Matthias Holweg
American Standard Companies Professor of Operations Management,
Saïd Business School, University of Oxford
Organisations globally are making unprecedented capital investments in artificial intelligence (AI), yet corporate leadership is increasingly asking a fundamental question: where is the return on investment?
Despite significant investments, realised productivity gains consistently remain elusive. The explanation for this stagnation lies in a historical lesson that is routinely overlooked: we are treating AI as an isolated, plug-and-play solution, forgetting that technology only generates economic value when effectively integrated into an optimised business process.
The modern productivity paradox
To diagnose why AI so far has failed to transform bottom lines, it is worthwhile to revisit the early deployment of personal computing.
During that era, economists identified a “productivity paradox,” noting that massive information technology investments yielded negligible macroeconomic productivity improvements. This occurred because businesses merely digitised existing manual workflows rather than re-engineering how work was performed.
Corporate correspondence provides a case in point. Historically, drafting a physical letter required significant time and administrative effort; consequently, organisations communicated selectively, ensuring each document possessed high utility.
The advent of email made communication instantaneous and virtually costless. Instead of capturing efficiency gains, however, inboxes were promptly flooded with a high volume of low-value electronic messages.
AI is now driving a structurally identical, yet far more severe, operational loop. Generative AI enables the mass production of content, resulting in the widespread “slopification” of business processes.
As employees offload critical thinking to algorithms, they generate polished but low-utility “workslop”. This forces downstream recipients to deploy intensive human labour just to verify and validate the output – an extra burden that routinely negates any initial productivity gains.
we must look beyond what is technologically feasible and analyse AI models
within the broader intra- and interorganisational workflows in which they reside
Mechanics of AI value creation
The operational disconnect stems from a fundamental misunderstanding of the mechanism through which AI generates value. Stripped of all hyperbole, AI functions strictly as a predictive engine.
The core financial and operational value of the technology is unlocked when historical data is used to identify latent patterns, which are then translated into a prediction. This prediction is intended to inform a concrete corporate decision, which subsequently allows management to execute a more efficient allocation of organisational resources.
Reversing this logic reveals a critical boundary condition. If a business process does not conclude with a concrete decision capable of being informed by historical data, AI cannot yield value — a reality that directly surfaces two fundamental limitations for leadership: process variation and structural innovation.
Modern AI is built upon machine learning, which is a subfield of statistics. Hence, its performance is bound by core statistical principles: more data and lower process variation directly impact the quality of the output.
As enterprises transition toward agentic AI – deploying autonomous software agents to execute multi-step workflows – this statistical dependency becomes a tangible operational challenge. Processes that are standardised and highly repeatable can be successfully “learnt” and, in turn, augmented or even automated. Conversely, inherent process variability and structural exceptions will severely limit the productivity gains agentic AI can offer.
And for businesses relying on novel strategy, market intuition or operating in highly volatile contexts devoid of historical precedents, AI-generated predictions have little to offer. Because these systems function on probabilistic pattern matching, they are structurally biased toward what can be termed “backward similarity”: evaluating new ideas based entirely on what has already succeeded in historical datasets.
Consequently, AI will struggle to recognise a genuine strategic breakthrough or a novel operational paradigm; it merely replicates and enforces past conventions.
Beyond the hype of shiny AI models
To capture true economic value, we must look beyond what is technologically feasible and analyse AI models within the broader intra- and interorganisational workflows in which they reside. The real productivity return of AI is not measured by the speed of its algorithms, but by the net efficiency gains it brings to business processes.
Until we stop marvelling at the novelty of new AI tools and start systematically improving the workflows they support, AI will remain a compounding capital expense rather than a structural value driver.
