The Real Reason Your AI Initiative Stalled Has Nothing to Do With the Model
- Michael Privat
- 2 hours ago
- 7 min read
I have spent more than two decades inside healthcare IT, one of the most compliance-heavy, data-dense environments in enterprise technology, and I can tell you this with complete confidence: the AI failure epidemic is not a technology problem. It never was. I lead an organization of over 500 engineers across the US and India, and I have watched the AI gold rush play out in real time, from the breathless launch announcements to the quiet, uncomfortable budget-review conversations that come after. The pattern repeats itself almost everywhere, and it points to the same place every time: leadership.
Here is the number that should be keeping executives up at night. Ninety-five percent of enterprise generative AI pilots are delivering zero measurable P&L impact, despite somewhere between $30 and $40 billion in enterprise investment behind them (Source: Yahoo Finance, 2025). MIT's own research is careful to note that the failure is not about model quality or regulation. It is about what organizations do, and more precisely, what their leaders decide before a single model is deployed. At the same time, 74% of organizations say they want AI to grow revenue, while only 20% have actually seen it happen (Source: Deloitte, 2026). That gap between ambition and reality is where careers and budgets get quietly eaten alive.
So the question worth sitting with is not "why is AI failing?" It is "who made the decisions that guaranteed it would?"
The Pilot Was Designed to Feel Like Progress
Most AI initiatives kick off with a genuine burst of excitement, and honestly, that excitement is the first warning sign. Not because enthusiasm is a bad thing, but because excitement has a habit of bypassing the hard conversations about what success is actually supposed to look like. A pilot gets greenlit. A vendor gets brought in. A working demo gets built. Someone in the room calls it "promising." And then, about six months later, the organization realizes it has a very polished demo that has never touched a revenue line or a cost center in any meaningful way.
This is not a technology story. It is an accountability story. Pilots are structured, almost by design, to succeed in a narrow sense: small team, clean data, limited scope, and very little organizational pressure to perform. Boston Consulting Group surveyed more than a thousand senior executives across 59 countries and found that only 4% of organizations consistently generate value from AI across functions (Source: BCG, 2024). The 4% are not working with fundamentally better models. They are working with fundamentally better questions. They established what success meant before the pilot was worth running, and they defined "measurable impact" before anyone wrote a single line of code.
Seventy-three percent of failed AI projects had no agreed definition of success before work began, and 61% were approved on projected ROI that was never actually measured after launch (Source: MIT Sloan, 2025). BCG adds that 60% of companies never defined or monitored financial KPIs tied to AI value creation at any point in the process (Source: BCG, 2025). Read those two numbers together and a clear picture forms: the other 96% did not fail at execution. They failed at the decision that comes before execution, the moment where a leader should have asked what success looks like, who owns it, and when the organization expects to know. That conversation got skipped, replaced by a directive to "explore AI" and a budget line that nobody wanted to challenge. That is not a strategy.
I have watched this play out from every angle. The teams I work with are talented and genuinely motivated, but even the best engineers cannot manufacture a business outcome that leadership never defined in the first place. When the mandate is vague, the delivery will be vague. When accountability is spread across a committee rather than owned by one person, urgency disappears fast. These are not engineering failures. They are leadership failures that happen to show up in engineering.
The Problem Nobody Wants to Say Out Loud
Even when an organization manages to push an AI initiative past the pilot stage, a second challenge tends to surface, and most leaders are not ready for it: managing teams whose work is now heavily shaped by AI-generated outputs. It is a subtle shift, which is exactly what makes it so easy to miss until it becomes a real problem.
Output has gotten cheaper to produce. A developer generates code faster. A knowledge worker drafts faster. A data analyst surfaces patterns faster. On the surface, this reads as a productivity win, and in some ways it genuinely is. But what gets lost in that framing is this: faster output is not the same as better output, and the speed at which something gets produced has almost no relationship to the judgment required to evaluate whether it is actually correct.
Eighty-one percent of leaders report finding AI investments hard to quantify, and 79% say untracked AI budgets are becoming a real accounting problem (Source: Larridin, 2025). A big part of that measurement gap is cultural. Leaders built their productivity expectations around output volume and speed, both of which are now artificially inflated by AI assistance. What stops getting measured is quality, accuracy, and judgment, which are precisely the things that matter most as AI handles more of the mechanical work.
Think about what this looks like in practice. If an engineering team is using AI tools to produce twice the code in the same amount of time, a surface-level read of the metrics says things are going great. A more honest read asks: who is reviewing that code with the same rigor the original pace required? Who actually owns the quality of what gets shipped? At too many organizations right now, the answer is nobody in particular, and when something is nobody's explicit job, it does not get done with any consistency.
Performance standards need to evolve to reflect this reality. In an AI-assisted environment, what someone produces is increasingly a commodity. What is not a commodity is their ability to catch what the AI got wrong, to understand the reasoning behind a recommendation before accepting it, and to apply domain knowledge that no model has been trained to replicate. Those are the capabilities worth developing and measuring. Leaders who are still evaluating their teams primarily on output volume are measuring the wrong thing, and they typically will not realize it until something breaks in production or, worse, in front of a client.
The organizations actually extracting value from AI, the 5% that MIT's research identifies as generating measurable P&L impact, share one consistent trait: they define success criteria before deploying, build measurement into the system from day one, and hold specific people accountable for outcomes. Governance that arrives after deployment is not governance. As research from Boston University points out, by the time a governance review happens on a live system, accountability gaps have already formed and the cost of fixing them is substantially higher than it would have been if oversight had been built in from the start (Source: Boston University Questrom School of Business, 2026).
The difference between leaders who are winning with AI and those stuck in pilot purgatory is not access to better technology. It is the decision to treat accountability as a non-negotiable part of every initiative, not something to figure out later.
What the Leaders Getting This Right Actually Do
The leaders I have seen navigate this well share a few habits that are easy to describe and genuinely hard to stick to.
They decide before they build. Before any AI initiative gets a budget, they define the specific business outcome they are going after, the metric that will confirm whether it happened, the person who owns that metric, and the date when they will have an honest answer. Not a demo. Not a presentation. An actual number that moved. This sounds straightforward, and it is almost never done.
They treat adoption and value as two completely different things. Usage numbers, seat counts, and completed pilots are not indicators of success. They are indicators of activity. The organizations pulling real returns from AI measure value realization directly, connecting AI performance to KPIs that exist independently of the AI itself. Revenue influenced. Cost reduced. Time saved at a specific point in a specific workflow. Anything short of that is a narrative, not a result.
They change how they evaluate their people. This is the most uncomfortable shift, and it is the one most leaders quietly avoid because it means rewriting the unspoken agreement between management and their teams. Evaluating an engineer on code output alone when they are using AI assistance is a bit like evaluating an airline pilot only on whether the flight departed on time, while ignoring whether they would know what to do the moment the automation failed. In an AI-assisted workplace, judgment, critical thinking, and the ability to verify what AI produces are not soft skills. They are core technical requirements.
So, Is Your Next Initiative Actually Ready?
Before the next AI initiative gets approved, one question is worth asking out loud: can you name the person accountable for the outcome, define the metric that will confirm it worked, and commit to a date when you will have an honest answer? If those three things do not have clear answers before the project starts, the initiative is not ready. The technology is not the problem. It never was.
The organizations that will look back on this period as the moment they pulled ahead are the ones being honest with themselves right now about what leadership failure actually looks like. It does not look like a bad model. It looks like a kickoff meeting where nobody asked hard enough questions. That is a conversation I find myself having constantly, and if it is one you want to continue, you can find me on LinkedIn or follow along on Substack.
References
Michael Privat, Chief Data and Engineering Officer, Availity
Michael Privat is a Chief Data and Engineering Officer leading a global team of 500+ engineers. With 25 years in tech, he helps organizations unlock speed, clarity, and accountability. He turns stalled engineering teams into high-performing systems built on ownership, discipline, and modern AI-driven workflows through his “accountable autonomy” model.
