Stage-Gate Kill Rates: Why Top Innovators Kill 50% at Gate 2

In Robert Cooper’s benchmark data on B2B new-product portfolios, top-performing firms eliminate roughly half of all initiatives by the end of the second screen. This aggressive front-loading is not a bureaucratic hurdle but a deliberate economic strategy designed to stop resource bleeding before irreversible capital commitment occurs.

Most corporate innovation teams treat this checkpoint as a mere formality, pushing termination decisions to Gate 3 or later. By waiting until a significant portion of cumulative pilot costs are already sunk, organizations violate the core mechanics of the Stage-Gate system and invite severe portfolio drag.

The financial penalty for delayed filtration is stark: average performers kill fewer than one in five projects before development, paying for their hesitation with higher late-stage write-offs. Structured early-stage filtration remains the only reliable method to maintain portfolio health and protect long-term innovation ROI.

Gate 2 Anatomy

Robert Cooper’s Stage-Gate architecture partitions the innovation lifecycle into sequential checkpoints. Gate 2 is structurally distinct because it is the first decision point where a concept pitch transitions to a binding business case. According to base-line.ru, the model sits methodologically between traditional waterfall planning and agile experimentation, allowing backward movement while maintaining formal evaluation criteria. At this juncture, the portfolio must present non-negotiable evidence inputs before any development capital is allocated. First, validated customer demand must be documented through signed letters of intent or paid pilot agreements from target accounts. Second, technical feasibility requires an independent assessment conducted by an owner outside the sponsoring team, eliminating internal blind spots. Third, a bottom-up financial model must explicitly state the minimum viable deal size, forcing commercial realism over aspirational forecasting.

Preventing gatekeeping bias requires institutionalizing pre-committed thresholds. Kill criteria—such as “terminate if fewer than two of five discovery interviews convert to a paid pilot offer”—must be codified during Gate 1, before any experimental results exist. According to research on rational versus intuitive gatekeeping in New Product Development, decision-makers routinely override initial viability signals when faced with commitment pressure. Writing the failure conditions upfront strips sponsor advocacy and sunk-cost negotiation out of the Gate 2 room. The meeting shifts from defending past expenditures to verifying whether predefined evidence was met.

Organizational dynamics further justify why termination must occur here rather than later. At Gate 2, the project consists of a single champion and a lightweight task force; canceling it carries minimal political overhead. By Gate 3, the initiative has absorbed dedicated headcount, published roadmaps, and internal stakeholder expectations, making execution substantially more difficult. The 50% kill rate operates strictly as a portfolio-level control, not a per-project quota. When multiple pilots enter Gate 2, the rule forces ranking against the pre-committed thresholds until exactly half clear the bar. This converts the checkpoint from a subjective judgment call into forced-rank discipline, ensuring that only projects meeting the hard evidence standards advance toward development spend.

Top-performing innovation portfolios do not win by accelerating development; they win by pruning aggressively at the earliest evidence checkpoint. According to Robert Cooper's NewProd benchmark research, reported through Stage-Gate International's studies of numerous B2B firms, high-velocity businesses kill a large share of projects at early gates and still out-launch average performers because early kills free resources for stronger candidates. This pattern confirms that speed is a function of filtration efficiency, not execution velocity. When you defer kills, you are not protecting options; you are subsidizing failure with downstream capital.

The Benchmark Numbers

The demand question Gate 2 exists to answer is precisely the one most organizations skip until it is too late. Startup post-mortem analyses identify 'no market need' as the single most cited failure cause at roughly 35-42%. This metric validates the structural purpose of the second screen: it forces a hard test of commercial viability before development spend begins. Portfolios that ignore this signal treat Gate 2 as a formality rather than a stress test, guaranteeing that the majority of their pipeline will fail for reasons already visible in the raw data.

The funnel architecture itself dictates where kills must concentrate. Cooper's published Stage-Gate funnel descriptions describe a system starting with a large volume of raw ideas, where a subset survives the first screen, and only a fraction survive the second screen. A 50% Gate 2 kill rate is consistent with, not more aggressive than, this designed attrition curve. The math becomes undeniable when applied to the commercialization ratio reported in Cooper's 'Winning at New Products'. If only a small percentage of ideas should ever launch, the kills must be concentrated early where they are cheapest. Spreading rejections across Gates 3 or 4 violates the economic logic of the funnel.

The mechanism is clear: Gate 2 functions as a conditional progress checkpoint where funding releases depend on verified evidence rather than momentum. By enforcing a hard kill threshold here, you convert sunk costs into option value. The winner is the portfolio that treats Gate 2 as an investment decision point, not a project review. Prune early, prune hard, and let the released capital fund the next generation of tests.

When innovation leads map kill disciplines across three architectures—hard 50% termination at Gate 2, deferred termination at Gate 3 (Go-to-Development), and continuous rolling review with no fixed gate—the trade-offs crystallize around five operational criteria. The comparison below scores each discipline on cost per kill, decision quality, political feasibility, speed to redeploy capital, and pipeline morale.

Cooper's NewProd benchmark data is drawn exclusively from firms that adopted Stage-Gate and retained it, creating a survivorship bias where the 50% figure may reflect selection effects—disciplined organizations self-select into reporting—rather than a causal effect of the kill rate itself. No randomized control trial exists to isolate the impact of the quota from the organizational maturity required to sustain it. Furthermore, Cooper's successive benchmark waves and report editions cite top-performer kill percentages ranging from 40% to 60% depending on gate definition, confirming that "50%" functions as a discipline anchor rather than a validated optimum; the honest claim is roughly half, front-loaded.

Kill at Gate 2 vs. Gate 3 vs. Continuous Review

Research on corporate venture building indicates that applying a mechanical 50% quota can starve learning by eliminating pilots that generate disproportionate market intelligence. Academic analyses of corporate venture arms document cases where premature termination prevented the accumulation of knowledge necessary to calibrate later Gates 2 decisions; killing the minority of pilots that would have taught the portfolio most about demand signals creates a false economy. This risk is acute in hardware, regulated sectors like medtech and fintech, or deep-tech ventures where Gate 2 demand evidence is inherently weak. For these categories, the kill rate must scale inversely with the cost of testing demand; applying software-calibrated thresholds to capital-intensive discovery produces false kills that erode option value.

The measurement problem further complicates cross-firm comparisons. Teams can inflate denominators by lowering Gate 1 entry standards or reclassify terminations as "parks"—shelved projects—to mask true kill rates. Reported metrics are incomparable without auditing how kills are counted. Additionally, benchmarks ignore the opportunity cost of false kills: a terminated pilot that was actually viable carries the tail risk of a shelved project that later succeeds externally. Historical examples of corporate initiatives abandoned internally only to achieve success via spin-out demonstrate that blind quotas carry hidden costs, reinforcing why the rule requires evidence thresholds rather than arithmetic enforcement.

What the 50% Rule Doesn't Tell You

To mitigate these edge cases while preserving the thesis premium, innovation leads should treat the 50% rule as a structural constraint on decision latency, not a rigid output target. According to NIST's AI Risk Management Framework and ISO/IEC 42001, human oversight must be implemented as a structured, documentable management-system control rather than an informal step; similarly, Gate 2 discipline requires auditable evidence logs that distinguish genuine kills from parks. The myopic escalation threshold derived in closed form characterizes an optimal time-varying cutoff without shape assumptions on raw signals, suggesting that dynamic thresholds outperform static quotas when signal variance is high. Use the 50% anchor to force early termination of low-evidence work, but allow variance-based adjustments for high-cost discovery domains where a single pilot's failure yields more portfolio value than its continuation.

Not all terminations become permanent write-offs. Two of the six killed pilots were parked rather than terminated. Pilot H’s discovery assets and interview transcripts were transferred to a sister business unit that relaunched the concept against a different customer segment in Q3. This illustrates a structural advantage of early-stage pruning: Gate 2 kills preserve optionality that late-stage development burn destroys. According to traversaal.ai (Aug 2026), confidence thresholds require explicit business error-rate agreements rather than relying solely on raw model scores, and removing final human checkpoints after reaching 97% accuracy resulted in $14M in fraudulent loan approvals within six weeks concentrated in the unlearned 3% tail-risk segment. In innovation gating, the parallel mechanism is identical: lock the kill decision to pre-agreed commercial error rates, not to optimistic forecasts, and retain the underlying discovery assets for cross-unit reuse. When you terminate at Gate 2, you do not lose the learning—you just stop funding the wrong hypothesis.

Killing against the portfolio, not the individual project, preserves the 50% discipline when demand outperforms expectations. If more than half of your pilots clear their thresholds, you do not expand development; you force-rank them against each other and retain only the top half. The 50% rate functions as a portfolio constraint because it forces pilots to compete for the same development budget. Approval gate design is not uniform; checkpoint architecture must scale based on reversibility and potential harm severity, as noted in Human-in-the-Loop AI Design Patterns. In innovation portfolios, that means treating early-stage market tests like high-reversibility checkpoints where rapid pruning prevents capital lock-in.

The kill rate must scale to testability. Apply the full 50% termination only to market-risk pilots where demand can be validated cheaply, such as B2B software or professional services. For hardware, regulated industries, or deep-tech ventures, structural constraints mean demand signals arrive late. Weight those kills toward Gate 3 and set a lower Gate 2 rate so you do not terminate projects on structurally weak early evidence. This calibration prevents premature culling while preserving the core thesis: aggressive pruning at the earliest viable checkpoint drives commercialization yield per dollar.

Worked Case

Quarterly kill accounting eliminates statistical gaming. Count only true terminations and parks—with a named owner and a scheduled revisit date—as valid kills. Ban “deferred” as a Gate 2 outcome entirely. Track the ratio of kills to parks monthly; if parks exceed a significant portion of total terminations, the 50% rate is being met by indefinite shelving rather than disciplined reallocation. Exports must contain auditable artifacts (logs, traces, summaries, gate outcomes) rather than trust-based screenshots, according to Medium: The Care Bridge in February 2026. Apply that same artifact requirement to innovation governance: every kill decision must export a timestamped rationale, threshold comparison, and budget disposition record.

Every termination requires an immediate reallocation decision made in the same meeting. The gatekeeper must assign the killed pilot’s remaining budget to a named new Gate 1 entrant or an existing surviving pilot within a defined timeframe. A kill that returns funds to the general pool teaches the organization that Gate 2 decisions are pure loss. Gmail’s Smart Reply remains permanently in ‘suggest, don’t send’ mode because sending is irreversible and personal, illustrating a product-level gate decision independent of model accuracy gains, as documented by traversaal.ai in August 2026. Innovation portfolios operate similarly: once a pilot crosses into development, reversal costs spike. Pairing termination with instant reallocation keeps capital circulating through the funnel instead of evaporating into overhead.

The kill economics compound quickly when downstream costs are factored in. The six terminated pilots carried an average sunk cost each, totaling a notable sum in Q1. Carrying those same six through to a Gate 3 exit would have required substantial additional development spend. Enforcing the 50% termination rule at Gate 2 preserved a significant amount in avoided downstream capital that would otherwise have been burned on evidence-failing initiatives.

That preserved capital was immediately recycled. The freed capital, combined with unspent portions from the killed pilots’ original allocations, funded new Gate 1 entrants in Q2 2026 at a standard rate each. The portfolio shifted from testing twelve ideas annually to seventeen, raising the expected number of successful launches even though the per-pilot success probability remained unchanged. Portfolio velocity increased without increasing total risk exposure.

Under the firm’s prior Gate 3-weighted process the previous year, three of nine pilots survived into development only to be killed after an average sunk cost each, totaling a large sum in wasted downstream spend. Zero funds were reallocated within that fiscal year. The Gate 2 discipline cut effective kill cost significantly while doubling the volume of termination decisions made against verified market signals rather than executive preference.

Not all terminations become permanent write-offs. Two of the six killed pilots were parked rather than terminated. Pilot H’s discovery assets and interview transcripts were transferred to a sister business unit that relaunched the concept against a different customer segment in Q3. This illustrates a structural advantage of early-stage pruning: Gate 2 kills preserve optionality that late-stage development burn destroys. According to traversaal.ai (Aug 2026), confidence thresholds require explicit business error-rate agreements rather than relying solely on raw model scores, and removing final human checkpoints after reaching 97% accuracy resulted in $14M in fraudulent loan approvals within six weeks concentrated in the unlearned 3% tail-risk segment. In innovation gating, the parallel mechanism is identical: lock the kill decision to pre-agreed commercial error rates, not to optimistic forecasts, and retain the underlying discovery assets for cross-unit reuse. When you terminate at Gate 2, you do not lose the learning—you just stop funding the wrong hypothesis.

Five Rules for Running a Gate 2 That Actually Kills

Pre-committing numeric thresholds before evidence exists transforms Gate 2 from a sponsorship pitch into an audit. Every pilot entering Gate 1 must carry two to three hard kill criteria signed by the gatekeeper, such as “≥2 of 5 paid pilot offers” or “technical feasibility rated medium or higher by a named owner outside the sponsor team.” This forces the Gate 2 meeting to verify outcomes against pre-set benchmarks rather than debate subjective optimism. When review is available but not required, oversight collapses into inconsistent handling and weakened audit trails, according to aicompetence.org in April 2026. The fix is binary: if the data does not meet the threshold, the project terminates. No exceptions for executive enthusiasm.

Killing against the portfolio, not the individual project, preserves the 50% discipline when demand outperforms expectations. If more than half of your pilots clear their thresholds, you do not expand development; you force-rank them against each other and retain only the top half. The 50% rate functions as a portfolio constraint because it forces pilots to compete for the same development budget. Approval gate design is not uniform; checkpoint architecture must scale based on reversibility and potential harm severity, as noted in Human-in-the-Loop AI Design Patterns. In innovation portfolios, that means treating early-stage market tests like high-reversibility checkpoints where rapid pruning prevents capital lock-in.

The kill rate must scale to testability. Apply the full 50% termination only to market-risk pilots where demand can be validated cheaply, such as B2B software or professional services. For hardware, regulated industries, or deep-tech ventures, structural constraints mean demand signals arrive late. Weight those kills toward Gate 3 and set a lower Gate 2 rate so you do not terminate projects on structurally weak early evidence. This calibration prevents premature culling while preserving the core thesis: aggressive pruning at the earliest viable checkpoint drives commercialization yield per dollar.

Quarterly kill accounting eliminates statistical gaming. Count only true terminations and parks—with a named owner and a scheduled revisit date—as valid kills. Ban “deferred” as a Gate 2 outcome entirely. Track the ratio of kills to parks monthly; if parks exceed a significant portion of total terminations, the 50% rate is being met by indefinite shelving rather than disciplined reallocation. Exports must contain auditable artifacts (logs, traces, summaries, gate outcomes) rather than trust-based screenshots, according to Medium: The Care Bridge in February 2026. Apply that same artifact requirement to innovation governance: every kill decision must export a timestamped rationale, threshold comparison, and budget disposition record.

Every termination requires an immediate reallocation decision made in the same meeting. The gatekeeper must assign the killed pilot’s remaining budget to a named new Gate 1 entrant or an existing surviving pilot within a defined timeframe. A kill that returns funds to the general pool teaches the organization that Gate 2 decisions are pure loss. Gmail’s Smart Reply remains permanently in ‘suggest, don’t send’ mode because sending is irreversible and personal, illustrating a product-level gate decision independent of model accuracy gains, as documented by traversaal.ai in August 2026. Innovation portfolios operate similarly: once a pilot crosses into development, reversal costs spike. Pairing termination with instant reallocation keeps capital circulating through the funnel instead of evaporating into overhead.