We read forty practitioner accounts of corporate–startup pilots, pulled every conversion statistic we could find back to its primary source, and catalogued twenty-two cases where a pilot actually converted. The pattern was not what the field says it is. The pilot is rarely the problem. The handover is.
of corporates see fewer than 1 in 4 pilots become a commercial deal. Nearly half see fewer than 1 in 10.
500 Startups, 2017 · n=100+ Fortune 1000 executivesof corporate–startup engagements reach a signed deal, measured across 1,500+ engagements at 120+ corporates.
European Innovation Council, 2025spread between the lowest and highest published pilot conversion rates — from 6% to 90%. That spread is the finding.
This report’s synthesis of published ratespractitioner accounts, out of forty, that credit anything done during a pilot with saving it. Every save was structural, and every one was set up before day one.
Original finding · Arc Shift corpus, 2026The published conversion rates look like six contradictory measurements of the same thing. They aren’t. They are one clean measurement of a single variable: how much production budget authority the person who signed the pilot actually had. Sort them by that, and the contradiction disappears into a ladder.
Everything usually blamed — clock-speed mismatch, risk aversion, not-invented-here, “we speak different languages” — is downstream of one structural fact: the pilot was funded from an innovation cost centre, and has to convert into a business-unit operating plan that was locked six to twelve months before the pilot began.
The part everyone manages, staffs and worries about. It is comparatively survivable — the technology usually works, and both sides are motivated.
The moment the pilot has to become a budget line owned by someone who never signed up for it. Almost nobody staffs this stage. It is where the deal dies.
Corroborating anchors at the scale decision: BMW converts roughly 20% of its engagements; BCG measured 22%; McKinsey found under 30%. Honest caveat: the 60–80% band is a synthesis across sources, not a single survey line. The direction is well supported; the precise number is not.
Across twenty-two catalogued saves, not one was rescued by effort, enthusiasm or escalation after the pilot had started. They were won by structures agreed before day one. These are the ones that came up repeatedly, from independent sources.
The honest caveat: this corpus has real survivorship bias — nobody publishes a blog post titled “we rescued it in week eight.” The finding is that no practitioner in this corpus credits mid-pilot heroics. It is not proof that mid-pilot intervention never works.
We tried to trace every statistic to its primary source. Several of the most repeated ones do not have one. If you have been handed these in a deck, they came from somewhere other than where the deck says.
“Structured pilots with measurable KPIs are 3× more likely to scale”
Attributed to BCG
Misattributed Not in the BCG report — which never uses “3×” and never says “proof of concept.” The real finding is from Sapphire Ventures: POCs lasting under three months convert 3× better. The variable was quietly swapped from duration to “structure.”
“More than half of proofs of concept fail”
Attributed to BCG
Misattributed Also absent from BCG. It originates as a Sapphire Ventures blog headline, restating a different measurement entirely — and with an undisclosed sample.
“70% of transformations fail”
Attributed to McKinsey
Zombie Traces to Hammer & Champy in 1993, who described it themselves as an unscientific estimate about reengineering. A 2011 review of five published instances found no valid empirical basis for any of them. McKinsey never measured it.
“43% of enterprise deals stall in security review”
Attributed to two security vendors
Unverifiable Traces to a vendor blog citing two sources with no link, page, method or sample — and then uses the figure to compute the ROI case for its own $149/month product.
Worth knowing how this happens: the page laundering the BCG misattributions carries source links tagged utm_source=chatgpt.com. Unchecked machine-generated citations, quoted onward by people who reasonably assumed somebody had verified them.
Whether you’re a corporate with a pilot that needs a path to a P&L, or a startup trying to get one over the line — this is the problem I work on. I’ve led three corporate spin-outs through exactly this transition, and I sit on both sides of the table across eleven APAC markets.
If you’re building partnerships between corporates and startups, reach out. Happy to compare notes even if there’s nothing to sell.
Take the 15-question pilot diagnostic →
Related: Corporate venture building → · For corporate venture teams →
Forty practitioner accounts of corporate–startup pilots, read and coded. Every conversion statistic we could find traced back toward its primary source, and graded: A for a study with a disclosed sample, B for vendor-published or undisclosed-sample figures, EST for a reasoned estimate. Twenty-two cases where a pilot converted, catalogued by mechanism.
Two limits worth stating plainly. First, survivorship bias is real — nobody publishes a post called “we rescued it in week eight,” so the absence of mid-pilot saves in this corpus is not proof that they never happen. Second, the structural-over-cultural argument is the central falsifiable claim here: it is an argument from the shape of the evidence, not a controlled demonstration. Where a number is soft, it is labelled soft on the page rather than rounded into confidence.