By Here is a number that should make every CIO uncomfortable: 88% of organizations now use AI in at least one business function. 80% of employees say it improved their personal productivity. Yet only 37% of organizations report any positive EBIT impact from AI — and that number has not moved in two years. The genuine high performers, organizations attributing at least 5% of EBIT to AI, represent exactly 6% of respondents in McKinsey’s 2026 State of AI survey across 1,719 organizations and 97 countries.
The productivity gains are real. The financial returns are not. Something significant is being lost in translation.
Three Layers, Three Leaks
The ROI gap is not a single problem — it has a structure. Productivity gains exist at the individual level. They leak before reaching the team level. What makes it to the team level leaks again before reaching the organization’s profit-and-loss statement. Each transition drops value.
The reason is straightforward: most organizations layered AI onto existing workflows rather than redesigning those workflows for AI. They sped up broken processes instead of eliminating them. The McKinsey data is blunt about this: nearly three-quarters of high performers fundamentally redesigned their workflows because of AI. Only about one-quarter of other organizations did the same. This single decision separates the 6% from everyone else more clearly than any other variable in the dataset.
The returns on each approach are not close. Analysis from AIJourn puts it starkly: layering AI onto existing workflows delivers roughly 30% ROI. Redesigning those workflows for native AI execution delivers 30x. The difference is not incremental — it is architectural.
Design Failure, Not AI Failure
Writer’s 2026 Enterprise AI Adoption Survey, covering 1,200 C-suite executives and 1,200 employees through independent research firm Workplace Intelligence, confirms the pattern from a different angle. 54% of C-suite executives admit that adopting AI is creating significant internal conflict. 78% report tension between IT and business lines. 48.6% cite legacy system integration as a blocker.
These are not complaints about the models. They are complaints about the organizational plumbing around the models.
Stanford’s Digital Economy Lab found that escalation-based agentic systems delivered 90% productivity gains in IT operations and 71% in customer support — but specifically in organizations that redesigned workflows before deploying agents. The same agents, operating inside unchanged legacy workflows, produced marginal results. The tool is not the variable. The structure surrounding the tool is.
This matters for how the problem gets framed. The ROI gap is not evidence that AI does not work. It is evidence that most organizations skipped the hard part and expected the easy part to be sufficient.
The Inference Cost Ceiling
There is a second constraint that the McKinsey data flags for the first time in 2026: AI operating costs are now actively constraining AI use at approximately one in five organizations. This is new. A year ago it did not appear in the dataset.
The mechanism is counterintuitive. Per-token costs dropped roughly 98% since early 2024. Enterprise AI bills are still rising. The reason: agentic workflows use between 5 and 30 times more tokens per task than standard chatbot queries. A single agentic interaction triggers 10 to 20 model calls. The cost per agentic customer service interaction went from $0.04 in 2023 to $1.20 in 2026 — a 30x increase driven entirely by architectural complexity, not model pricing.
The Uber case is instructive. Claude Code adoption across Uber’s 5,000-engineer organization went from 32% to 84% between December 2025 and March 2026. By April, the entire annual AI budget had been consumed. Monthly per-engineer API costs ran $500–$2,000. Inference now represents over 85% of total enterprise AI budgets in organizations that have scaled agentic workflows. Scaling without cost controls is not a growth strategy — it is a budget crisis scheduled in advance.
How Tool Sprawl Makes Everything Worse
Fragmented tool stacks amplify the ROI gap. Most large organizations deployed five to seven AI point solutions per department between 2024 and 2025. At that level of fragmentation, agentic workflows face data silos that prevent cross-process operation, integration failure points that multiply with each added tool, and compounding token overhead as agents re-establish context they cannot share between systems.
The market has started correcting this. Developer tool counts dropped from over 20 specialized tools in 2024 to roughly 8 consolidated tools in 2026. Enterprise vendor counts fell from 12–18 AI tools to 3–6 platforms. Tool consolidation from 20 to 5 is now the dominant enterprise IT procurement strategy. 44% of organizations have already cut their stacks; 72% report improved outcomes after doing so.
What the 6% Actually Did
The high performers are not using better models or spending more money. They made different decisions at the architectural and organizational level before most companies even started deploying AI broadly.
- Redesigned workflows for native AI execution rather than overlaying AI on existing processes
- Defined revenue and innovation goals alongside efficiency goals — not just cost reduction
- Built data foundations — clean, accessible, connected — before deploying agents that depend on that data
- Measured AI impact with defined processes and visible senior-leadership accountability
These are not technical decisions. They are organizational ones.
What This Means for Developers
If you are in an organization inside the 94%, you are not there because the technology failed. You are there because somewhere upstream, a decision was made to skip the redesign work and ship the tools. The Stack Overflow 2026 survey found that AI favorability among developers dropped from 77% in 2023 to 60% today. That declining sentiment tracks with the experience of using powerful tools inside processes that were never designed to leverage them.
The question worth asking in your next planning conversation is not which AI tool to evaluate next. It is which process your organization is willing to actually redesign. That is where the returns are. The McKinsey data is clear: the 6% did not discover a secret tool. They did the unglamorous structural work first. The other 94% are still waiting for the tool that does it for them.
