Organizational capital and the value of AI

AI produces value through decisions and workflows, not through model performance in isolation.
Organizational capital includes the routines, authority, data definitions, incentives, and feedback systems that allow technical capability to affect action.
Firms that invest in models without redesigning these complements may automate existing friction or scale an invalid decision.

The Missing Input

An organization can hire capable data scientists, purchase software, acquire compute, and build accurate models without producing a meaningful operational gain.

The result is often described as an adoption problem. That phrase can be too shallow. It suggests that the technical product is complete and the rest of the organization merely refuses to use it.

In practice, the model may not correspond to the decision. Data definitions may differ across teams. No one may own the intervention. Incentives may reward a conflicting objective. The workflow may not collect the feedback required to learn whether the model helped.

These are problems of organizational capital.

Defining Organizational Capital

Organizational capital is the stock of routines and relationships that allows people, knowledge, and technology to operate as a system.

It includes:

• stable definitions and data ownership;
• decision rights;
• processes for escalating uncertainty;
• documentation and reproducibility;
• incentives aligned with the intended outcome;
• interfaces among technical and domain teams;
• mechanisms for learning from error; and
• institutional memory that survives turnover.

Unlike a server or a credential, organizational capital is difficult to count. Its absence becomes visible through delay, rework, conflicting reports, unused models, and repeated failure.

A Production Function with Organizational Capital

Let firm output be

where $H$ is human capital, $I$ is intellectual capital, $P$ is physical and technological capital, and $O$ is organizational capital.

The marginal product of AI-related physical capital is

The return to technology rises with $O$. The cross-partial derivative

captures the complementarity. A better organizational system increases the value of technical investment.

From Prediction to Decision

Suppose a model predicts outcome $Y$ from features $X$:

The organization still needs a decision rule:

The utility function $U$ contains prices, capacity, risk, fairness, customer effects, and strategic priorities. A predictive model does not determine these values.

Table 1. Organizational capabilities required for productive AI

Model output	Missing organizational decision
Probability of customer churn	Which customer receives which intervention at what cost?
Forecast of product demand	How should inventory and capacity change?
Fraud risk score	Which cases are blocked, reviewed, or allowed?
Candidate ranking	Who defines acceptable evidence and appeal?
Equipment failure probability	When does expected downtime justify maintenance?

The value is created when a prediction changes an action under a defensible objective.

The Workflow as a Causal System

AI adoption changes behavior. Once a score affects a decision, it can change the future data.

Suppose a lender approves applicants when $\widehat p(x)\geq\tau$. Repayment is observed primarily for approved applicants. The policy shapes the next training sample:

If the organization treats the model as a static forecasting object, it may overlook selective labels, feedback loops, and changes in applicant behavior.

Organizational capital is needed to define:

• who monitors the data-generating process;
• which outcomes are observed;
• when thresholds change;
• how overrides are recorded;
• who investigates distribution shift; and
• when the system should be suspended.

Model governance is part of production, not an administrative afterthought.

Automation Can Scale Weakness

Before automation, a flawed process may be slow and inconsistent. Automation can make it fast and consistent without making it correct.

Let each decision create expected value $v$ and let the system make $n$ decisions. Total expected value is

If $v>0$, scale is beneficial. If $v<0$, scale increases loss.

This simple equation explains why efficiency metrics are insufficient. An AI system that processes twice as many cases may produce twice as much value or twice as much harm.

The organization must validate the sign and distribution of $v$, not only the speed of execution.

The Role of Management Practices

Management affects whether knowledge travels and whether difficult information reaches decision makers.

Read as a diagnostic rather than a second scorecard, the framework becomes clear: Clear targets — Contribution to AI value: Connects model evaluation to an operational objective. Data ownership — Contribution to AI value: Maintains definitions and quality. Decentralized expertise — Contribution to AI value: Allows specialists to challenge assumptions. Structured review — Contribution to AI value: Detects failure before scale. Post-decision feedback — Contribution to AI value: Measures realized outcomes. Documentation — Contribution to AI value: Preserves knowledge across teams and time. Incentive alignment — Contribution to AI value: Prevents local metrics from defeating the system objective.

A firm can have capable managers and still lack these practices. Organizational capital is embodied partly in people but also in repeatable arrangements.

Building Organizational Capital

Organizational capital accumulates through investment and depreciates through turnover, system change, and neglected documentation:

where $J_t$ is deliberate investment in processes and interfaces, $F_t$ is learning from feedback, and $\lambda$ is the rate at which feedback becomes institutional memory.

Meetings alone do not necessarily raise $O$. The investment must leave a reusable change: a clarified definition, a decision log, a tested interface, a revised incentive, or a governance rule.

Feedback raises organizational capital only when it alters future behavior.

Small pilots are useful when they are treated as experiments in organizational design. A pilot should test not only whether a model can be trained, but whether data ownership is clear, users respond to the output, overrides are recorded, and outcomes return to the technical team.

The pilot creates organizational capital when its lessons become a reusable interface or rule. If a successful demonstration remains dependent on the people who improvised it, the firm has evidence of technical possibility but little institutional capacity to scale.

An AI Adoption Diagnostic

Before building a model, a firm can examine the decision system.

Expressed as a practical comparison rather than another table, the distinctions are clear: What decision will change? — Weak answer signals: “We want insights”. Who owns the action? — Weak answer signals: Several teams assume another team will act. Which outcome defines value? — Weak answer signals: Only model accuracy is specified. How will the result enter the workflow? — Weak answer signals: A dashboard without a decision rule. What feedback will be observed? — Weak answer signals: No record of intervention or outcome. Who can stop or revise the system? — Weak answer signals: Authority is unclear. Which constraints matter? — Weak answer signals: Legal, capacity, and customer effects are deferred.

A weak answer does not always mean the project should stop. It identifies the organizational work that must accompany the technical work.

Human Capital Inside Organizational Capital

Specialists need more than technical depth. They must understand the institutional interfaces through which their work creates value.

This includes the ability to:

• translate an operational concern into a modelable question;
• identify missing stakeholders;
• distinguish predictive accuracy from decision value;
• communicate uncertainty to non-specialists;
• document assumptions and revisions;
• recognize incentives that distort data; and
• design a feedback loop.

These capabilities are not “soft” alternatives to mathematics. They determine whether the mathematics belongs to the decision.

At the same time, organizations should not expect one technical employee to repair every interface. Management is responsible for constructing the system in which expertise can operate.

A Worked Example: Demand Forecasting

Suppose a retailer reduces mean absolute forecast error from 18 to 12 percent.

The technical gain is clear:

The business value remains uncertain. Inventory decisions may still be fixed months in advance. Store managers may override forecasts without recording reasons. Purchasing incentives may reward availability rather than margin. Suppliers may impose order constraints.

The decision objective may be

where $q$ is order quantity, $D$ is demand, $c_h$ is the cost of excess inventory, and $c_s$ is the cost of shortage.

Forecast improvement creates value only if it changes $q$ and if the cost parameters reflect the organization's priorities. This requires organizational capital around the model.

Organizational capital also depreciates. Experienced reviewers leave, exceptions accumulate outside formal documentation, data definitions drift, and a once-useful approval rule becomes ceremonial. AI can hide this depreciation because the system continues producing outputs at speed. Maintenance must therefore include process memory: versioned definitions, recorded overrides, incident reviews, and clear ownership of model withdrawal. These practices look administrative, but they preserve the institution’s ability to understand why a decision system behaves as it does. Without them, the firm may retain the software while losing the knowledge required to operate it safely. Investment accounting should recognize this maintenance burden. Implementation is not complete at launch; it creates a recurring obligation to preserve data meaning, reviewer competence, escalation capacity, and the link between model output and accountable action. A system that cannot fund those recurring complements is not cheaper simply because its initial model invoice is low. Its deferred cost will appear later as drift, duplicated investigation, review bottlenecks, or decisions that nobody can reconstruct. Recognizing the liability early allows management to compare automation projects on their full recurring lifecycle cost. It also prevents maintenance work from disappearing between technology, data, risk, and operating teams, each of which may otherwise assume that another function owns the problem.

This missing input now appears repeatedly in current evidence. From AI Access to Organizational Capability models the fixed investments required to integrate AI into work. Beyond Robot Relations identifies governance capabilities for managing an AI-dependent firm, and The AI Premium Is About Implementation, Not Access finds a stronger market signal from substantive use. Korea’s AI Paradox provides the macro-level warning: high adoption can coexist with weak productivity when saved time is not reorganized into output.

Conclusion

AI value is an organizational output.

Models, compute, data, and skilled people are necessary inputs. They become productive when decision rights, definitions, incentives, workflows, and feedback allow them to work together. These arrangements constitute organizational capital.

This explains why some firms appear to adopt AI rapidly but obtain little value. They have automated an interface without redesigning the system. It also explains why apparently modest models can be valuable inside strong organizations: the decision path is clear and learning is cumulative.

The managerial question is not only, “How accurate is the model?” It is, “What institution turns this model into a better decision?”

References

Erik Brynjolfsson, Lorin M. Hitt, and Shinkyu Yang, “Intangible Assets: Computers and Organizational Capital”, Brookings Papers on Economic Activity, 2002.
Nicholas Bloom and John Van Reenen, “Measuring and Explaining Management Practices Across Firms and Countries”, NBER Working Paper 12216, 2006.
Ajay Agrawal, Joshua Gans, and Avi Goldfarb, Prediction Machines, Harvard Business Review Press, updated ed., 2022.
Swiss Institute of Artificial Intelligence (2026) ‘Beyond Robot Relations: Managing, Measuring and Organizing the AI-Dependent Firm’, SIAI Working Papers, 9 August.
Swiss Institute of Artificial Intelligence (2026) ‘From AI Access to Organizational Capability: Pricing the Corporate AI Transition’, SIAI Working Papers, 9 August.
Swiss Institute of Artificial Intelligence (2026) ‘The AI Premium Is About Implementation, Not Access’, SIAI AI Memo, 11 August.
The Economy Editorial Board (2026) ‘Korea’s AI Paradox: High Adoption, Low Productivity’, The Economy Review, 29 August.