What a 1969 Cockroach Study Predicted About McKinsey's 2026 AI ROI Gap
McKinsey's 2026 State of AI report lands on a finding that should make every C-suite uncomfortable: organizations are deploying agentic coding tools at accelerating rates, capturing substantial individual productivity gains, and still struggling to translate that into measurable organizational ROI. The workers are winning. The companies are not. McKinsey's 2026 report frames this as an adoption maturity problem. That framing is wrong, and a behavioral scientist named Robert Zajonc figured out why in 1969, using 72 cockroaches.
This is not a metaphor. Zajonc's experiment is the most direct mechanistic explanation for the 2026 enterprise AI ROI gap available, and the research predates the problem by 57 years. Understanding why that is true changes how you think about AI deployment strategy, and it changes which companies you should expect to pull ahead.
The Organizational Productivity Split McKinsey Cannot Quite Explain
The core finding in McKinsey's current State of AI data is a split. Individual workers using AI tools, particularly coding assistants and generative AI for content and research tasks, report consistent productivity improvements. Organizational-level ROI measurement tells a murkier story. Companies are pouring capital into AI infrastructure and struggling to demonstrate that the spending is converting to competitive advantage at the enterprise level.
The explanation typically offered is that organizations need more time to reconfigure workflows, train managers, and build governance frameworks. That may be partially true. But it does not explain why the gap is concentrated in specific task types rather than spread evenly across AI deployments. Companies capturing AI ROI are doing it in narrow, well-defined task categories: code review acceleration, customer service deflection, document processing. Companies struggling to capture ROI tend to be those attempting AI deployment in broader strategic or judgment-intensive workflows.
That pattern is not about adoption maturity. It is about task complexity, and the interaction between task complexity and observation pressure has been studied extensively since the 1960s.
What Zajonc Found in 1969
Robert Zajonc recruited 72 cockroaches and set them to run mazes, some of which were simple straight paths and some of which required navigating turns. The twist: some roaches ran alone, while others ran with an audience of fellow cockroaches watching from clear plastic boxes alongside the course.
The result, as Roger Dooley's Neuromarketing summary documents, was precise: observed cockroaches got faster on simple paths and slower on complex mazes. The presence of an audience helped performance on easy, well-rehearsed tasks and degraded performance on cognitively demanding ones.
This is social facilitation, and it applies to humans just as directly. Zajonc's explanation was physiological: observation elevates arousal. Elevated arousal is useful for tasks that rely on automatic, well-practiced responses because it sharpens execution of known patterns. It is counterproductive for tasks requiring novel problem-solving or careful reasoning because it creates interference in exactly the cognitive processes those tasks depend on.
The 2020 replication of Zajonc's study found that the simple-task enhancement did not hold up as cleanly in cockroaches as originally reported. But the core split between observation effects on simple versus complex tasks has been replicated extensively in human performance research across decades. Athletes perform better-practiced skills better under observation and worse-practiced skills worse. Students who know material cold perform better on tests in crowded rooms; students who are fuzzy on material perform worse. The mechanism is real even if the original cockroach data is contested.
Why This Is the Entire AI ROI Story
Map Zajonc's framework onto enterprise AI deployment and the McKinsey finding becomes inevitable.
AI coding assistants, customer support deflection tools, and document processing pipelines are organizational equivalents of the straight maze. The task has a known structure. The output is measurable. The performance criteria are unambiguous. When these tools are deployed with monitoring, accountability metrics, and team visibility, the observation effect amplifies the AI productivity gain rather than suppressing it. Developers who know a manager can see their completion rate on AI-assisted tickets will use the tool more consistently. Customer service teams with visible AI-deflection dashboards will lean into handoffs more aggressively. The accountability and the simple-task structure compound.
Agentic AI applied to strategic planning, competitive analysis, product positioning, or novel problem-solving operates in complex-maze territory. These are judgment-intensive tasks with ambiguous success criteria, high stakes, and significant managerial observation. When an organization deploys an AI strategy tool with senior leadership watching the output, they have recreated the exact conditions Zajonc documented as performance-degrading. The system is under observation pressure while navigating complexity. In humans, this produces defensive conservatism, anchoring to familiar outputs, and avoidance of the novel synthesis the tool is actually capable of.
The organizational ROI gap exists because most companies are deploying AI uniformly across both task categories without recognizing that the observation conditions they apply are interacting differently with each one.
The Base Rate Problem That Compounds Everything
Nick Maggiulli at Of Dollars and Data recently published a piece about why he was wrong to be bearish on US stocks that is instructive here for a different reason. His core error was pattern-matching surface signals (SPAC filings, elevated price-to-sales ratios that turned out to be based on corrupted data) while ignoring the base rate: US stocks rise in roughly 7 of every 10 years. He was fitting his prediction to recent noise and discounting the most probable outcome.
Enterprise AI forecasting has the same structural problem. Companies see signals that look like ROI barriers (integration complexity, change management friction, governance gaps) and build deployment strategies around those signals, while ignoring the base rate question: over a large deployment of AI tools across a large employee population, what fraction of use cases are simple-task-equivalent versus complex-task-equivalent? That number is not close to 50/50. The realistic distribution skews heavily toward simple-task-equivalent use cases, which means the base rate prediction for AI ROI should be more optimistic than most organizational strategies reflect. The companies that are capturing ROI are, knowingly or not, deploying against the base rate rather than against the noisy surface signals.
The ones struggling are building governance frameworks and maturity models designed to solve a complexity problem that is actually a task-classification problem. They have diagnosed the wrong thing.
IBM and the Organizational Proof Case
Branding Strategy Insider's analysis of IBM's 2026 AI positioning offers a useful case of what the complex-task failure mode looks like at the organizational strategy level. IBM's leadership has framed AI as a "temporary disruption" to its core business and positioned the company's mainframe infrastructure as ultimately AI-enhanced rather than AI-threatened. The argument is substantively defensible. The strategic behavior it produces is not.
When you frame a structural shift as temporary disruption, you are performing Zajonc's complex maze under observation pressure. Every strategic decision IBM makes about AI investment is watched by analysts, boards, and customers who have already priced in the company's historical positioning. The observation pressure creates incentives toward outputs that confirm prior frameworks rather than outputs that require genuinely novel synthesis. IBM is not simply being slow to adopt AI; it is in an organizational state where the conditions for capturing AI's highest-value use cases are specifically counterproductive.
The three requirements Branding Strategy Insider identifies for IBM's future (enduring growth strategy, thought leadership through innovation, positive momentum rather than legacy coasting) are all complex-task-equivalent goals. They require sustained novel synthesis in an environment of maximal observation pressure. Zajonc would not be optimistic about the outcome under current conditions.
The contrast with companies capturing AI ROI is not primarily about technical infrastructure or product quality. It is about whether the organizational context surrounding AI deployment is creating simple-task-equivalent conditions (clear criteria, measured outcomes, low-ambiguity execution) or complex-task-equivalent conditions (strategic stakes, leadership scrutiny, ambiguous success metrics).
What the Zajonc Framework Actually Recommends
The practical implication is not that organizations should reduce AI accountability or stop measuring AI performance. It is that task classification should precede deployment strategy, and the observation conditions applied to each task category should be designed around the known psychology rather than against it.
For simple-task-equivalent AI deployments, visibility and accountability are features. Build dashboards. Set measurable targets. Make the productivity gains observable to managers and teams. The observation amplifies the gains precisely because the task structure is known.
For complex-task-equivalent AI deployments, the approach inverts. We have written here before about the trust gap that emerges when AI agents operate under excessive scrutiny in judgment-intensive domains. The solution is not less measurement but different measurement: outcome-based evaluation on longer time horizons, with the observation conditions reduced during the synthesis phase and applied at the evaluation phase. Protect the complex-task work from performance pressure during execution. Evaluate it after.
Most organizations are doing neither of these things deliberately. They are applying uniform AI governance frameworks, designed for compliance and risk management, to both task categories simultaneously. The result is that simple-task deployments are under-exploited relative to their potential (governance overhead slows adoption) and complex-task deployments are structurally compromised (observation pressure suppresses the novel synthesis that would justify the investment).
The caution around agentic AI adoption that has emerged across industries is, in part, a rational response to observing complex-task AI deployments fail under exactly these conditions. The caution is justified, but the diagnosis is usually wrong. The problem is not agentic AI. The problem is agentic AI applied to complex tasks in high-observation organizational environments.
The Original Claim the Data Supports
Here is the inference no individual source in this article makes: the McKinsey 2026 AI ROI gap is mechanically predictable from Zajonc's 1969 task-complexity research, and it will persist for exactly as long as organizations continue deploying AI without classifying task complexity first. Companies that close the gap are not closing it through better models, better governance, or better change management. They are, whether they realize it or not, deploying against the social facilitation gradient: high observation for simple tasks, protected environments for complex ones.
The implication for investment and competitive positioning is direct. Look for companies where the AI use cases showing ROI are narrow, well-defined, and heavily instrumented. That is the social facilitation effect working as designed. Be skeptical of companies announcing broad AI strategy integration with senior leadership sponsorship and a lot of measurement infrastructure. That is the organizational equivalent of Zajonc's complex maze with an audience. The cockroaches slow down.
STI Research publishes independent analysis on decision intelligence, brand strategy, and behavioral economics. If you want to see how these frameworks apply to specific strategic decisions, our research library is here.