AI ROI Is an Operating Model Problem

Microsoft’s 2026 Work Trend Index found that organizational factors account for roughly two-thirds of the variance in AI performance outcomes. Your AI ROI problem is an operating model problem.

The Belief That Won’t Die

Every conversation about AI underperformance follows the same script. The technology isn’t advanced enough. Teams don’t have the right skills. Vendors oversold the platform. These three explanations absorb nearly all the budget, attention, and executive frustration aimed at closing the AI ROI gap.

They also account for roughly a third of the actual problem.

What 20,000 AI Users Revealed

Microsoft’s 2026 Work Trend Index surveyed 20,000 AI users across 10 countries and measured which factors are associated with AI performance outcomes (1. Microsoft, 2026). Organizational factors, including how decisions get made, how workflows are designed, and how goals get communicated, account for approximately 67% of the variance. Individual factors, including skill, training, and personal adoption habits, account for 32%.

The study reports these as statistical associations, not proven causal relationships. But the magnitude is hard to dismiss. The organizational environment around the AI user matters roughly twice as much as the user’s own capability.

The same study found that 31% of AI users are misaligned, actively using AI tools in ways disconnected from their organization’s goals (1. Microsoft, 2026). These aren’t reluctant adopters or undertrained staff. They’re people using the technology without organizational direction on what to use it for. The implementation worked. The organization around it never changed to support it.

The Constraint You’re Not Addressing

I’ve argued before that fixing AI performance starts with operationalizing marketing itself. Microsoft’s data quantifies why. And the pattern holds beyond their study. McKinsey’s 2025 State of AI survey found organizations extracting value are three times more likely to have redesigned workflows and three times more likely to have senior leaders demonstrating direct ownership of AI initiatives (3. Singla et al., 2025). Fewer than one in ten of McKinsey’s respondents qualified as high performers (3. Singla et al., 2025).

The technology works. And the organization extracts a fraction of the value because its operating model was designed for a world where humans performed every step. Decision rights, data governance, measurement contracts, coordination protocols: all built for human-only execution.

Kathy Katz, VP of Revenue Operations at Clari, identified the mechanism: “The measurement infrastructure needed to prove AI ROI is the same infrastructure needed to run a clean GTM operation” (2. Omoregie, 2026). She’s describing the compounding nature of organizational capability. Every operating model improvement you make for AI generates returns that extend beyond AI.

Clean data governance supports better attribution. Clear decision rights reduce coordination overhead across every initiative, not just AI-driven ones. The capability you build compounds across everything the organization touches. Platforms depreciate the day you sign the contract. The organizational capacity you build around them doesn’t.

Andrea Hornaday, VP of Marketing Operations at Bain & Company, diagnosed it directly: the AI performance gap “is a systems problem” (2. Omoregie, 2026). How the organization coordinates around its technology determines whether AI performs or stalls.

The distinction between integration and orchestration explains why. Most organizations have integrated their AI tools. The systems exchange data, the APIs connect, the dashboards consolidate. Far fewer have orchestrated them, designing the operating model so that AI-generated outputs flow through decision frameworks, accountability structures, and measurement systems that make the outputs actionable. Integration is a technology achievement. Orchestration is an operating model achievement. The ROI lives in the second one.

The CMO’s First Move

The temptation is to respond with an enterprise-wide AI strategy document. Resist it. Operating model redesign at scale is a multi-year transformation, and the data is clear that most organizations aren’t ready for that scope.

Start with one workflow. Pick the workflow where AI coordination currently depends on human handoffs between disconnected tools. The one where someone manually moves an AI-generated output from one system to another, applies judgment, routes it forward, and tracks the outcome in a spreadsheet. That workflow is where your operating model constraint is most visible and most fixable.

Redesign three things for that single workflow:

Decision rights. Who approves what the AI produces? Who overrides it? Who’s accountable when the output is wrong? If the answer to any of these is “it depends” or “whoever’s available,” the operating model is the bottleneck.

Data access. Does the AI agent have access to the context it needs, or does a human assemble information from three systems before the agent can do its work? Every manual data assembly step is an operating model failure masquerading as a technology limitation.

Measurement. Define the business outcome this workflow exists to produce before the AI touches it. Adoption rate and time saved measure technology activity, not business value. Build the measurement infrastructure first. Per Katz’s insight, you get both the proof mechanism and the operational backbone (2. Omoregie, 2026).

One workflow. Three redesigns. The capability you build transfers to the next workflow, and the one after that. The organizational muscle compounds in ways a platform purchase never will.