65% of Companies Still Haven't Scaled AI: The 2026 Trough | NEXAIRI

Why are most companies still stuck in AI experimentation mode in 2026?

By almost every headline metric, AI adoption is a runaway success. Investment is up, models are better, and tools are cheaper than they were two years ago.

The on-the-ground reality inside most organizations tells a different story. Stanford University's Human-Centered AI institute (HAI) published its 2025 AI Index in April 2025, reporting that 78% of organizations use AI in at least one business function—up from 55% the year before. That number looks impressive. But it masks a harder truth: using AI in one function and scaling AI across your organization to prove measurable ROI are two entirely different stages of maturity, and most companies are still at the first one.

PwC's 2026 AI Business Predictions report put it bluntly: "Success is becoming visible," but that success "has been concentrated in so few." The consulting firm found that most organizations "spread their efforts thin, placing small sporadic bets" on AI rather than committing to deep, top-down programs in a few key areas. The result is impressive adoption numbers that seldom produce transformation.

This is the 2026 adoption trough. The models are ready. Most workforces aren't, and most governance frameworks aren't either.

What do the adoption numbers actually say?

The Stanford HAI 2025 AI Index put global private AI investment at $109.1 billion in 2024—nearly 12 times China's $9.3 billion. Generative AI alone attracted $33.9 billion globally, an 18.7% increase from 2023. That's a lot of money moving into the space.

Usage is accelerating too. The 78% of organizations reporting AI use in 2024 (per McKinsey data cited in the Stanford HAI report) represents a 23-percentage-point jump in a single year. By that measure, AI adoption is one of the fastest technology transitions in enterprise history.

But the adoption surveys track deployment, not transformation. They count companies that have used an AI tool in any capacity in any one department. A team using ChatGPT for email drafts counts the same as an organization that has restructured its core operations around AI agents with measurable P&L impact. The surveys don't distinguish, and that's where the headline numbers mislead.

PwC's 2026 report found that while many companies "are also experiencing measurable ROI" from AI, "their outcomes are often modest—some efficiency gains here, some capacity growth there, and general but unmeasurable productivity boosts." These results, PwC noted, "can pay for themselves and then some. But they don't add up to transformation."

Across multiple consulting surveys and enterprise research, a consistent pattern emerges: roughly 15–20% of organizations can demonstrate scaled AI ROI that changes their competitive position. The rest are running pilots.

Why does the adoption gap persist even as models improve?

The standard narrative blames model quality: AI isn't good enough yet, so companies are waiting. That narrative is false, and it's important to say so directly.

The blockers that keep companies from scaling AI are organizational, not technical. PwC and other researchers have found the same three categories of friction at essentially every company that gets stuck.

Governance and data quality. AI systems require clean, well-labeled, accessible data. Most enterprise data doesn't meet that bar. It lives in silos, it's inconsistently formatted, and ownership is unclear. Building an AI agent that does something meaningful with that data requires data infrastructure work that predates the AI investment itself. Companies that skip this step get agents that hallucinate or produce outputs no one trusts.

Skills gaps and ownership gaps. Knowing how to evaluate an AI model is a different skill set than knowing how to deploy one in a critical workflow, and both are different from knowing how to manage teams of agents at scale. Most organizations don't have enough people who can do all three. The gap becomes worse when there's no clear owner for AI programs—when initiatives live in IT, or in a single enthusiastic team, rather than as a CEO-sponsored priority with executive accountability.

Sequencing errors. Companies that rush to deploy agentic AI before demonstrating value with simpler automation often build expensive systems that break in production. The more durable path—chatbots and classification systems before agents, departmental wins before enterprise-wide rollout—feels slower but compounds. PwC described this as "go narrow and deep," focusing on "a few key workflows where payoffs from AI can be big" before expanding.

PwC's 2026 predictions also flagged a cultural dimension: "AI feels easy to use," which creates overconfidence. Early wins from basic AI tools—a better first draft, faster code completion—mask the deeper challenges of building AI into mission-critical processes with governance and monitoring in place.

The maturity estimates above synthesize findings from PwC's 2026 AI Business Predictions and broader enterprise research patterns. Individual sectors vary—financial services and technology companies tend to skew toward the scaling and mature tiers; manufacturing and government tend to skew toward the experimenting tier.

Where does the scaling actually work?

The companies that have cracked enterprise AI at scale share a few structural patterns, not just good luck or deep pockets.

In software development, Cursor's agentic coding environment offers the clearest data point available. A November 2025 study from the University of Chicago found that, after companies adopted Cursor's agent as their default development tool, they merged 39% more pull requests. Upwork reported over a 25% increase in PR volume and over a 100% increase in average PR size—translating to roughly 50% more code shipped, according to Anton Andreev, a Principal Software Engineer at Upwork.

Cursor's enterprise page reports that 64% of Fortune 500 companies now use the platform, across more than 50,000 enterprises globally. Jensen Huang, CEO of NVIDIA, has noted publicly that all 40,000 of NVIDIA's engineers are AI-assisted through Cursor. Brian Armstrong, CEO of Coinbase, said that by February 2025, every Coinbase engineer had adopted Cursor, with "single engineers now refactoring, upgrading, or building new codebases in days instead of months."

What Cursor demonstrates is the value of embedding AI into an existing workflow where professionals already spend their time—VS Code, JetBrains IDEs—rather than adding a separate AI tool on the side. The friction of context-switching disappears. The agent knows the codebase. The output is measurable in pull requests, not in sentiment surveys.

Google Workspace's March 2026 updates follow the same logic for knowledge workers. If your organization runs on Gmail, Drive, and Docs, Gemini doesn't require a workflow change—it lives where you already work and uses data you've already created. That's why Google's internal benchmark for "Fill with Gemini" in Sheets—9x faster data entry on 100-cell tasks—is a real productivity gain for existing Workspace users, not a number you'd get starting from scratch.

What's different about startups and small teams compared to enterprises?

The adoption trough is largely an enterprise problem. Small teams and startups don't face the same headwinds, and that creates a structural window that's worth being direct about.

Large organizations deal with legacy systems, procurement cycles, compliance reviews, and risk aversion at every layer of the organization. Their data is fragmented across a decade of M&A activity. Their governance models were built for a world where software was deterministic and auditable. Agentic AI is neither.

A startup or small team building AI-native from the beginning skips most of those problems. There's no legacy data architecture to retrofit. There's no committee to approve the AI toolchain. The team moves fast by default. When the tools are good—and in 2026, the tools are genuinely good—a five-person team can run workflows that previously required a department.

PwC's 2026 report noted that "agentic workflows are spreading faster than governance models can address their unique needs." For enterprises, that's a risk. For small teams, it's a competitive window.

The practical implication: while enterprises are running six-month pilots to decide whether to expand an AI program from one department to two, small teams can instrument an entire operation with agents in weeks. The gap in operational velocity isn't permanent—but it's wide enough right now to matter.