The Pilot Graveyard: Why Enterprise AI Keeps Dying Before It Gets Started
The money is real. The urgency is real. The pressure from boards, from investors, from competitors real or imagined, is very real. Global AI spending is forecast to hit $1.5 trillion in 2025, growing nearly 50% year on year, according to Gartner, a figure set to exceed $2 trillion in 2026. And yet, across industries and continents, the same scene keeps playing out: a promising AI pilot is stood up, generates excitement, impresses the room, and then quietly disappears into the organisational equivalent of a drawer that no one opens. No production deployment. No scale. No return on the investment that was supposed to change everything.
The scale of the wreckage is not incidental. S&P Global Market Intelligence’s 2025 survey of over 1,000 enterprises across North America and Europe found that 42% of companies abandoned most of their AI initiatives that year, a dramatic spike from just 17% in 2024. On average, organisations scrapped 46% of AI proof-of-concepts before they ever reached production. McKinsey’s State of AI 2025 report found that approximately 62% of organisations are still experimenting with AI agents while only around 23% report scaling them in production environments. And RAND Corporation, in one of the more sobering assessments of the field, found that over 80% of AI projects fail outright, twice the failure rate of non-AI technology projects.
This is not a technology problem. The technology, in most cases, works. What does not work is the organisation pointing it at the wrong thing, in the wrong way, with no structural foundation to catch what comes next. Three of the most senior technology strategists operating in enterprise AI today have watched this failure repeat itself often enough to have developed a precise vocabulary for it, and none of them are remotely optimistic that the industry has yet understood what it is doing wrong.
“People have underestimated what happens once the AI gets a solution,” said Jessica Constantinidis, Innovation Officer EMEA in the Chief Strategy Office at ServiceNow. The system delivers. The organisation stalls. “The UAT (user acceptance testing) is still manual, the project is waterfall. Whatever follows after AI gives you an idea or a solution is still extremely manual.” What looks like an AI problem is almost always a process problem wearing AI’s clothes. And the consequence is that enormous sums in deployment spend are, in practice, funding the automation of dysfunction rather than the transformation of it.
The Wrong Thing, Done Faster
Here is the version of events that most organisations tell themselves: we identified a use case, we ran a pilot, it underperformed, the technology was not ready. Here is what is more often actually happening: the use case was broken before AI touched it, the pilot succeeded on its own narrow terms, and the organisation discovered, at significant expense, that a competent pilot built on a flawed process produces a flawed outcome at greater speed.
Joe Dunleavy, Regional CTO and Global Head of Dava.X AI Group at Endava, has watched this play out across clients for more than a decade. “People take AI and AI-enable a thing, but not factoring in: is this the right thing to do, or is the process working today so that if you AI-automate it, you’re not in effect just compounding an already existing problem,” he said. The excitement moves faster than the assessment. The pilot is narrow, technically clean, and strategically irrelevant, impressive in a demo and invisible in production. “People’s excitement to do AI is quicker than their ability to assess which is the right use case.”
McKinsey’s 2025 AI survey confirmed this pattern, finding that organisations reporting significant financial returns were twice as likely to have redesigned end-to-end workflows before selecting modelling techniques. The sequencing matters enormously. Constantinidis drew the same conclusion with a sharper edge. “People automate in the same way: ‘send an email with all the details’ versus ‘send the facts.’ Did they change the process? Absolutely not. They changed the medium in how you receive it, but they didn’t change the process.” A spreadsheet pumped into a language model is still a spreadsheet. A broken customer journey automated through an agent is a broken customer journey that now runs faster.
Informatica’s CDO Insights 2025 survey put harder numbers around what is going wrong underneath the surface. It identified data quality and readiness as the top obstacle to AI success, cited by 43% of respondents, tied with lack of technical maturity. Shortage of skills came in at 35%. These are not technology failures. They are operational ones, the kind that no model upgrade resolves on its own.
The deepest version of this problem, Dunleavy argued, is structural. Most enterprise processes were designed around human limitations: the need for handoffs, stage gates, sequential approvals, the rough ceiling of six or seven pieces of information a person can hold in mind at once. None of that applies to agents. “Agents don’t have to follow the same script and don’t have to have the same set of things in place. They can be multifaceted, interact with multiple people at the same time, don’t get tired,” he said. Overlaying agentic AI onto human-shaped workflows is not transformation. It is constraint mistaken for caution. “Don’t design the AI work that you do based on the processes that you’ve had always. They’ve been built on the fact that they were done by people.”
The Methodology No One Has Built
If the use case problem explains why pilots start badly, the governance problem explains why they end badly. Moving from pilot to production requires a set of cross-functional decisions that most organisations have never developed a methodology to make. Levent Ergin, Chief Industry Strategist for Agentic AI, Regulatory Compliance and Sustainability at Informatica, identified this gap as the single greatest structural barrier in the space. “The biggest barrier is actually companies not having the right methodology in place to select the right use case for an agent or for an agentic AI workflow,” he said.
That methodology, where it exists at all, needs to account for far more than technical readiness. Risk classification, data quality, regulatory compliance, cyber security architecture, and integration testing all need to be resolved before an agent meets a production environment. The EU AI Act alone introduces a layered risk classification system that most organisations are not yet equipped to navigate at the speed they are deploying. Ergin described what he calls a “safe AI deployment committee,” a cross-functional leadership team whose job is to ask every question the pilot team did not.
“Are we covered from a legal perspective? Are we covered from a compliance perspective? Have we looked at which internal systems the agent will need to have access to? Have we built the cyber security controls to make sure the agent only has access to what it needs and nothing more?” he said. Without that committee, production is an aspiration, not a destination. “Without having that cross-functional engagement up front, it doesn’t matter how good the pilot goes. You won’t be able to take it into production because you haven’t engaged the right people.”
Constantinidis applied the same governance logic at the level of the individual project. Every AI initiative, she argued, needs a board with a precise composition: someone who understands the technology, someone who knows the operations inside out, a person who sees where the business is moving, and a strategist. All four are required. “Chuck them together and they need to make that decision. It needs to be at least a two-thirds majority,” she said. Crucially, that board needs KPIs for failure as much as success. “Once you hit $50,000 in AI credits, should you stop or should you continue? So it’s a positive KPI but also a negative KPI.” The absence of failure thresholds, she argued, is what turns a struggling pilot into a sunk cost. “If you change your mind 25 different times, it’s going to be an endless project with zero value.”
Ergin reinforced the point from the direction of measurement. Knowing that a pilot succeeded is meaningless if you cannot explain why. “Can you trust the responses you’re getting out of the AI? Can you explain it so that you can actually stand behind the outcomes?” he asked. Testing in isolation, with synthetic data, is not the same as deploying into the operational ocean of live systems, and the gap between those two environments is where the majority of production failures originate.
The operating model implications run deeper still. In the agentic enterprise, Ergin argued, every analyst will need to think like a manager, because the low-value tasks that once defined their role will be performed by agents. “When you turn a process that used to take a couple of days into 10 minutes, that impacts not just that process in itself. It could even give the organisation a completely new market where they’re able to operate,” he said. BCG research has put a number on what that shift is worth, finding that AI-native companies achieve five times higher revenue growth and three times greater cost reductions compared with peers slower to integrate AI into their operating models. The gap between those who have solved the pilot-to-production problem and those still circling it is not theoretical. It is beginning to show up in competitive outcomes.
The Cost of Hesitation
Given the complexity, a quieter argument has started circulating in boardrooms: should organisations simply wait until things settle? Wait for pricing stability, for model consolidation, for the field to mature. Dunleavy engaged with this directly and then disposed of it. “The option of waiting is not a particularly compelling one. Unfortunately this technology is fairly disruptive to everything and anything a business is doing,” he said. He was equally clear that a bubble exists, but drew a careful boundary around it. Infrastructure investment has outrun sense. “We are in a bubble on AI infrastructure and hardware for sure. Infrastructure is perishable. It’s like milk or eggs. You build it today and the chips in there, in three years’ time, will not be fit for purpose.” But a hardware bubble and an AI capability bubble are not the same thing. “Anybody thinking the genie is going to go back in the bottle: the worst AI is today. Tomorrow it’ll be better. The day after that, it’ll be better too.”
Gartner estimated that 40% of enterprise applications will embed AI agents by 2026, up from less than 5% today. IBM, meanwhile, projected that AI-enabled workflows will grow from 3% to 25% of enterprise operations by the end of 2025. Taken together, these figures describe a landscape that is restructuring faster than most organisations are moving. Those waiting for clarity will find themselves redesigning for a world that has already moved on.
Ergin translated the same logic into competitive consequence. “The market will vote with their money, which is really the customer. If I go to one bank who can approve me for a loan in two minutes versus another bank still using their current processes without AI, I’m personally going to go to the one that can give me a response straight away,” he said. The organisations holding out for certainty are not managing risk. They are accumulating it, quietly, while their customers make different choices.
Constantinidis took the argument a stage further. The real question, she argued, is not whether to adopt AI but whether organisations have understood what adoption actually demands. It demands that the AI be given an identity: values, strategy, context, not just a task. “You need more than one AI. The AI you need is your company’s AI, your values, what you want, what you stand for, programmed into it so it has a conscious understanding,” she said. A Claude, a Gemini, a Copilot can be taught rules and regulations, but rules are not the same as institutional DNA. “There is always this handover between the worker and the AI. And that blocks a lot of times, because either the worker doesn’t know how to ask AI or AI doesn’t know how to actually follow up on what the worker needs to do.”
The pilot graveyard, in the end, is not filled with technology that failed. It is filled with organisations that never quite decided what they were building, or why, and discovered, at significant cost, that AI has no answer to that question on their behalf.