While artificial intelligence (AI) adoption has spread rapidly, meaningful productivity gains remain elusive because organizations have conflated easy-to-deploy horizontal AI tools with the domain-specific vertical AI systems that actually drive results.
Predictions of equally utopian or dystopian paradigm shifts have followed the proliferation of artificial intelligence (AI) technologies almost from the start. Even as many commentators warn of imminent, permanent, and catastrophic job displacement, a commensurate number of counterparts predict trillion-dollar efficiency gains on the back of rapid automation and unprecedented productivity growth.
To date neither eventuality has been borne out. Across most industries, reality has been far more measured. The adoption of AI technologies is happening, but not at revolutionary speed. Productivity improvements are emerging, but not at a transformative scale. Most organizations remain on a continuum somewhere between pilot, proof of concept, and early operationalization.
This gap between expectation and reality is not, however, evidence that AI has been overhyped. Rather, it reflects a pattern researchers have documented for decades: major technologies rarely reshape productivity the moment they appear. Instead, time and again it has been demonstrated that emergent technologies require complementary innovations, redesigned processes, new skills, system-level integration, and organizational change before their economic benefits become widely apparent. Despite its contemporary novelty, AI seems bound for a similar trajectory, if new contributing factors may complicate the reasons why.
The Productivity Lag: A Feature, not a Bug
In a look at productivity, a recent Barron’s analysis focuses on what economists Erik Brynjolfsson and Chad Syverson first highlighted: technological revolutions often generate a “productivity J-curve” where early investment precedes measurable gains by years. Their research shows that general-purpose technologies—such as electricity, computing, and now AI—require extensive organizational adaptation before they yield significant economic impact.
This phenomenon is well documented, and long running. The key components of electrification appeared on the scene by the late 1880s, but impacts took decades to show up in factory-level productivity figures; a century later the Solow Paradox famously mused that “you can see the computer age everywhere but in the productivity statistics.”
While it seems unlikely that AI is poised to permeate as slowly—in gross, quantifiable terms—than earlier analogs, it may yet not wholly buck the historical pattern. Indeed, despite the widespread initial footprint, when it comes to industrial impacts, the Massachusetts Institute of Technology’s (MIT’s) enterprise AI research found that fewer than 5% of organizations deploying AI saw significant measurable return on investment (ROI). The drivers of this fact are a combination of familiar ones observed throughout the history of technological innovation, and new ones specific to the nature of AI tools and technologies.
The History of Technological Transformation and Productivity
Two major factors are important in understanding the contemporary slower-than-predicted AI adoption. The first relates directly to the adoption curves of past technological revolutions already mentioned. To wit: technological innovations require extensive organizational adaptation before they yield significant economic impact. This is demonstrably true in the case of electrification, whereby electrified factories—optimized for pre-electrification steam-era layouts—initially lost productivity. Only when manufacturers redesigned workflows around electricity’s flexibility did productivity accelerate. Paul David’s comparison of the economic impacts and prerequisites to productivity gains inspired by the dynamo and the computer demonstrate a similar need for substantive, intentional complementary organizational change before the potential of each new technology could be realized.
There is little question that AI will at least, to some extent, hew to the now-established paradigm. Brynjolfsson, Rock, and Syverson, for example, argue that AI requires even more extensive “co-invention” than past technologies because it reshapes cognitive tasks and decision structures. Organizational redesign, data architecture, skills development, and new workflows take time, and transformation cannot begin until those foundations are in place.
AI’s early hype cycle certainly underestimated the time required for meaningful integration. The technology dawned quickly; the organizational transformation it requires will not. That said, a lesser-discussed factor may in fact be responsible for MIT’s much-discussed findings, in concert with the historical trend.
Horizontal AI: Visibility vs. Transformation
AI deployments can be characterized as one of two typologies: horizontal or vertical. Failing to differentiate between the two plays no small role in the gap between permeation in the market, which comes on the back of the former, and productivity, which should be expected to arise as a function of the latter.
Simply put, horizontal solutions, including generic chatbots, copilots, and summarizers—the tools that the average user may now interact with every day—are easy to deploy but fundamentally limited in their ability to influence operational performance. Horizontal AI technologies have certainly created awareness of AI tools, at least conceptually. Designed to perform general reasoning across broad domains, they can generate content, summarize text, or assist with simple tasks. In many cases, they may be utilized to expedite existing workflows. But they are not built with the domain-specific logic needed to support real industrial decision-making.
This limitation explains why early adopters have struggled to scale impact. Organisation for Economic Co-operation and Development analysis found that most firms fail to capture meaningful gains from AI until systems are paired with high-quality operational data and deep domain integration. Research by McKinsey similarly showed that while generative AI can boost productivity in isolated knowledge tasks, it delivers far greater economic value only when embedded into the core value chain of an industry.
Vertical AI: The Rise of Technology-Driven Productivity
The next phase of AI adoption will be driven not by horizontal systems, but by vertical AI: solutions engineered for specific industries, built on domain data, and aligned with operational context. Vertical AI technology represents an entirely different set of tools, applications, and deployments, and, properly implemented, its impact tends to show up directly on organizational bottom lines, and in turn, in measures of productivity.
Industry research points to the same conclusion. Deloitte, Accenture, and BCG have all reported that the majority of economic value created by AI in industrial sectors comes from deeply integrated, domain-aligned systems, rather than general-purpose assistants.
If the key to productivity gains and the ROI that otherwise seems so elusive is vertical AI technologies and solutions, why is it that we’re still predominantly utilizing horizontal ones? The answer lies in unmanned aerial vehicles (UAVs). Or pizza. Or chocolate chip cookies. Any of the three can be obtained off the shelf of your preferred retailer, and many variations of each can demonstrate utility. UAVs can be utilized for sport, or entertainment, or photography, etc. Even cheap pizza frozen within a cardboard box tides a college student over indefinitely. All three of these examples can usefully be likened to horizontal AI: any number, variety, and configuration of each example are readily available, attainable, and deployable to and by any consumer. And, at least from a distance, might be conflated with other, more advanced, professionally deployed versions of the same.
Advanced unmanned aerial systems, however, in this day and age of electronic jamming and countermeasures, fiber-optic controls, and high-explosive payloads, can and should readily be differentiated from what may appear on your holiday shopping list for nieces and nephews. That’s to say nothing of autonomous stealth bombers. Your last visit to your preferred hole-in-the-wall or high-end eatery is all that’s required to reinforce the difference between mass produced and consumed consumables, and the uniquely American thick-crusted delicacy produced to meet your specifications by people who are experts in its crafting. And nothing in your life will approximate a chocolate chip cookie straight out of your grandmother’s oven. All three of these examples, in this analogy, would represent vertical solutions. Solutions crafted by experts for specific domain applications, which in reality share only superficial qualities with their horizontal counterparts.
In this vein, vertical AI solutions may utilize broadly familiar tools, including large language models, chain-of-thought problem solving, and optical character recognition (OCR) applications recognizable to many workplace participants. But whereas your enterprise-licensed GPT (generative pre-trained transformer) or copilot might struggle to accurately extract data from an unstructured data source fed page by page into your user interface, advanced tools can extract hundreds of thousands of assets from hundreds of piping and instrumentation diagrams via automated pipelines, wherein Computer Vision, OCR, and reasoning tools work in concert not only to perceive the data but to understand it; to extract not only textual data but to parse and decipher the meaning behind images; and to render all in a vectorized database demonstrating true semantic understanding of the underlying data, and, more importantly, that which it depicts and represents.
While the best horizontal AI tools will produce incomplete and inaccurate outputs, engineered vertical solutions will efficiently and accurately produce a queryable database wherein a semantic search tool crafted by data scientists will breadcrumb an end user to the appropriate data and insights—determined automatically based on their role and use history—to aid them in their unspoken task. These tools understand what’s upstream, or downstream, of a given asset; how the pressure and thermodynamics pertaining to a particular asset influence its potential maintenance needs and lifecycle; and surface appropriate preventive maintenance tasks, automate appropriate work orders, and autonomously identify gaps in data, flag outlier readings, and propose actions, all in real time. And all of this is accomplished as easily as a novice asks their favored copilot for help with a task for which it will produce an empirically incorrect response that the user in turn may or may not perceive, appreciate, or leverage.
This is the crux of the dissonance between perception of AI ubiquity and productivity: conflation. The misunderstanding of one set of simple, available, productivity aids with a stealth bomber, world-class Detroit-style pizza, and grandma’s famous homemade cookies.
Toward Productivity Gains
While it remains true that, in pursuit of the eventual promise of productivity tomorrow from today’s implementations, organizations must build complementary capabilities, redesign workflows, modernize infrastructure, and integrate technology into core operations, AI will, in fact, follow its own curve. AI technologies themselves, in many cases, will represent both the problem and the solution, when it comes to data organization and automating workflows in particular. Requisite capability development underpinning the most advanced technological solutions, this time, ironically enough, will pivot on a lynch pin of human capital. In this revolution, data scientists, domain subject matter experts, and developers are the hard limiters, and all are finite resources.
For better or worse, there are no cheat codes or shortcuts when it comes to leveraging the most advanced technologies to their furthest and highest uses. AI systems, however, do not represent mere incremental potential improvements. Properly deployed they reshape workflows, compress timelines, reduce downtime, and improve decision quality. They create the very structural productivity gains economists have long associated with general purpose technology diffusion.
Ultimately, while horizontal AI may have jumpstarted the conversation, vertical AI will deliver the results. The slower-than-expected pace of AI-driven transformation is not evidence of failure. It is evidence that, as has always been the case, it will take time for industries to invest appropriately in the resources required to fully integrate AI into complex operational environments.
As the novelty era, dominated by chatbots and copilots, approaches its end, the era of engineering, defined by vertical AI and unprecedented productivity gains and ROI, is in its infancy. We yet expect it to mature with stunning alacrity.
—Chris Wiles, PhD is a data science consultant with 1898 & Co., a part of Burns & McDonnell.
