Only 19% of enterprises can demonstrate that their AI initiatives have met or exceeded their ROI goals. That's 81% of organizations pouring budget into AI and struggling to show the board a single clean number. The Foundry 2026 State of the CIO study surveyed 662 IT leaders and landed on that finding just this spring. Meanwhile, Anthropic reported that Claude now writes more than 80% of code merged at one of the most sophisticated AI companies on earth. Both numbers are true. The distance between them is the most important thing an IT leader needs to understand about AI right now.
That distance is not a contradiction. It's a lesson — and most enterprises are reading it wrong.
The Model Is Not the Problem
When most executives see the 81% statistic, their first instinct is to blame the technology. The models aren't ready. The vendors overpromised. The hype outran reality. It's a comfortable explanation because it outsources the problem. If the models are the issue, you wait for a better model.
But look at the Foundry data more carefully. The top three barriers named by IT leaders were: lack of in-house expertise (40%), ill-defined ROI metrics (32%), and murky corporate AI strategy (31%). Not one of those is "the model isn't good enough." Not one respondent in a survey of 662 senior IT leaders said the underlying technology was the primary obstacle.
So if the models aren't the problem, what is?
The answer is hiding in plain sight inside the Anthropic data. Claude's coding gains aren't happening because coding models are fundamentally superior to the same models answering support tickets or reviewing contracts. The underlying LLMs are identical. What's different is everything around the model: the infrastructure that software development built over decades.
Code has governance baked into branch protection and code review. It has observability through version control and CI/CD pipelines. It has evaluation through automated test suites. It has persistent context through commit history. Developers didn't build any of that for AI. They built it for themselves. But when AI agents arrived, that infrastructure was already there — the exact scaffolding needed to make an agent's output trustworthy and measurable.
That's the substrate. And most enterprise workflows don't have one.
What 81% Failure Actually Looks Like
The Foundry data shows something that should concern every CIO: enterprises are not failing at AI because they're disorganized. They're failing despite being very organized.
Eighty-three percent of surveyed IT leaders have either stood up cross-functional AI steering committees or are in the process of building them. Just over half have some form of AI approval process in place. Forty-seven percent have formal AI success metrics defined, with another third actively building them. On paper, this looks like serious institutional readiness.
Yet the ROI isn't appearing. The governance machinery is running. The organizational chart has AI titles in it. The committees are meeting. And four out of five enterprises still cannot point to AI delivering on its promised value.
Here's what that tells you: the leakage isn't at the governance layer. Steering committees and KPI dashboards govern the organizational structure. But value — and waste — happen at the task level. Inside the workflow. At the moment an AI agent produces an output, is that output measurable? Is there a signal, anywhere in the system, for whether the agent got it right?
In software development, that signal is automated. A failing test is a failing test. The CI/CD pipeline either goes green or it doesn't. There's no ambiguity, no committee required, no executive judgment call. The workflow itself closes the feedback loop.
In most enterprise workflows being targeted for AI today — procurement, contract review, customer communications, financial analysis — that feedback loop doesn't exist at the task level. AI output goes into a human review process. The human either catches errors or doesn't. If an error gets through, it's often untraceable back to the model. There's no persistent signal the system can learn from. There's no automated test that proves value was delivered.
The TIAA Reality Check
TIAA is about as prepared as an enterprise can be. The firm is three years into its AI journey, runs both generative and agentic use cases across fraud detection and call centers, and has 85% of its workforce on an internal AI platform called TIAA Gate. It has the full governance stack that most CIOs are still assembling.
And yet its chief operating, information, and digital officer told CIO.com this spring that the ROI gap is still real. "You need to understand the full cost of operations," he said, "the efficiencies of running tokens or how you're handling traffic or RAG." Three years in, with near-universal platform adoption, TIAA is still working through what it actually costs to run AI at scale and what the measurable returns look like at the task level.
This is not a criticism of TIAA. They're further ahead than almost anyone. The point is that organizational readiness and platform adoption are necessary but not sufficient. The thing that closes the gap is workflow observability — the ability to see, at the level of the actual task, whether the AI output was right and what it cost to produce.
The Expertise Gap Is a Substrate Signal
The sectors with the largest in-house expertise gaps are healthcare (52%), retail (51%), and manufacturing (49%). That's not a coincidence. Those are also the sectors whose core work looks least like a software development lifecycle.
Healthcare workflows involve clinical judgment, regulatory documentation, and patient safety decisions — none of which have structured observability baked in. Retail operations span supply chain, merchandising, and customer interactions that are measured episodically, not continuously. Manufacturing floor processes produce physical outcomes that are hard to reduce to a digital signal.
The expertise gap is real and it matters — you cannot build workflow substrate without people who know how. But the gap is deepest where the substrate is hardest to build. That's consistent with substrate being the underlying variable, not just talent availability.
Compare that to financial services, where the NVIDIA 2026 State of AI report shows the strongest adoption outcomes. Financial workflows are measurable by design. A transaction either clears or doesn't. A fraud flag either matches a known pattern or doesn't. A portfolio rebalancing either hits its targets or doesn't. The financial sector built structured, observable workflows long before AI arrived. When AI agents showed up, the substrate was there.
What Sequencing Wrong Costs You
The prescription most AI strategies are built on goes something like this: identify where AI can create the most value, prioritize those use cases, and invest there first. It sounds rational. It's how most capital allocation decisions get made.
The problem is that value potential and substrate readiness are almost never correlated. The highest-value use cases — customer-facing personalization, complex contract negotiation, clinical decision support — are also the ones with the least structured workflows and the widest trust gaps. The cost of a wrong answer is asymmetric. A bad internal draft gets fixed before anyone sees it. A bad customer answer is the whole relationship.
What actually happens when you sequence by value: you end up with a sprawling portfolio of pilots, each tackling a high-value problem, none of them producing the structured feedback loop needed to demonstrate ROI. The TIAA lesson applies even at the portfolio level. More pilots does not mean more measurement. Andrea Ballinger, CIO at Rensselaer Polytechnic Institute, described this trap precisely in the Foundry report: "We are saying yes to everyone without stepping back and focusing on the business cases that show real value."
The alternative is counterintuitive: sequence by workflow readiness, not by value potential.
Start with use cases where the workflow already has — or can quickly be given — structured inputs, automated evaluation, and a clear signal for whether the output was right. Not because these use cases are more important, but because they're the ones where AI can actually demonstrate ROI instead of just promising it.
The Practical Framework
For CIOs and AI strategy leads working through their portfolio, this translates into three concrete questions for every AI initiative:
1. Where is the feedback loop?
Before deploying AI into any workflow, identify the point in that workflow where you can measure whether the AI output was correct. If that point doesn't exist or requires human judgment with no structured capture, the AI's performance is invisible to the organization. Build the measurement layer first. This is not a technology problem — it's a process design problem.
2. What is the substrate?
Audit the workflow for observability: structured inputs, version control equivalent, automated evaluation, persistent context. Software development has all four. Most other enterprise workflows have none. Each one you can add before deployment increases the probability of measurable ROI. Prioritize substrate investments alongside model investments.
3. Are you piloting or deploying?
A pilot without a defined feedback loop is not a learning experiment. It's an investment with no measurement plan. If you cannot describe how you will know whether the pilot succeeded — with a specific, automated, task-level signal — it should not proceed as a pilot. It should proceed as a substrate-building project first.
What the 19% Are Doing Differently
The enterprises in the top quintile — the ones actually demonstrating AI ROI — share a pattern that emerges from the Foundry and NVIDIA data. They're not smarter about AI. They're smarter about workflows.
NVIDIA's 2026 State of AI report found that 76% of large enterprises (over 1,000 employees) are actively using AI, versus 64% overall. The gap is not explained by bigger models or better vendors. Large enterprises have more structured internal processes, more established data architectures, and more existing instrumentation. They came to AI with substrate already in place.
The 19% who are meeting their ROI targets are deploying AI where the workflow can already tell them whether the AI is doing its job. The 81% are deploying AI where the workflow can't.
The Strategic Takeaway
Enterprise AI ROI is not a technology problem. It is a workflow readiness problem disguised as a technology problem.
The models are capable. The governance structures are being built. The executive commitment is there — average enterprise AI spend hit approximately $7 million in 2025 and is projected to jump 65% to $11.6 million in 2026. The investment is real.
What's missing is the substrate layer between the model and the business outcome. The feedback loops that make AI performance visible. The evaluation infrastructure that tells you, at the task level, whether the agent got it right.
Software engineering built that infrastructure over decades without knowing it would one day be the critical success factor for AI deployment. Every other enterprise function now has to build it on purpose.
The enterprises that close the ROI gap fastest won't be the ones with the biggest AI budgets or the most advanced models. They'll be the ones that treat workflow observability as a prerequisite — not an afterthought — for every AI initiative they launch.
The Foundry data is telling you something important: you're not failing because AI doesn't work. You're failing because you're deploying AI into workflows that can't tell you whether it's working.
Fix the workflow. The ROI follows.
Sources: Foundry 2026 State of the CIO (662 IT leaders surveyed); Anthropic Recursive Self-Improvement research; NVIDIA State of AI 2026 (3,200+ respondents across financial services, retail, healthcare, telecom, manufacturing); NVIDIA State of AI in Financial Services 2026.
