Eighty-one percent of enterprise AI initiatives are failing to meet their business goals. That's not a pundit's pessimistic take — it's from CIO.com's 25th annual State of the CIO survey, which canvassed 662 IT leaders and 249 line-of-business executives in 2026. Only 19% say their AI programs have met or exceeded expectations. Another 18% admit fewer than one-third of their use cases are hitting defined targets.
The irony is that most organizations aren't short on AI projects. They're drowning in them.
The Pilot Purgatory Problem
If you've sat in a quarterly business review lately, you've heard some version of this: "We've run 40 AI pilots this year. Twelve are in extended evaluation. Six are in POC stage 2. Three are being presented to the steering committee." And when someone asks how many are in production delivering measurable value? The room gets quiet.
Researchers have a name for it: pilot sprawl. Organizations spin up isolated experiments that never scale — fragmented tool ecosystems, disconnected workflows, and data foundations too weak to support production-grade AI. A recent Teradata study of 1,000 senior technology and data leaders found that only 7% of enterprises have reached enterprise AI scale. Seven percent.
The problem isn't that enterprises lack enthusiasm for AI. It's that enthusiasm has outpaced execution discipline.
"We are saying yes to everyone without stepping back and focusing on the business cases that show real value," noted Andrea Ballinger, CIO at Rensselaer Polytechnic Institute, in the CIO.com survey. "No one is measuring ROI on an ongoing basis because we are facing counterpressures from every vice president and line-of-business domain looking to implement AI for their own optimization."
That sentence describes the majority of enterprise AI programs in 2026.
Three Root Causes (and None of Them Are Technical)
The State of the CIO survey surfaced three organizational barriers that keep showing up — and none of them are about the models or the infrastructure.
1. Ill-defined ROI metrics (32%) — Almost a third of CIOs cite fuzzy success criteria as a primary obstacle to scaling AI. When you don't define what "success" looks like before you start, you can't measure progress, can't justify continued investment to the CFO, and can't kill projects that should be killed. AI projects without measurable targets run indefinitely on hope.
2. Murky corporate AI strategy (31%) — Nearly as many cite unclear organizational strategy as their blocker. This usually manifests as competing AI agendas across business units with no central prioritization framework. Sales wants AI for pipeline forecasting. Finance wants it for contract risk analysis. HR wants it for recruiting. Legal wants it for contract review. All simultaneously, with no agreed-on sequencing or shared data infrastructure.
3. Lack of in-house expertise (40%) — The most-cited barrier is talent. Four in ten CIOs say they don't have the people to execute effectively. This isn't just a technical talent gap — it's also a gap in AI product management, business case development, and change management. You can hire all the ML engineers you want; without people who can translate between business needs and technical capabilities, pilots will keep stalling before production.
Notice what's missing from that list: model quality, compute costs, vendor selection. Those are solvable with money. The real blockers are structural and organizational.
The Technical Reality CFOs Aren't Hearing
Here's a conversation I've had with business leaders more than once: "The pilot showed a 3x productivity improvement. Why isn't the full rollout delivering that?"
The answer usually comes down to what TIAA's chief operating, information, and digital officer Sastry Durvasula described in the survey: "What's on paper sometimes doesn't turn into real ROI given the reality of operational costs. Something may prove to be a successful pilot, but you need to understand the full cost of operations — for example, the efficiencies of running tokens or how you're handling traffic or RAG."
Let me translate that for non-technical executives.
A pilot runs against a small, clean dataset with one or two developers closely managing the system. Production runs against messy, real-world data at scale, with no babysitting, handling edge cases the pilot never encountered, and serving hundreds or thousands of users simultaneously. The token costs (what you pay the AI model per query) multiply with scale. The retrieval costs (pulling relevant context for each query) add up. The infrastructure costs — monitoring, logging, failover, security controls — weren't part of the pilot budget.
A pilot that costs $15,000 to run for three months can easily translate to $400,000 per year in production. If the business case was built on the pilot's cost structure, the CFO is going to ask hard questions when the real invoice arrives.
For technical leaders: document your total cost of ownership before any production recommendation. Break out model inference costs, retrieval/embedding costs, infrastructure, security, monitoring, and ongoing fine-tuning. Present the full picture. For business leaders: if your AI vendor or IT team isn't showing you production cost projections before you approve scaling, ask for them. The pilot cost is not the production cost.
What the 19% Are Doing Differently
The organizations that are actually delivering AI ROI share some structural patterns. They're not using better models or spending more money. They've built the organizational machinery to move AI from experiment to production systematically.
Cross-functional AI governance with real teeth. The State of the CIO survey found 83% of organizations either have or are implementing cross-functional AI steering committees. But having a committee isn't the same as having effective governance. The high performers — the 19% — have committees with decision-making authority over which projects get resourced, which get killed, and what success looks like. They're not advisory bodies that produce slide decks. They have veto power.
Stage-gated funding tied to outcomes, not deliverables. Thomas Prommer, a longtime CTO, CIO, and CAIO quoted in the survey, made one recommendation that I think every executive team should adopt immediately: fund outcomes, not deliverables. His firm doesn't fund "build a model." It funds "reduce returns by 8% in this product category" — with hard checkpoints at 90, 180, and 270 days. Projects that miss two checkpoints get killed. Prommer estimates they cut roughly a third of what they start. That's healthy. That's discipline.
Compare that to the typical enterprise approach: fund a project, schedule quarterly reviews, watch the team keep building past every missed milestone because killing it feels like admitting failure. Stage-gated funding forces the conversation earlier, when course correction is still possible.
Embedded AI teams, not centralized centers of excellence. Prommer's firm replaced a centralized AI Center of Excellence with AI squads embedded inside individual business units. The CoE model created a clearinghouse that nobody owned. Embedded teams force accountability at the point where business impact happens.
This matters because the people closest to business workflows know what needs to change. A central AI team building solutions for a business unit they've never worked in will consistently miss the operational reality. As Sriram Krishnasamy, former chief digital information and transformation officer at FedEx, put it: "If someone on the data science team builds a great model that provides insights on improving manufacturing efficiency, but it's so far removed from what the shop floor supervisor does in day-to-day life, it will never be used at scale."
A named technical and business sponsor for every project. Joint accountability, co-owned outcomes. Not "the AI team built it and handed it to the business." Not "the business unit requested it and threw it over the wall to IT." Both parties own the outcome from day one.
What Good Looks Like: First Student
First Student, one of North America's leading school bus transportation companies, offers a useful case study. Their CIO Sean McCormack credits early AI success to two things: a well-defined innovation framework and an AI-specific council with representation from business leaders and the C-suite.
"We have more discipline around business cases than most companies," McCormack said. "Everything is metrics-driven and dependent on proving value. By the time we put something into production, it's been through a series of proof of concepts, there's been a deep dive on financials, and we are able to move quickly and demonstrate value."
Note what he didn't say. He didn't say they have the best models, the fastest infrastructure, or the biggest AI budget. He said they have discipline. Financial rigor. A process that ensures only proven, value-generating projects reach production.
That's the unlock. Not technology. Process.
Five Actions for Leaders Who Want to Be in the 19%
If you're in the 81%, here's what needs to change:
1. Define ROI criteria before the first line of code. Every AI project should answer: What business metric will this improve? By how much? By when? If you can't answer those questions before you start, you're not ready to start.
2. Build a production cost model before approving scale. Require your technical team to present a full year-one production cost estimate before any pilot becomes a production commitment. Include model inference, retrieval infrastructure, monitoring, security, and support. If the numbers don't pencil out, the pilot is a dead end regardless of how impressive the demo looked.
3. Replace your CoE with embedded AI squads. Centralized AI teams create distance between capability and business impact. Embed your AI talent inside the business units where value gets generated. Keep a small central team for architecture standards and shared infrastructure, but the builders should sit next to the business stakeholders.
4. Implement stage-gated funding with hard exit criteria. Fund in 90-day tranches. Define what results are required to release the next tranche. Name the checkpoints where a project gets killed if it misses targets. Kill roughly a third of what you start. That's not failure — that's intelligent capital allocation.
5. Make the metrics public inside the organization. Only 47% of enterprises have established formal AI success metrics. Of those, 40% measure operational efficiency, 34% measure employee productivity, and 30% measure cost reduction. Whatever you pick, share the scorecard broadly. When everyone can see which AI projects are delivering and which aren't, the accountability dynamics change fast.
The Mandate Is Shifting
The era of "we're running AI pilots" as a sufficient answer to board questions is ending. Investors, CFOs, and operating leaders are asking harder questions: What did we spend on AI last year? What did we get? What's the ROI?
CIO.com's survey makes the situation clear: the pipeline of pilots and rampant experimentation is giving way to a new mandate — prioritize and scale AI solutions with the greatest propensity to deliver business value and impact the bottom line.
The 19% who are already delivering aren't smarter. They're not using better technology. They built the organizational structures, governance processes, and financial discipline to take AI from experiment to enterprise. That's learnable. That's repeatable.
The question for every executive team heading into Q3 is simple: are you running AI as an experiment or as a business? The companies that answer that question honestly — and restructure accordingly — will be in the 19% by this time next year.
The rest will be running more pilots.
Sources: CIO.com 25th Annual State of the CIO Survey (2026), 662 IT leaders and 249 line-of-business respondents; Teradata agentic AI Maturity Index study, 1,000 senior technology and data leaders.
