95% of enterprise AI pilots deliver no measurable ROI, and the fix isn't more tools

Ninety-five percent of generative AI pilots at major enterprises have delivered no measurable impact on the bottom line. That is the finding from MIT's Project NANDA, reported by Fortune, and it is the number that should be pinned above every AI steering committee agenda in 2026. After two-plus years of aggressive tool rollouts, most organizations have faster individuals and unchanged organizations.

The failure is not the models. It is the architecture underneath them. According to a Smartsheet analysis published via Axios in July 2026, enterprises have systematically confused two very different things: ad hoc intelligence, which makes a single employee quicker at a task, and institutional intelligence, which makes the entire organization smarter over time. They built the first and called it transformation.

A structural debt, now overdue

For years, fragmented enterprise software stacks were held together by people. Workers knew which SharePoint folder was actually current, which Slack thread overrode the formal policy document, and which approval had already been given verbally. That human connective tissue papered over real architectural gaps. AI cannot replicate it, because AI needs connected, governed context to be useful, and most enterprise stacks do not provide that.

Smartsheet's Chief AI Officer Drew Garner framed the problem directly in the Axios piece: individual task completion has gotten easier, but working across systems and across teams has not improved. The fragmented stack was always a liability, and AI deployment has made that liability impossible to ignore.

The compounding consequence is visible at the employee level right now. Workers are already routing company work through consumer AI tools, with or without IT authorization, according to the Axios report. Those tools have no way to distinguish a current approved document from a stale draft, or to enforce what a given user is actually permitted to see. The result is a governance gap that widens every week AI adoption grows.

Two failure modes hiding inside the productivity narrative

The Smartsheet analysis names two employee archetypes that enterprise leaders are almost certainly recognizing on their own teams. The first is the productivity hoarder: someone who quietly masters AI prompting, ships work at twice the previous speed, and shares nothing. That private capability is a single point of failure. When the person takes leave or resigns, the speed advantage leaves with them, and no organizational knowledge was ever captured.

The second archetype is what Smartsheet calls the "slop cannon": an employee who mistakes volume for value and uses AI to generate enormous quantities of output with no improvement in quality or relevance. Both behaviors share a root cause. When AI is deployed without governance, without a shared standard for what is true or what constitutes good output, individual tools amplify whatever habits already exist rather than elevating the organization.

Workvivo's analysis of AI adoption patterns echoes the same diagnosis: companies have spent heavily on AI tooling but have not addressed the implementation gaps that prevent those tools from producing organizational change. The gap between AI arrival and enterprise transformation is not a technology problem. It is a process and architecture problem.

The three layers most AI deployments skip

According to the Smartsheet framework, moving past the 95% wall requires capturing three categories of knowledge that have historically lived only in employees' heads. The first is context: a live, connected map of what work exists, who owns it, and what its current state is across every system. The second is intent: a clear picture of what outcomes actually matter, and early signals when execution is drifting away from the plan. The third is judgment: the decision frameworks carried by experienced leaders that determine when to escalate, when to wait, and when the formal policy points the wrong direction.

Most AI deployments skip all three. Without them, AI cannot compound its value. Every project that ends and every team that reorganizes resets the system back to zero, because the organizational knowledge was never encoded anywhere a machine could use it.

The practical implication for procurement and IT leaders evaluating AI platforms is significant. A tool that makes a single workflow faster is easy to demo and nearly impossible to justify at enterprise scale. A platform that connects to a governed, current, permissioned view of organizational work, and that leaves an auditable trail on every action, is harder to stand up but is the only architecture that produces the 5% outcome MIT identified.

What the 5% are doing differently

The organizations that have cleared the productivity wall share a common approach, according to the Axios report. Rather than accepting the false choice between locking AI down entirely or opening it up without controls, they have connected AI to a governed operational picture where every data read is scoped to the person asking and every action generates an auditable record. That is not a feature of any single AI model. It is a property of the underlying platform.

Smartsheet is positioning its work management platform as that governed layer, arguing that AI agents need a structured, permissioned workspace to act reliably rather than a raw connection to an undifferentiated data lake. The broader market signal is clear regardless of vendor: the enterprise AI investment that matters in 2026 is not the model or the chat interface sitting on top of it. It is the data architecture, the governance standard, and the workflow connective tissue underneath.

MIT's Project NANDA finding, that only 1 in 20 enterprise AI pilots reaches measurable ROI, puts a hard number on what many operations leaders have been sensing for months. The next procurement decision for most organizations should not be which large language model to license. It should be whether the platform running underneath that model actually knows what is true, what is current, and who is allowed to ask.