IT Business Value After 30 Years: We Are Still Asking the Wrong Question

I opened the binder to the productivity paradox section last week and found a coffee stain from February, which means I had not touched it since oral prep. Rereading Brynjolfsson (1993) after spending months with Melville et al. (2004), Yeh et al. (2025), and Baird and Maruping (2021) made the thirty-year arc visible in a way it was not before. The question has changed. The field has not.

Brynjolfsson named the productivity paradox in 1993. Organizations poured money into IT through the 1970s and 1980s, and macroeconomic productivity statistics barely moved. He gave four explanations: mismeasurement, because the statistics could not capture quality or variety improvements; time lags, because IT benefits require organizational restructuring that takes years; redistribution, because gains get passed to customers as lower prices rather than captured as higher profits; and mismanagement, because firms invested without aligning processes, skills, or governance. I wrote before about why the paradox never really went away, and I still think mismanagement is the explanation that keeps recycling. The point here is different. Brynjolfsson's real legacy was not the four explanations. It was the question he embedded into the field: does IT create value?

That question dominated the 1990s and early 2000s. Researchers ran regressions, built production functions, and argued over whether the coefficient on IT capital was positive or zero. The implicit frame was binary. Either IT produces returns, or it does not. Brynjolfsson and Hitt (1996) moved the debate to the firm level and showed that IT capital has a higher marginal product than non-IT capital, which resolved the paradox empirically for a while. But the binary frame survived. The field kept asking whether IT pays off, as if the answer could be a single yes or no.

Melville, Kraemer, and Gurbaxani (2004) broke that frame. Their integrative BVIT model refused the direct shortcut from IT spending to organizational performance. IT resources interact with complementary organizational resources, which improve business process performance, which then drives organizational performance, all moderated by the competitive environment. I have used this model repeatedly because it is the single best answer to oversimplified claims about IT returns. The Melville et al. question was not whether IT creates value. It was under what conditions. The answer turned out to be about capability, not spending. The unit of analysis shifted from the technology to the organization surrounding it.

The shift from "does IT create value?" to "under what conditions?" was real progress. But it left a gap. The BVIT model is elegant at the organizational level and silent at the individual level. It tells you that complementary resources matter, but it does not tell you what a data analyst actually does with the system at nine in the morning. The mechanism between IT investment and process improvement was still a black box labeled "organizational capabilities." You could fill that box with anything: culture, governance, training, leadership. The model was right about the chain, but it was thin on the specific link between a person and a system.

Yeh, Eden, Fielt, and Syed (2025) saw the gap clearly. Their BDA-Use-Value framework is an explicit attempt to bridge the business value of IT literature with the system use literature, which had been running on parallel tracks for decades. BVIT puts agency with senior managers who allocate resources and build complementarities. System use puts agency with individual users who decide whether and how to employ specific features. Yeh et al. argue that both levels matter and that the field needs a multilevel picture: system capabilities, individual use, organizational capabilities, complementary resources, learning loops, and environmental context. They built their framework for big data analytics, but the theoretical move is general. Value creation requires that the right people at the right level are doing the right things with the system, and neither the managerial layer nor the individual layer is enough by itself.

I think Yeh et al. made the most important advance in IT business value research since Melville et al. They also revealed the limit of the advance. Their framework still treats "use" as the central construct between the system and value. Use is defined as the user's employment of one or more features of a system to perform a task. That definition holds for databases, dashboards, and ERP screens. It does not hold for agentic systems that appraise a task, distribute subtasks between themselves and a human, and coordinate interdependencies over time. It does not hold for an AI agent that writes code on behalf of a developer, or a surgical tool that decides whether a tumor is malignant. The closer the system gets to acting autonomously, the more the human-system relationship looks like a transfer of rights and responsibilities, and the less it looks like feature employment.

Baird and Maruping (2021) gave the field the construct that was waiting at the end of this arc. They argued that for agentic IS artifacts, delegation replaces use as the foundational lens. Delegation is the transfer of rights and responsibilities for task execution and outcomes from one agent to another. The three mechanisms are appraisal, distribution, and coordination. I have written extensively about why this matters for agentic AI, but the implication for IT business value is broader. Delegation is not a special case for chatbots and copilots. It is the mechanism that was always operating beneath the surface of use, and agentic AI simply made it visible enough to measure.

When an organization buys an analytics platform and gets nothing, the standard explanation from Brynjolfsson's mismanagement category is that the firm failed to align processes, skills, or governance. Melville et al. would say the complementary organizational resources were missing. Yeh et al. would say the multilevel use was not coordinated. All three are correct but incomplete. The deeper explanation is that the organization never built the capacity to delegate. Managers did not appraise which analytical tasks the system could handle. The firm did not distribute decision rights between the system and the analysts. Nobody coordinated the handoffs. The dashboards sat there because nobody had transferred the right responsibilities to them. As I argued in my post on why the productivity paradox disappears when you measure delegation, the paradox does not disappear because IT creates more value. It disappears because we finally count the value-creating mechanism instead of its proxy.

The thirty-year arc from Brynjolfsson through Melville et al. to Yeh et al. traces a gradual decentring of the technology itself. The question went from "does IT pay off?" to "what conditions make IT pay off?" to "what do people actually do with IT and at what levels?" The next step, which Baird and Maruping opened, is to ask what the organization delegates, to whom, and whether the coordination routines around that delegation are functional. Delegation theory gives us the vocabulary to ask that question without collapsing everything into a usage metric or a spending figure.

I do not think the IS field has fully absorbed what this means. Gartner forecasts global IT spending at six trillion dollars in 2026. McKinsey finds that only seven percent of organizations have fully scaled AI. The spending is measurable. The delegation is not. We are still running regressions of IT capital on firm performance, still producing variance explained numbers that pretend the black box is shallow, still treating the productivity paradox as a puzzle about investment when it is really a puzzle about rights and responsibilities. The question that should have replaced Brynjolfsson's paradox was never "does IT create value?" It should have been "how does the organization measure what it values, and who holds the rights and responsibilities for that measurement?" Delegation theory does not answer that. It gives us the vocabulary to ask it without flattening every human-system interaction into a login count.