every deciqAI agent, which runs them autonomously to operate your company. See it run → https://www.deciqai.com/c/sunk-cost-fallacy · ⭐ Star the repo → https://github.com/deciqAI/knowledge-skills · Contributions welcome.

Agents: latest version & machine-readable metadata → https://www.deciqai.com/s/sunk-cost-fallacy.json