AI Underperforms in Business: The Knowledge Gap Bottleneck

Artificial intelligence adoption across enterprise organizations routinely stumbles over performance bottlenecks because a substantial portion of the expertise required to perform daily work remains excluded from company records. According to research conducted by scholars at Harvard and MIT across 4043 full-time workers, roughly 40% of practical job knowledge resides strictly in employees’ minds rather than in documented repositories. Deprived of this operational ground truth, language model performance degrades sharply, shedding light on why numerous enterprise technology investments fail to produce tangible financial returns.

Industry benchmarking confirms this disconnect between capital expenditure and realized value. Boston Consulting Group reports that 60% of companies register minimal revenue or cost gains from artificial intelligence deployments. Concurrently, the IBM Institute for Business Value finds that merely 25% of executive-led initiatives hit expected return-on-investment thresholds. The core impediment stems not from algorithmic compute constraints, but from a persistent deficiency of structured operational context throughout enterprise software architectures.

The working paper by Zoë Cullen (Harvard), Danielle Li (MIT), and Shengwu Li (Harvard), titled “Labor as Capital: AI and the Ownership of Expertise” (HBS Working Paper 26-063, revised July 23, 2026), investigates this friction through an empirical survey paired with a formal microeconomic model. Drawn from a stratified sample of the United States full-time workforce via Prolific, the dataset captures a systematic divergence between written documentation and real-world workplace execution.

The distance between technical capacity and practical deployment

Automation forecasts often focus narrowly on theoretical task exposure. Standard methodologies evaluate whether a model can technically execute an isolated, formalized activity, such as summarizing text or generating software code. While this framing highlights baseline model capabilities, it overlooks whether the enterprise possesses the workflow data required to execute specialized, high-stakes tasks in practice.

The authors contrast survey findings against the Missing AI Index, a metric derived from real-world usage logs on Anthropic’s Claude. The resulting correlation stands at 0.58 standard deviations: occupations marked by a wider knowledge gap exhibit a far larger disparity between theoretical AI capabilities and observed workplace adoption. Systems stall not because the models lack cognitive capacity, but because they lack the localized context, edge-case protocols, and informal heuristics that govern department workflows.

How the knowledge gap is distributed across roles and processes

The volume of uncodified workplace knowledge varies significantly across economic sectors:

• Creative fields, legal practice, and computer science show the largest average gaps, spanning between 46% and 51%.
• Sales and logistics register smaller uncaptured reserves, ranging from 28% to 36%.

The most striking divergence, however, occurs within identical job titles across different organizations. The size of this information deficit correlates neither with an employee’s years of experience nor with their formal education. Instead, it tracks organizational choices: the knowledge gap widens in environments with fewer systematic process-monitoring and logging systems in place.

This tacit knowledge carries direct market valuation in emerging training data economies. Tracking postings on the talent marketplace Mercor between January and July 2026, the authors found that hiring volumes and compensation packages for domain experts concentrate in professions with the highest self-reported knowledge gaps. Artificial intelligence developers actively purchase human expertise precisely where enterprise infrastructure has failed to capture it systematically.

Worker agency in knowledge transmission

Organizational documentation depth is far from fixed. Ninety percent (90%) of surveyed employees confirm they can take deliberate steps to reduce the volume of work data captured by employers, while 96% identify concrete actions to expand it. Common methods include moving operational conversations to unlogged communication channels (71%) and producing intentionally streamlined documentation (47%).

Worker assessments indicate that these withholding choices fundamentally determine whether an external substitute can operate effectively:

• Under baseline conditions, relying solely on archived company records, a substitute performs at 45 percentage points below the incumbent.
• If an incumbent deliberately withholds knowledge to lower operational observability, the performance penalty widens to 54 percentage points.
• Under active worker collaboration with thorough procedure logging, the gap contracts to 21 percentage points.

This 35-point delta proves that an individual’s discretionary decision to supply knowledge shapes operational reproducibility as heavily as the organization’s entire cumulative record repository.

Theoretical implications of the economic model

To evaluate these dynamics, the researchers formulate a two-period microeconomic model. In period one, a worker delivers routine output while choosing how much proprietary expertise to supply into company systems. The employer collects these records to train automated systems. In period two, the firm and the worker renegotiate compensation and retention.

In this formal model, comprehensive knowledge surrender improves the employer’s outside bargaining option: the automated tool can perform specialized tasks independently or assist lower-cost personnel. Consequently, retaining procedural nuance is not irrational resistance to technological progress; it represents a calculated economic defense of future bargaining power. Absent clear incentive alignment, this dynamic induces a coordination failure that depresses data fidelity and stalls organizational productivity.

Evaluating alternative data ownership regimes

The strategic crux lies in data governance. When presented with three alternative regulatory models:

• 62.8% of workers favor individual data ownership with the explicit right to monetize their workplace data.
• 57.7% advocate binding restrictions on the secondary use of passively monitored logs.
• 46.7% endorse collective data ownership frameworks.

The model uncovers an unexpected flaw in individual data rights. Sixty-four percent (64%) of respondents acknowledge that a colleague’s work records would prove just as useful as their own in training a replacement. When an individual sells their personal logs to the employer, the resulting model captures procedures shared across the department. This unilateral sale triggers a competition externality that undercuts the bargaining leverage of peers performing similar roles, precipitating a race to the bottom in data pricing.

Collective bargaining over training data rights emerges as the only framework within the model capable of securing Pareto efficiency. By coordinating knowledge supply across entire operational units, collective governance curbs mutual wage deflation, restores incentives for accurate data capture, and guarantees that downstream efficiency dividends are shared with the professionals whose expertise trained the systems. Structured process mapping must precede any technical integration to ensure that human domain capital is captured cooperatively. To audit your internal operational workflows and assess technological readiness, connect with our integration team via our Contact us page.

Frequently Asked Questions

What is the knowledge gap identified by Harvard and MIT?

The knowledge gap measures the disparity between the practical expertise workers apply during daily execution and the formal knowledge documented in manuals, CRM platforms, and corporate databases. In this study of US full-time workers, this uncaptured component represents approximately forty percent of total job knowledge.

Why do enterprise AI implementations frequently fall short of expected returns?

Benchmark surveys by Boston Consulting Group and IBM reveal that most deployments fail to reach projected ROI targets. The primary bottleneck is the lack of rich operational context and real-world exception-handling data, both of which reside within experienced staff rather than corporate records.

Can employees restrict the operational knowledge learned by AI systems?

Yes. Ninety percent of respondents report having direct agency over the data they supply, using tactics such as off-the-record messaging or submitting minimal procedural notes. When employees choose to withhold knowledge, the performance deficit of an AI-trained replacement widens by nearly ten percentage points.

Does individual data ownership protect employee earnings?

According to the model developed by Cullen, Li, and Li, individual data monetization triggers a competition externality. Because core capabilities are shared among colleagues in similar functions, an isolated data sale allows the firm to automate the broader workflow, reducing compensation across the entire peer group.

Sources

• Harvard University, Massachusetts Institute of Technology, Zoë Cullen, Danielle Li, Shengwu Li, “Labor as Capital: AI and the Ownership of Expertise”, Harvard Business School Working Paper Series, No. 26-063, draft dated July 23, 2026. Available at Harvard Business School Faculty & Research.
• Agenda Digitale, Maurizio Carmignani, “AI, così la conoscenza dei lavoratori diventa capitale”, September 7, 2026. Available at Agenda Digitale.
• Boston Consulting Group (BCG), “The widening AI value gap: Build for the future 2025”, September 2025 (enterprise survey on revenue and cost returns in corporate AI adoption), cited in Cullen et al. (2026), footnote 3.
• IBM Institute for Business Value, “CEO study: Double down on AI while navigating enterprise hurdles”, May 2025 (global survey of 2,000 CEOs on enterprise AI ROI metrics). Available at IBM Newsroom.
• Michael Polanyi, “The Tacit Dimension”, Anchor Books, 1967 (theoretical foundation for the non-codifiable nature of human tacit knowledge, cited in the Harvard/MIT working paper).