AI Pilot Purgatory
AI is probably the first technology in history that, at the moment of use, systematically asks you to doubt it. That organizational paradox — deploying a tool that warns you not to trust it — maps directly onto why most AI projects stall before they ever reach production.
AI is everywhere: in press releases, in board presentations, in budgets approved with enthusiasm in January and already under review by June. It has become the mandatory reference point for any conversation about the future of the enterprise, the term that transforms an ordinary roadmap into a strategic vision, the word that in the right slide is worth a round of applause on its own.
“Everywhere,” however, does not mean much: right now, AI is in production in only 8.6% of companies.
The figure comes from the Recon Analytics AI Pulse, a continuous survey of over 120,000 organizations conducted between March 2025 and January 2026. 63.7% do not have even a single formalized AI initiative. 14% are in the pilot phase: that place where projects exist, consume resources, and are presented to stakeholders with reassuring slides, without ever taking the next step. Analysts call it pilot purgatory, and the name is more accurate than it might seem: there is something theologically precise about the idea of a place where one remains suspended, neither damned nor redeemed, waiting for a resolution that never arrives.
What follows is an attempt to understand why organizations end up there, and above all how they get out.
The Data Problem: The Fuel That Is Not There
The Lucidworks “Dawn of the Agentic AI Era” report, which combines a survey of over 1,600 AI leaders with an independent direct assessment of more than 1,100 companies, arrives at a conclusion that has the merit of being uncomfortable: 65% of organizations do not have the foundations necessary to build meaningful AI into their platforms. Lucidworks CEO Mike Sinoway uses an automotive metaphor: trying to build a Formula One car around a go-kart engine, with the tank nearly empty. It is an image that works, even if the reality inside most companies is often even more prosaic.
Knowledge bases are distributed across ten different systems, none of which communicates with the others. Documents exist in triplicate, in different versions, in repositories that no one updates systematically. Historical data is there somewhere, but extracting anything useful from it requires a project in itself. In this context, AI does not fail because the model is wrong or because the vendor overpromised. It stalls upstream, on the oldest and least glamorous problem in enterprise technology: data quality and consistency.
The starting point, therefore, is not choosing an AI platform. It is understanding what you actually have, and what is missing. A well-executed assessment at this stage is worth more than any benchmark on available models.
The Governance Problem: Who Is Responsible for This Thing?
The Deloitte “State of AI in the Enterprise 2026” report, conducted with 3,235 leaders across 24 countries between August and September 2025, captures a situation in which enthusiasm for agentic AI — systems capable of planning, deciding, and acting autonomously — is running far ahead of the capacity to manage it. Only 21% of organizations have a mature model for governing agentic AI. The remaining 79% are in practice deploying tools that make decisions without having defined who is responsible for those decisions, how they are verified, and what happens when something goes wrong.
This figure intersects in an interesting way with what emerges from the Global Trust Study 2025, conducted by the University of Melbourne in collaboration with KPMG across 48,340 people in 47 countries: nearly half of employees admit to using AI in ways that violate company policies, and more than half present AI-generated outputs as their own work without disclosing it. Only 40% of workplaces have a formalized policy or guidance on the use of generative AI.
The picture that emerges is one of organizations that have distributed access to tools without distributing accountability for their use. Governance, in this sense, is neither an IT topic nor a legal one in the strict sense: it is the precondition for any AI project to scale without becoming an operational risk. Defining who controls what, how outputs are audited, which data can be used and in which context — these are not bureaucratic questions. They are the questions that separate projects that survive contact with reality from those that stop at the first incident.
The Integration Problem: The AI Living in a Drawer
Back to Deloitte. 37% of the organizations surveyed use AI at a superficial level, without intervening on the underlying processes. The image that emerges is that of a tool added on top of an existing workflow without touching it: a chatbot that answers generic questions, a tool that summarizes documents, an assistant that helps write emails faster. All useful, certainly. But all light-years away from AI that transforms an operation, accelerates a supply chain, reduces response times in a contact center, or lowers risk in a predictive maintenance process.
The problem, in most cases, is not the quality of the chosen model. It is that the model is not connected to anything real: it does not talk to the ERP, it does not know the contracts, it has no access to the operational knowledge base, it knows nothing of the specific workflows of that organization. It is like hiring a brilliant consultant, keeping them in isolation, and then being surprised that their advice sounds generic.
The Traceability Problem: Trust Is Not Enough
The most underrated finding from the Melbourne/KPMG Global Trust Study is this: 66% of people use AI with some regularity, but only 46% are willing to trust it. Trust, in other words, does not grow with use. In the period from 2022 — before ChatGPT — to today, adoption has exploded and trust has declined. People use tools they do not trust, because they have no alternative, or because organizational pressure is stronger than individual skepticism.
There is something structurally peculiar about this: AI is probably the first technology that, even in production, systematically and invariably hedges on its own functioning. Just below the input field, any mainstream tool warns that the system can make mistakes and that responses should always be verified. It is not a disclaimer buried in the footer: it is the first message every user sees, every time. No other technology has ever communicated its own limitations so explicitly at the moment of use. The result is an organizational paradox: adopting a tool that, first and foremost, asks you to doubt it.
In companies, this translates into something very concrete: 83% of AI leaders surveyed by Lucidworks report “major” or “extreme” levels of concern about their progress — an eightfold increase since 2023. The concern is not about the power of the models, which in the meantime has grown spectacularly. It is about control: knowing what produced a given output, why, on the basis of which data, with what margin of error.
In regulated environments — finance, healthcare, public administration, critical infrastructure — this is not an abstract concern. An AI output that cannot be audited is not simply unreliable in certain contexts: it is unusable. Traceability is not an accessory feature to be added later. It is an architectural requirement that must be designed in from the beginning, or it does not exist at all.
Getting Out of Purgatory
Deloitte 2026 registers an encouraging signal: the number of organizations that have moved 40% or more of their pilots into production is expected to double in the next six months. The momentum is there, and Recon’s data confirms the direction: the share of companies with agentic AI in production rose from 7.2% in August 2025 to 13.2% in December 2025. Something is moving, even if the starting point was low enough to make any progress relative.
What distinguishes those who scale from those who stand still, however, does not emerge from data on which models were chosen or how much was invested. It emerges from the quality of the work done beforehand: data assessment, clear definition of use cases, building governance that works as an operating system rather than a document in a shared folder, genuine integration with existing processes, and output traceability. Pilot purgatory is the natural consequence of having skipped steps in the wrong order. Getting out requires reversing that sequence: slowing down intelligently, in order to accelerate sustainably.
How We Approach This at Geckosoft
When an organization brings us in on an AI project, the first thing we do is resist the temptation to answer immediately. The question that arrives is almost always technical — which model to use, how to integrate an LLM into existing processes, how to build an internal assistant — but the useful answer rarely is. It is, instead, analytical: understanding where the data is, who owns it, how it is governed, which processes genuinely need artificial intelligence and which ones simply need to be redesigned.
This is the work we do at Geckosoft before touching any tool. And only after that — only when the context is clear and the use cases are defined with precision — do we enter the build phase.
Rational AI is the platform we developed to address the four problems described in this article in an integrated way: centralized governance, a coherent and queryable knowledge base, output traceability, integration with existing systems without requiring what already works to be dismantled. For organizations operating on distributed supply chains or managing critical infrastructure, Everplan and Kairos extend this approach to the most complex operational processes: those where the margin for error is low and the pressure on continuity is high.
It is not a proposal for everyone, and we make no claim that it is. But if you are reading this article with the feeling that it describes a familiar situation, it is probably worth a conversation.
Originally published on Geckosoft website, translated with Sonnet 4.6 through Rational AI.