We use cookies to keep this site running. Analytics cookies are optional and off by default. Privacy Policy
ALI Insights
Why most AI initiatives never reach the income statement, and what boards and CEOs must own to change that.
It starts with the way these things always start. Another round of strategy and budget planning meetings. Every leader with a deck, a case to make, and a quiet hope that no one presses too hard on the numbers. AI is on every slide this year. Move or fall behind. Do more with less. Buy or build. The same scarce capital, more hands reaching for it, each function certain its initiative is the one that makes the company a market leader.
Then a board member sets down the deck and asks the simplest question in the room. "What are we doing with AI that actually moves the needle?" A pause. "And I mean the income statement. Not pilots, not press. Revenue or cost. What has actually moved?"
For a moment, nobody has a clean answer. There is plenty of activity to point to. Pilots, proofs of concept, a budget line that grew. But the honest answer, the one no one wants to give first, is that almost none of it has changed anything that shows up in the numbers.
That silence is the real state of enterprise AI right now. And it is worth being precise about what it means because the usual explanation is wrong. The projects underneath that silence did not stall because the technology failed. The models mostly work. They stalled because of what happened, and did not happen, in rooms exactly like that one.
The risk is not that AI does not work. The risk is that it works, at scale, on the wrong problem, without control or ownership.
This guide is about the part of AI that does not live in the technology. The judgment about where it belongs, the people who must change how they work, and the controls that decide whether any of it survives contact with reality. Those three things are the executive's job, and they are the three things that no vendor, no model, and no consultant can own on your behalf. At ALI Consulting, we have witnessed this pattern repeatedly across executive teams: the failure is rarely the technology.
The board member's instinct was correct, and the data backs it. In 2025, S&P Global Market Intelligence surveyed more than a thousand enterprises across North America and Europe and found something that should stop any leadership team cold.
That is not a technology failure curve. That is a decision-making failure curve. Abandonment more than doubled in twelve months, not because the models got worse, but because organizations kept funding initiatives that were never going to reach the income statement and eventually had to admit it.
A separate study makes the same point from the other direction. MIT's 2025 research, analyzing three hundred deployments, found that roughly 95 percent of enterprise generative-AI pilots delivered no measurable impact on profit and loss. Only about five percent produced rapid revenue gains. The authors were direct about the cause. It was not model quality. It was the gap between the tool and the organization, the integration, the workflows, the ownership. In other words, the part that was the company's job, not the technologies.
For a board, those two numbers reframe the whole conversation. The question is no longer "are we using AI." Almost everyone is. The question is whether any of it is material and whether anyone in the building can prove it.
The most common mistake is upstream of any specific project. AI gets framed as a technology decision when it is really an operating-model decision. That framing rarely originates with the CEO. It arrives through the existing ecosystem and the feedback loops around it, where the problem gets quietly scoped as a purchase, and the broader view never makes it into the room.
The tell is in the questions that come first. What tools should we buy? Which model provider? Should we run a few pilots? Should we run more than one for resiliency? Every one of those assumes the value is in the tool, that the risk is mostly technical, and that adoption is incremental. For AI, all three are wrong.
AI changes the unit economics of the work itself. When a task drops from ten dollars to fifty cents, and from thirty minutes to thirty seconds, the win is not doing that task faster. It is eliminating it, combining it, or rebuilding the workflow around it. Treat that as a tool purchase, bolt it onto the old process, and the income statement never moves. That is how you end up in the 42 percent.
This is also why the demo is so dangerous. A board sees an impressive AI presentation and reasonably assumes the hard work is done because surely it would not be presented otherwise. But a demo is built to show capability and hide reliability. It runs on clean, curated data. It does not show the variance, the silent failures, or how badly the thing degrades when the inputs are real. AI is powerful, and unreliable, at the same time. The gap between those two is where most of the money goes to die.
Suppose the framing is right and the technology is sound. There is still a failure mode that sinks more initiatives than any model limitation, and leaders consistently underestimate it. It is not communication. It is identity.
When AI arrives, people do not think "I need new skills." They think "Am I still valuable here? Is my job safe?" That is a different problem, and it does not yield to the standard playbook of clear messaging and training. Those tools assume the barrier is understanding. The real barrier is status, and the fear of being exposed.
The pattern is consistent, and counterintuitive. The most capable people are often the most resistant because the thing that made them valuable is the first thing the machine does well. Training completion runs high. Behavior barely moves. In private, the language shifts from "this is confusing" to "this makes my role unclear." It is not a skills gap. It is a threat to how people define their own worth.
The instinct to reassure ("humans plus AI, not replacement") is incomplete, because employees hear "not yet." What works is redefining value out loud. A great analyst is no longer the fastest at first-pass analysis. A great analyst is now the one who can frame the problem, direct the AI, and catch where it is wrong. If leadership does not define the new standard, people define it for themselves, and they always assume they are losing. The contact center is the cautionary tale: "AI will replace the agents" is both wrong and expensive. The work changes, the advancement is new, but the failure to lead people through it is very old.
There is a predictable second act. The board aligns, the budget unlocks, the organization buys in. Then every line of business discovers a critical use case, each argues its own is the priority, and the urgency narratives spike. We are falling behind. It feels like momentum. It is fragmentation, and the loudest voices win instead of the highest-value ones. That road leads straight back to a pile of pilots and no measurable impact.
At this stage, the failure is not strategy. It is governance discipline under pressure. The organizations that hold the line do one thing clearly: they separate who goes first from who benefits first. Going first is a sequencing decision, ranked on income-statement impact, workflow maturity, data readiness, and risk. So, when a leader insists on priority, the answer is not political. It is "here is where you rank, and here is why." They treat the first projects as capability-building, because those set the patterns and guardrails everything else inherits. And they constrain capacity on purpose, because limiting the number of active initiatives is what forces real prioritization. Without that constraint, everything looks important and nothing gets owned.
A CEO does not need to become an engineer. A CEO needs a reliable way to tell a real capability from a confident demo. The market makes this harder than it should be.
That figure comes from healthcare, but the principle is universal: you cannot verify what you cannot audit. Most AI tools look good in a controlled setting because they optimize for fluency, completeness, and speed. They fail in the real world because they lack auditability, a control framework, and an accountability model. Here is the filter for the room.
That filter is not theoretical. It comes from a quarter century of building technology inside regulated financial institutions, where being wrong did not mean a bad quarter, it meant a finding from the Office of the Comptroller of the Currency. In that world you internalize one rule: if you cannot explain it, trace it, and defend it, it does not exist. Every decision must answer what data it used, what logic it applied, what controls were in place, who reviewed it, and whether it can be reproduced. If any of those answers are weak, you are exposed.
So, the first question about an AI system is not "is it impressive." It is "would it survive an audit." That is a far higher bar, and it leads to one design principle that should govern every deployment: the model is not the system. The system is everything built around the model to make it auditable.
This is the discipline behind the AI that ALI builds. ALI Knowledge Intelligence turns a body of organizational knowledge into cited plain-language answers. ALI Contract Analyzer surfaces contract risk for attorney review. Neither was designed to be clever. Both were designed to be trustworthy where being wrong is expensive. In practice that means a few specific things, and they are the same things a board should expect of any AI it relies on. Citations are a control, not a feature: no assertion without a traceable source. Human oversight is formal accountability because accountability cannot be handed to a probabilistic system. The boundaries are explicit, so "here are the relevant clauses and risks" is in scope and "you should sign this" is not. And the system fails closed, not open: when confidence is low, it stops, because a confident wrong answer is worse than no answer at all.
The mindset is the whole point. Most of the market asks how to make AI more powerful. The regulated instinct asks how to make it safe enough to trust when the stakes are real. Those are different goals, and they build vastly different systems. One of them survives the board member's question. The other becomes part of the 42 percent.
This is the part to bring into the room. One anchor question, ten areas to probe, one closing question. The goal is not to collect answers. It is to find out whether management has a coherent story, one where the problem, the economics, the data, the controls, and the path to scale actually connect. If they do not, you will get activity, and you will not get outcomes.
Anchor question: If this system is wrong at scale, how bad does it get, and how quickly would we know? What is the financial, regulatory, and reputational cost of failure, how fast would we detect it, and what stops or contains it?
Closing question: Where, specifically, will this show up in our financials, and when? What metric moves, by how much, and on what timeline?
Read those two bookends together, the anchor and the close, and you have the board member's original question made rigorous. What are we doing with AI that moves the income statement, and can we prove it survives scrutiny. Everything in between is how a disciplined organization earns the right to answer yes.
ALI Consulting advises CEOs and boards on exactly these decisions, and the perspective here is not theoretical. The firm's founder spent twenty-five years as a chief information officer inside regulated financial institutions, turning technology from a cost center into a measurable business driver: more than thirty-five million dollars in enterprise savings, cybersecurity exposure cut from over ten thousand vulnerabilities to under a thousand, and zero repeat findings across OCC, FDIC, and CFPB examinations. He led two bank merger integrations, replaced a projected seventeen-million-dollar integration with a reusable-services model delivered for roughly one and a half million, and was brought in to rescue a regulatory remediation two prior teams had failed to close, resolving it inside fifteen months. ALI also builds production AI, which is why the advice on telling signal from noise comes from shipping systems that have to survive scrutiny, not from watching demonstrations. If your organization is weighing where AI actually belongs, that is the conversation we are built for.
Start with one anchor question: if this is wrong at scale, how bad does it get and how fast would we know? Then test for a coherent story linking the problem, the income-statement impact, the data, the controls, and the path from pilot to production. If management cannot say where it shows up in the financials and when, it is not ready for approval.
It is a business decision, specifically an operating-model decision. Treating it as a tool purchase is the most common reason initiatives never reach the income statement. AI changes the unit economics of work, so the real question is how the work gets redesigned around it, not which model to buy. The technology choice is real but downstream.
Rarely because the model is bad. Pilots are run as experiments with no path into core workflows, systems of record, or accountability, so nothing scales. And prioritization is driven by internal pressure rather than economic impact, so effort fragments. S&P Global found AI abandonment jumped from 17 to 42 percent in a single year. The fix is sequencing discipline and embedding AI in real workflows.
Ask how it fails and what happens when it is wrong. Real systems answer that, show their work with traceable citations, keep a human accountable for consequential decisions, and prove themselves on messy real-world data rather than a clean demo. Confident black-box claims with no error handling are the clearest warning sign.
Recognize that you are managing identity and job-security fear, not a skills gap. People are asking whether they are still valuable. Redefine out loud what good looks like in the new model, create low-risk room to experiment, lead visibly rather than delegating, and tie the change to individual upside, not just company efficiency.
The model is not the system. Trust comes from what is built around it to make it auditable: citations as controls, formal human oversight, explicit boundaries on what the system claims, validation around the probabilistic core, and a fail-closed design that stops when confidence is low. The test is simple: would it survive an audit?
If you want to answer the board's question with confidence, a conversation with ALI Consulting is the place to start. We work from the operator side of the table, where we were accountable for the outcomes, not advising from the outside. If you cannot answer that question clearly today, your board will ask it soon. We will help you walk into that conversation with a defensible answer. No charge, no obligation. Start the conversation.
Start With a Conversation
Whether you are setting AI strategy, governing risk at the board level, or trying to move from pilots to measurable impact, we can help. Let's start the conversation.
Let's Talk →