Pilot Purgatory

Definition

Pilot purgatory is the state in which a technology project stays in pilot or proof-of-concept mode for an extended period and never reaches full deployment. The pilot isn’t cancelled and it isn’t scaled. It keeps running at small scale, often for a year or more, without delivering results the business can measure.

The term came out of manufacturing. Around 2018, consultancies including McKinsey used it to describe Industrial Internet of Things projects that stalled before scale-up. McKinsey’s work with the World Economic Forum on the Global Lighthouse Network gave it wide circulation, with a report stating that more than 70% of companies were stuck in pilot purgatory on advanced manufacturing technologies.

It’s now applied mostly to artificial intelligence. Since generative AI tools became widely available in late 2022, many organizations have launched pilots faster than they can move them into production, and “pilot purgatory” has become the common label for the resulting backlog.

See also: Proof of Concept (PoC), Minimum Viable Product (MVP), Stage-Gate Process, Crossing the Chasm

Why it matters for marketing

Marketing is one of the easiest places to start an AI pilot. A content team can test a writing assistant in an afternoon. A demand generation team can trial an AI lead-scoring tool on one segment. Starting is cheap, so pilots multiply.

Scaling them is a different job. Moving an AI content tool from five volunteers to a 60-person department means settling brand and legal review, integrating with the content management system, training everyone and deciding who pays for licenses. If nobody planned for those steps, the pilot stays where it is.

The cost shows up in a few ways. Budget and staff time go to experiments that don’t change results. Teams lose interest after the third pilot that went nowhere. And a marketing leader who has reported “we’re piloting AI” for two years has a harder time getting funding for the next request.

How pilots get stuck

Research on stalled pilots points to a consistent set of causes.

CauseWhat it looks like
No defined success criteriaThe pilot has no target metric, so nobody can say whether it worked
No path to productionIntegration, security review and support weren’t scoped at the start
No ownerThe pilot belongs to an innovation team, and no business unit has agreed to adopt it
Data that doesn’t scaleThe pilot ran on a small, hand-cleaned dataset that doesn’t reflect production data
Unclear business valueResults are described as “promising” but aren’t tied to revenue, cost or time
No scale-up budgetFunding covered the experiment and nothing after it
Too many pilotsResources are spread across a dozen experiments, and none gets enough to finish
Skills and change gapsThe wider team wasn’t trained or consulted and doesn’t adopt the tool

Gartner named four of these when it predicted in July 2024 that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025: poor data quality, inadequate risk controls, escalating costs and unclear business value.

McKinsey’s 2018 survey of about 400 manufacturers reported similar reasons for stalled IoT pilots. Respondents cited too few resources, too little knowledge and investments that were hard to justify at scale.

How to measure it

Pilot purgatory has no formal threshold. Organizations usually watch two indicators.

Pilot-to-production rate = (pilots moved to production ÷ pilots started) × 100

An illustrative example: a marketing organization started 12 AI pilots over 18 months and moved 3 into production. The rate is 3 ÷ 12 × 100 = 25%.

Pilot age is the time since a pilot started without a decision to scale or stop. In McKinsey’s 2018 survey, 84% of companies stuck in pilot mode had been there for more than a year, and 28% for more than two years.

Other signs are harder to quantify. A pilot that has been extended more than once, has no named executive sponsor, or has no scheduled decision date is at risk.

How to get out, and how to avoid it

Set exit criteria before starting. Write down the metric, the target and the date. Decide in advance what result leads to scaling and what result leads to stopping.

Name a business owner. Someone who will run the tool in daily operations should sponsor the pilot from the beginning.

Scope production requirements early. Involve IT, security and legal at the start so integration and review aren’t surprises later.

Use real data. Test on data that looks like what the tool will see in production.

Budget for scale. Estimate full deployment cost alongside pilot cost, and get conditional approval for both.

Limit the number of pilots. Fewer, better-resourced pilots are more likely to finish.

Hold a decision review. At the end date, choose one of three outcomes: scale, stop or extend once with a specific reason.

Stop pilots that don’t meet the bar. A stopped pilot isn’t in purgatory. It frees up people and budget.

Plan the rollout as a change effort. Training and communication for the wider team are part of scaling.

TermWhat it meansRelationship to pilot purgatory
Pilot purgatoryA pilot that continues indefinitely without scalingThe subject of this entry
Proof of Concept (PoC)A small test of whether an idea is technically feasibleThe stage where projects most often stall
PilotA limited real-world deployment with actual usersThe stage after a PoC, and the one the term is named for
Minimum Viable Product (MVP)The simplest version of a product released to real usersAn MVP is in production. A stuck pilot isn’t
Stage-Gate ProcessA project method with go or no-go decisions between phasesA structural way to prevent it
Crossing the ChasmThe gap between early adopters and the mainstream marketA similar gap, but in market adoption instead of inside one organization
Proof-of-concept fatigueLoss of enthusiasm after repeated pilots without resultsA common consequence

“PoC purgatory” and “POC graveyard” are variants with the same meaning. A project that’s formally cancelled has been abandoned, which is a different outcome. Purgatory is specifically the absence of a decision.

Best practices

Treat a pilot as a decision tool. Its purpose is to produce a scale-or-stop decision, and it should be designed to produce one.

Tie the pilot to a business metric. Use a measure the business already tracks, such as cost per lead or content cycle time. “Hours saved” is hard to convert into budget.

Pick use cases with a clear owner and clean data. These are more likely to reach production than more ambitious ones without either.

Time-box it. A fixed length of 60 to 90 days is common. The right number depends on the use case, but there should be one.

Design for the full department from day one. Ask how the tool would work for everyone, not only the five people in the test.

Report pilots as a portfolio. Track how many are active, their age and their status in one view, so stalled ones are visible.

Record what was learned from stopped pilots. The findings are useful for the next one.

Agentic AI is following the same pattern. In June 2025 Gartner predicted that more than 40% of agentic AI projects would be cancelled by the end of 2027, extending its earlier warning on generative AI.

Abandonment is rising. S&P Global’s 2025 survey of about 1,000 respondents in North America and Europe found that the share of companies abandoning most of their AI initiatives rose to 42%, from 17% a year earlier. The average organization scrapped 46% of its AI proofs of concept before production.

Scrutiny of return on investment. A 2025 report from MIT’s NANDA initiative drew wide attention for its finding that about 95% of generative AI pilots had produced no measurable profit-and-loss impact. The methodology has been questioned by some commentators, but the report sharpened executive attention on proving value.

Fewer, larger bets. Organizations are consolidating scattered experiments into a smaller number of funded programs with executive sponsors.

Vendor-built AI shortens the path. AI features delivered inside existing platforms, such as a CRM or marketing automation system, avoid much of the integration work that stalls custom pilots.

Maturity models focus on this gap. Research from MIT CISR identifies the move from pilots to scaled ways of working as the transition with the largest financial impact.

Frequently asked questions

What is pilot purgatory? It’s when a technology pilot keeps running at small scale for a long time without being rolled out or shut down.

Where did the term come from? From manufacturing. Consultancies including McKinsey used it around 2018 for stalled Industrial Internet of Things projects, and the World Economic Forum’s Global Lighthouse Network reports popularized it.

Why do AI pilots get stuck? Common reasons are unclear success criteria, no business owner, data quality problems, integration and security work that wasn’t planned, and no budget for scaling.

How long is too long for a pilot? There’s no set limit. A pilot that has passed its planned end date without a scale-or-stop decision is at risk. In McKinsey’s 2018 survey, most stuck pilots had run for more than a year.

What share of AI pilots reach production? Estimates vary by study and definition. S&P Global reported that the average organization scrapped 46% of its AI proofs of concept before production in 2025.

Is a failed pilot the same as pilot purgatory? No. A pilot that’s evaluated and stopped has reached a decision. Purgatory is the absence of one.

How is a pilot different from a proof of concept? A proof of concept tests whether something can work, usually in a controlled setting. A pilot tests it with real users on real work at limited scale.

How can a marketing team avoid it? Define the success metric and decision date up front, assign a business owner, involve IT and legal early, and budget for rollout as well as the test.

Does pilot purgatory only apply to AI? No. The term began with manufacturing technology and applies to any technology project. AI is where it’s most often used today.

What should happen to a pilot that’s already stuck? Hold a review with a deadline. Either commit an owner and budget to scale it, or stop it and record what was learned.

  1. Proof of Concept (PoC)
  2. Minimum Viable Product (MVP)
  3. Stage-Gate Process
  4. Crossing the Chasm
  5. Time to Value (TTV)
  6. Return on Investment (ROI)
  7. Key Performance Indicator (KPI)
  8. Machine Learning Operations (MLOps)
  9. Organizational Change Management (OCM)
  10. Build vs. Buy

Sources

Was this helpful?