IBM recently surveyed more than 2,000 CEOs worldwide and the headline number is hard to ignore: only 25% said their AI initiatives delivered the ROI they expected. Three out of four AI investments underperformed. Not because the models were weak or the vendors were dishonest — but because the organizations deploying AI skipped the foundational work that determines whether any technology investment pays off.
The “random acts of AI” problem
The phrase that keeps surfacing in enterprise AI circles is “random acts of AI” — a pattern where companies launch disconnected pilots across departments, each solving a different problem in a different way, with no shared framework for measuring success and no connection to the metrics that actually drive the business.
It looks productive. Teams are experimenting. Budgets are moving. Vendors are engaged. But none of it compounds. Each pilot exists in isolation. When one succeeds, nobody can explain why in terms that transfer to the next project. When one fails, the lessons die with it because there was never a shared diagnostic framework to begin with.
This pattern is not limited to companies that are new to AI. Some of the most sophisticated technology organizations in the world have reported the same dynamic. The scale of AI spending — Gartner projects $2.59 trillion globally in 2026 — means the waste is industrial-grade. And the root cause is remarkably consistent: organizations are choosing what to automate based on enthusiasm, vendor pitches, or the loudest internal complaint, rather than on a rigorous assessment of where AI actually creates value.
Why the failure rate is so high
The IBM finding sits alongside a body of research that tells the same story from different angles. MIT found that 95% of generative AI pilots produce no measurable profit. S&P Global reported that 42% of companies abandoned most of their AI projects in 2025. Researchers attributed the gap not to model capability but to “poor workflow design, not weak AI models.”
That last finding is the critical one. The technology works. GPT-class language models, computer vision systems, and agentic frameworks are genuinely capable. The failure happens upstream — in the decision about what to point them at and the preparation that should precede deployment.
Three specific patterns account for most of the 75% that underperform:
No process map exists
Most organizations have never documented their actual workflows end to end. They have assumptions — org charts, job descriptions, standard operating procedures — but these describe how the company imagines it operates, not how it actually operates. The real work includes workarounds that started as temporary fixes and became permanent, handoffs that stall in places nobody monitors, and steps that exist only because a system limitation from five years ago forced them into existence.
When you automate a process you have not mapped, you are automating your assumptions. The AI faithfully executes a workflow that was broken before it arrived — faster, at scale, and harder to undo.
No tie to a P&L lever
AI initiatives that cannot be connected to a specific revenue, cost, or risk lever are noise. They may produce impressive demos. They may generate internal excitement. But when the CFO asks what the investment returned, the answer is a narrative, not a number.
This is not a reporting problem. It is a design problem. If the project was not scoped against a P&L outcome from the beginning, no amount of post-hoc analysis will produce a convincing ROI case. The measurement has to be designed into the initiative before the first line of code is written.
Too many pilots, too little focus
The temptation to launch 20 experiments across the organization is understandable — it feels like maximizing optionality. In practice, it maximizes distraction. Each pilot requires executive attention, change management, integration work, and evaluation. Spread across 20 efforts, none of them gets enough of any of those resources to succeed.
The companies that land in the successful 25% do not have better AI. They have better discipline: fewer bets, deeper preparation, and a clear connection between every initiative and a measurable business outcome.
The five-point prescription
The pattern behind organizations that do achieve ROI from their AI investments is not mysterious. It is not a proprietary framework or a secret technology stack. It is a set of practices that successful teams follow consistently — and that the 75% consistently skip.
1. Map end-to-end workflows before touching technology
Before selecting tools, before writing requirements, before talking to vendors — map the work. Not the org chart version, not the process document that lives in SharePoint. The actual work: who touches what, in what sequence, with what tools, over what timeframe, and at what cost.
This mapping requires structured interviews with the people who do the work, not just the managers who describe it. It surfaces the gap between how an organization thinks it operates and how it actually operates. That gap is where the highest-value automation opportunities live — and it is invisible without the map.
A thorough mapping exercise across an organization takes two to three weeks. It is not glamorous work. It produces a document that looks more like an engineering diagnostic than a strategy deck. And it is the single highest-return investment an organization can make before spending a dollar on AI technology.
2. Tie every AI initiative to a P&L lever — or kill it
Every project in the automation catalog should be connected to one of three outcomes: revenue increase, cost reduction, or risk mitigation. If a proposed initiative cannot be tied to one of those levers with a specific, quantified projection, it should not be funded. Not because it is a bad idea, but because without that connection, there is no way to measure whether it worked and no way to justify continued investment.
This discipline eliminates a surprising number of projects that feel urgent but are actually low-impact. The process that generates the most complaints might consume 2% of total labor hours. The quiet process that nobody talks about might consume 30%. Without the map and the P&L connection, companies automate the complaint, not the opportunity.
3. Pick 3–5 enterprise priorities, not 20 scattered pilots
Concentration beats diversification in AI deployment. The organizations that succeed pick a small number of high-impact, well-scoped targets and give each one the resources, attention, and governance it needs. They build one, measure it against the baseline, prove the ROI, and then expand.
This approach has a second benefit: it builds organizational confidence. When the first project produces measurable results, the second project gets better internal support, faster adoption, and less resistance from the teams whose workflows are changing. That compounding effect is impossible when 20 pilots are competing for attention and none of them has enough support to succeed convincingly.
4. Design for humans in the loop and governance from day one
AI governance is not a phase you add after the system is built. It is a design constraint that shapes every architectural decision from the beginning. Which actions does the AI take autonomously? Which require human approval? What happens when the model produces an output that falls outside expected parameters? Who reviews the logs? How are decisions audited?
These questions are not bureaucratic overhead. They are engineering requirements. An AI system without governance is not deployed — it is released. And the first time it produces a bad output with real consequences, trust evaporates. Rebuilding trust after a governance failure costs more than building governance correctly in the first place.
The governance framework should include scoped AI sessions with controlled data access, human-in-the-loop gates on consequential decisions, auditable activity logs, and clear escalation paths. This is not optional infrastructure. It is what separates a system your organization can depend on from one it merely tolerates.
5. Build model-agnostic architecture so you are not locked to today’s technology
The AI landscape changes quarterly. Models that are state-of-the-art today will be superseded. Vendors that dominate today may pivot. Pricing structures will shift. An architecture that is tightly coupled to a specific model or vendor is an architecture with an expiration date.
Model-agnostic design means building your automation logic, your data pipelines, your governance layer, and your evaluation framework in a way that allows you to swap the underlying model without rebuilding the system. It means treating the model as a component, not as the foundation.
This is not theoretical prudence. It is practical engineering. We have seen organizations locked into model-specific architectures that could not adopt better, cheaper alternatives when they became available — because the entire system would need to be rebuilt. The switching cost exceeded the value of the improvement. That is an avoidable trap, but only if model independence is a design principle from the start.
The diagnostic before the surgery
We describe this entire approach with a simple analogy: run the diagnostic before the surgery. No competent surgeon operates without imaging, lab work, and a clear understanding of what they are looking at. The diagnostic does not slow down the treatment — it makes effective treatment possible.
The same logic applies to AI investment. A structured discovery phase — process mapping, stakeholder interviews, ROI scoring, governance planning — produces the operational intelligence that makes every subsequent dollar of AI spending measurably productive. It identifies the three to five initiatives that will actually return their investment and names the projects that sound exciting but would quietly fail.
The IBM survey confirms what we have observed over 12 years of enterprise AI work: the gap between the 25% that hit ROI and the 75% that do not is not budget, talent, or technology. It is preparation. The organizations that map first, score by ROI, concentrate their bets, and build governance from day one are the ones that extract real value from AI. Everyone else is funding random acts of AI — and hoping that volume compensates for strategy.
It does not.