Why AI pilots stall - the operating change that fixes it

We ran this content engine as a pilot before we ran it as a production loop. It stalled. The moment it stalled taught us more in an afternoon than six months of better prompting had, because the problem was never the model. We had pointed good AI at a workflow we had never redrawn, and a faster version of a broken handoff is still a broken handoff.

The stall is an operating-model problem: organizations layer AI onto the workflow they already have instead of redesigning the workflow for AI, and that sequencing error predicts whether a pilot scales or dies.

Short answer: Most AI pilots stall because the org wraps AI around the old workflow rather than redesigning the workflow for AI. McKinsey found that AI high performers are nearly three times as likely to fundamentally redesign workflows. The fix is structural: change how work is organized before you change the tools running it.

Why do AI pilots stall before they scale?

The pilot stalls because it was bolted onto a process nobody redrew, and the process was the constraint all along. The model performed. The organization around it did not change shape, so the gains had nowhere to go.

The failure rate is not a rumor. Gartner predicted in July 2024 that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025. By the time the data came in, Gartner reported more than half had been abandoned, on poor data quality, thin risk controls, escalating costs, and unclear business value. The proof of concept worked. The path to production did not exist.

The graduation math is worse. The IDC and Lenovo CIO Playbook 2025, which surveyed 3,120 IT and business decision-makers, found that 88% of AI proofs of concept never reach production. For every 33 pilots a company launches, four make it to production. The other 29 die somewhere between the demo and the org chart.

Why does the demo work but the rollout fail?

A demo tests the model. A rollout tests the organization, and the organization is the part that was never rebuilt. That is the whole gap, and it shows up the moment the pilot has to survive a real handoff instead of a curated one.

Stanford's Digital Economy Lab studied 51 successful enterprise AI deployments across 41 organizations and found that 77% of the toughest implementation challenges were invisible and intangible: change management, data quality, and process redesign. The AI technology itself rarely topped the list. The hard part sat in the seams between the tool and the people, which is exactly where a demo never goes.

Failure is also so normal it is almost a prerequisite. The same Stanford research found that 61% of the successful projects had a failed AI attempt before they succeeded. The first pass dropped AI into the existing process and stalled. The second pass redrew the process. That is the move that separates a screenshot from a system.

What percentage of AI pilots actually reach production?

Roughly one in eight, and the share that moves the P&L is smaller still. The number depends on where you draw the line, and every credible source lands in the same low range.

McKinsey's State of AI 2025, a survey of 1,993 respondents across 105 countries, found that 88% of organizations use AI in at least one function, yet only about 6% qualify as AI high performers, the group where AI drives 5% or more of EBIT. Two-thirds of AI users sit in pilot mode with no enterprise-level EBIT impact. Adoption is near-universal. Impact is rare.

One often-quoted figure puts the success rate even lower. MIT's NANDA initiative reviewed more than 300 AI initiatives and interviewed 52 organizations, and reported that only about 5% of AI pilots achieved rapid revenue acceleration, with the rest delivering little measurable P&L impact. Read that one as directional. It is preliminary research rather than a peer-reviewed study, and it points the same way as the rest.

What separates the pilots that scale from the ones that die?

The pilots that scale redesign the workflow before they scale the tool. The ones that die insert AI at an existing handoff point and hope the rest of the process keeps up. This is the single most consistent split in the data.

McKinsey put a multiple on it. AI high performers were 2.8 times more likely to have fundamentally redesigned their workflows: 55% of them against 20% of everyone else. The figure is self-reported survey data, so treat the exact number as directional, but the direction is unambiguous and it matches every other source here. The companies that moved the P&L are the companies that redrew how the work runs.

You can see the same pattern in usage depth. OpenAI's State of Enterprise AI 2025 found that frontier firms in the 95th percentile generate about twice the messages per seat and roughly seven times the model interactions of a median enterprise. They do not have better access to the tools. They built the work around them, so the tools sit inside the daily flow instead of beside it.

There is a name forming for the missing layer. IBM's Institute for Business Value reports that enterprises which build an AI orchestration layer are far more likely to scale successfully, and frames the stall as the absence of an AI operating model: a structural governance and workflow layer that sits above the individual tools. MIT CISR, drawing on 39 interviews and a 152-organization survey, reaches a related conclusion. Newer AI is blurring the line between IT and the business, and that demands a different operating-model structure to manage it. Different sources, one finding: the unit of change is the operating model.

*If this is the part you came for, we write about how operators rebuild these workflows. *Get the next piece.

What is the one operating change that fixes the stall?

Map the workflow before you scale the tool. Decide which steps AI can own end to end, then rebuild the handoffs around those steps, instead of dropping AI into the handoff points the old process already had.

Most rollouts do the reverse. They find the spot where the work currently changes hands and slot an agent into it, which leaves every other step, approval, and exception exactly where it was. The agent now runs faster inside a structure built for the speed of people. Nothing downstream can keep up, so the gain stalls at the first human gate it hits.

Redesigning for AI means asking what the workflow would look like if a capable agent owned its routine stretches outright. Which steps need no human in the loop. Where the handoff actually belongs once the routine is automated. Where a person still has to make the judgment call, set the policy, or own the exception. You answer those first, then you scale. The pilots that skip this step are the 29 in 33 that never graduate.

The two-question test before you scale a pilot

Before a pilot earns a scale decision, it has to pass two questions. If the answer to either is no, the pilot is not ready, and scaling it will only buy you a faster version of the problem.

The question	What a "yes" requires	If the answer is "no"
1. Have you mapped what the workflow actually does, step by step?	A written map of every step, handoff, approval, and exception in the real process, not the idealized one.	You are scaling a black box. Stop and map it. The seams are where pilots die.
2. Have you decided which steps AI owns end to end versus hands off to a person?	A clear line, step by step, between what an agent owns outright and where a human keeps the judgment, the policy, or the exception call.	You are inserting AI at an old handoff point. Redraw the handoffs around what AI can own, then revisit the scale decision.

The test is deliberately boring. The companies that scale AI are not the ones with the best models. They are the ones that did the unglamorous mapping work before they wrote the check, which is why Stanford found process redesign and change management harder than the technology in nearly four out of five cases.

How is AI pilot failure different from regular software failure?

A normal software project fails on the build. An AI pilot usually succeeds on the build and fails on the fit, because the technology works in the demo and then meets a process nobody redesigned to receive it. The code is rarely the problem.

Traditional software ships against a spec. You know what it should do, you build it, it works or it does not. AI lands differently. The model performs in the pilot, then stalls when it has to live inside real handoffs, real exceptions, and real data quality. Stanford's finding that 77% of the hardest challenges were invisible and intangible is the tell. The failure is organizational, so the fix has to be too.

FAQ

Why do AI pilots succeed in demos but fail in production? A demo tests the model on a curated path. Production tests the whole workflow: the handoffs, approvals, exceptions, and data quality the demo skipped. Stanford found 77% of the toughest deployment challenges were invisible and intangible, sitting in process and change management rather than the technology. The model works. The surrounding process was never redrawn to receive it.

How do you move an AI pilot to production at enterprise scale? Map the workflow step by step, decide which steps AI owns end to end, then rebuild the handoffs around those steps before you scale. McKinsey's high performers were 2.8 times more likely to have fundamentally redesigned workflows. Redesign comes before the rollout. Inserting AI at the old handoff points is the move that stalls.

What percentage of AI pilots actually succeed? Roughly one in eight reach production. IDC and Lenovo found 88% of AI proofs of concept never get there, and McKinsey found only about 6% of organizations qualify as AI high performers where AI drives 5% or more of EBIT. Adoption is near-universal. P&L impact is rare.

What is an AI operating model and why does it matter for scaling? An AI operating model is the structural layer that governs how work, decisions, and accountability are organized around AI. IBM frames the stall between experimentation and scale as the absence of this layer. MIT CISR reaches the same conclusion: scaling AI needs a different operating-model structure, not just adoption.

What does redesigning workflows for AI actually mean in practice? It means redrawing the work so AI owns its routine steps end to end, then rebuilding the handoffs around that. You map every step, decide what needs no human in the loop, and place the remaining judgment calls deliberately. OpenAI's frontier firms drive roughly seven times the model interactions of median enterprises because AI sits inside the daily flow, not beside it.

How is AI pilot failure different from regular software project failure? Software projects usually fail on the build, against a known spec. AI pilots usually pass the build and fail on the fit: the model performs in the demo, then stalls inside real handoffs, exceptions, and data quality. Stanford found 77% of the hardest challenges were organizational rather than technical, which is why the fix is structural.

Before you scale your next pilot

Run the two-question test on it first. We built a free Claude skill that makes AI-drafted writing read like a person wrote it, the same one behind every article we publish, and you can have it. Download the human-writing skill. It is email-gated because the download is the opt-in.

If you would rather not download anything, join the newsletter and we will send the next piece on rebuilding workflows for AI when it is out.

Sources

1. Gartner, "Gartner Predicts 30% of Generative AI Projects Will Be Abandoned After Proof of Concept by End of 2025" (July 2024). https://www.gartner.com/en/newsroom/press-releases/2024-07-29-gartner-predicts-30-percent-of-generative-ai-projects-will-be-abandoned-after-proof-of-concept-by-end-of-2025
2. Gartner, "GenAI Project Failure" (updated analysis). https://www.gartner.com/en/articles/genai-project-failure
3. IDC and Lenovo, "CIO Playbook 2025," via CIO.com (3,120 decision-makers, Sept-Oct 2025). https://www.cio.com/article/3850763/88-of-ai-pilots-fail-to-reach-production-but-thats-not-all-on-it.html
4. McKinsey, "The State of AI in 2025" (November 2025, n=1,993, 105 countries). Workflow-redesign figures are self-reported survey data; treat the exact percentages as directional. https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai
5. MIT NANDA, "The GenAI Divide" (August 2025), via Fortune. Preliminary, non-peer-reviewed research; the ~5% figure is directional. https://fortune.com/2025/08/18/mit-report-95-percent-generative-ai-pilots-at-companies-failing-cfo/
6. Stanford Digital Economy Lab, "The Enterprise AI Playbook" (March 2026, 51 deployments across 41 organizations). https://digitaleconomy.stanford.edu/publication/enterprise-ai-playbook/
7. IBM Institute for Business Value, "AI Orchestration Layer" and "The AI Operating Model" blueprint (2026). https://www.ibm.com/thought-leadership/institute-business-value/en-us/report/ai-orchestration-layer and https://newsroom.ibm.com/2026-05-05-think-2026-ibm-delivers-the-blueprint-for-the-ai-operating-model-as-the-ai-divide-widens
8. MIT CISR, "Enterprise IT Operating Models" (December 2025, 39 interviews, 152-org survey). https://cisr.mit.edu/publication/2025_1201_EntITOperatingModels_ThorogoodWoerner
9. OpenAI, "The State of Enterprise AI 2025." https://openai.com/business/guides-and-resources/the-state-of-enterprise-ai-2025-report/

Christopher Kliebenstein builds AI operating models and production content engines for operators at Kliebenstein AI Studio.