History Predicted It Already. AI Investment Won't Produce Immediate ROI

AI transformation moves at the speed of human change, not at the speed of technological capability. This is a hill I am willing to die on.

A frustrated CEO recently shared something that captures what’s happening in many organizations across every industry: ‘We’ve spent $2 million on AI in the past two years. Our adoption metrics look great; people are using these tools, but I cannot point to a single business outcome that improved greatly because of this investment. What are we doing wrong?’

The honest answer is that, like many, he is probably not doing anything wrong that can be easily fixed. He is encountering one of the most predictable patterns in technological history, a pattern that goes by a name most executives have never heard but should understand deeply if they want to make sense of their AI investments.

The pattern is called diffusion lag, and it explains why even the most powerful technologies take decades rather than years to produce the economic benefits their proponents promise. This was made popular in an article titled ‘AI as Normal Technology’ by two professors at Princeton University. It explains why electricity was ‘everywhere but in the productivity statistics’ for forty years after its commercialization. It explains why computers showed up in offices for decades before they actually changed how work happened. And it explains why your AI investments aren’t producing the returns you expected, even when the technology itself is working exactly as designed.

If you’re a leader trying to understand why AI ROI is so elusive despite genuine technological capability and substantial investment, the diffusion lag framework will reframe your thinking in ways that change how you approach the entire challenge.

The Statistics That Reveal the Real Story

A study released last year by researchers at the National Bureau of Economic Research made headlines because of an attention-grabbing finding: 40% of U.S. adults reported using generative AI by August 2024. This number was widely cited as evidence that AI adoption was unprecedented in speed, faster even than the adoption of personal computers in earlier decades.

But buried in the same study was a finding that received far less attention. Despite 40% adoption, generative AI use accounted for only 0.5% to 3.5% of work hours. The estimated impact on labor productivity was just 0.125 to 0.875 percentage points. In other words, most of the people who reported using AI were using it infrequently, briefly, and for tasks that didn’t fundamentally change their work output.

This is the gap between adoption and diffusion that most analysts miss. Adoption measures whether people have tried something. Diffusion measures whether it has actually changed how they work, what they produce, and what economic value they create. By the adoption metric, AI is spreading rapidly. By the diffusion metric, AI is barely making a dent in actual productivity despite billions of dollars in investment.

When researchers compared this pattern with the adoption of personal computers, the superficial comparison made AI seem faster. Forty percent of U.S. adults adopted generative AI within two years of the first mass-market product release, compared to 20% adopting PCs within three years.

But this comparison ignores critical differences. While PCs required substantial financial investment in an era when that mattered, AI is largely free at the point of use. PC adoption metrics measure ownership and regular use. AI adoption metrics measure people who have tried it once or twice. When you account for intensity of use rather than mere access, it’s entirely possible that AI is actually diffusing more slowly than PCs did, not faster.

This matters enormously for your business expectations. If you are benchmarking your AI ROI expectations against the assumption that AI is the fastest-diffusing technology in history, this could be farther from reality. Practically, AI is diffusing through your organization at roughly the same pace human and institutional change can absorb, which is to say, much more slowly than the technology’s capabilities would suggest.

The Inconvenient History of Transformative Technology

The story of electricity provides the clearest illustration of how transformative technologies actually create economic value, and why the timeline rarely matches the hype.

Thomas Edison opened his first central electrical generating station in 1882. Within two decades, electricity was widely available in industrial cities across America. By every reasonable adoption metric, electricity was spreading rapidly. Factories were being built, and equipment was being electrified. This technology was more flexible, cleaner, easier to control, and capable of powering individual machines rather than entire factories, superior to the steam power it replaced.

Yet for nearly forty years, electricity produced almost no measurable gains in productivity in American manufacturing. Robert Solow, the Nobel-winning economist, famously observed that you could see the computer age everywhere except in the productivity statistics, but this same observation had applied to electricity decades earlier. Economist Paul David documented this phenomenon in detail, showing that the productivity gains from electrification didn’t fully materialize until the 1920s, despite electrification being widely available since the 1890s.

What was happening during those four decades of seemingly wasted potential? Factory owners were trying to use electricity within the framework of existing industrial production. They were essentially using electric motors to replace steam engines while keeping everything else the same: the factory layouts, the workflow patterns, the management approaches, the labor organization, the training systems. The technology was working, but the surrounding system wasn’t designed to return value from it.

The breakthrough came when manufacturers fundamentally redesigned factory operations. The introduction of assembly line production, which became possible only when individual machines could be electrically powered rather than mechanically driven from a central source, required completely new factory layouts. The new layouts required new workflow patterns, and the new workflows required new management approaches. The new management approaches required new training systems and different kinds of workers with different skills. Each of these changes took years to develop, test, refine, and spread across industries.

By the 1920s, when these systemic changes had finally been worked out, manufacturing productivity exploded. The gains weren’t from electricity itself—electricity had been available for decades. The gains were from finally redesigning the entire system of industrial production around the logic of what electricity made possible. The technology had been ready since the 1890s. The human, organizational, and institutional changes required to capture its value took another thirty years.

This pattern is not unique to electricity.

It applies to every general-purpose technology in history. Railroads required new business structures and labor arrangements before transforming commerce. The telegraph required new information practices before reshaping how organizations operated. Personal computers required new workflow designs before changing knowledge work. Each time, the technology arrived faster than humans could figure out how to use it well, and the gap between technological capability and economic impact stretched across decades rather than years.

If you’re expecting AI to follow a fundamentally different pattern, to produce dramatic ROI within months because the technology is so impressive, history says that you may be wrong.

Why Software Doesn’t Solve the Diffusion Problem

A reasonable objection at this point would be that AI is fundamentally different from electricity or railroads because of how it’s distributed. AI software can reach billions of devices instantly through the internet; it doesn’t require building physical infrastructure or distributing physical products.

Surely this changes the diffusion equation?

Actually, it doesn’t. The availability of a technology is not the same as its effective use. Even when AI becomes instantly accessible to everyone on the planet for free, the human work of figuring out how to use it productively remains. That work moves at the speed of human change, not at the speed of software deployment.

Consider what is actually required to derive value from an AI tool in a knowledge-work context. An individual professional needs to learn what the tool can and can’t do. They need to develop intuition for when to use it and when to do something else. They need to integrate it into their existing workflow without disrupting the work they are already doing successfully. They need to develop quality control practices to catch the mistakes the tool will inevitably make. They need to figure out how AI-assisted work fits into the broader systems they operate within, how to share AI outputs with colleagues who haven’t yet adopted similar tools, how to maintain quality standards across hybrid human-AI work, and how to evolve their own role to remain valuable as AI capabilities grow.

None of this happens in the moment someone signs up for ChatGPT or Claude, or when a company builds their own agent or automation. It happens over months and years of experimentation, adjustment, and learning. It happens at different paces for different individuals, depending on their job, their cognitive style, their willingness to change established habits, and the support they receive from their organization.

Now multiply this individual-level adaptation by the organization-level changes required. The workflows need to be redesigned, the processes rebuilt, the decision rights clarified, and quality control systems developed. Governance frameworks need to be established, training programs need to be developed, career paths need to be reconsidered, compensation systems need to be reviewed, and performance metrics need to be updated.

Each of these organizational changes takes months or years to design, implement, refine, and absorb.

And then there’s the institutional level, the changes that have to happen across industries, in regulatory frameworks, in educational systems, in professional standards bodies. Industries need to develop a shared understanding of what good AI practice looks like. Regulators need to figure out what oversight is appropriate. Educational institutions need to redesign programs to prepare people for AI-augmented work. Professional associations need to update standards and ethical guidelines. These institutional changes happen even more slowly than organizational changes, and they ultimately shape what’s possible at the organizational and individual levels.

The technology can reach billions of devices in days. The human, organizational, and institutional adaptation required to turn that technology into economic value may take decades. This is not a problem to be solved through better technology. It is a structural feature of how transformative technologies actually create value, and pretending otherwise leads to exactly the kind of disappointed AI ROI conversations happening in boardrooms right now.

What This Means for Your AI ROI Expectations

If you are a senior leader trying to make sense of why your AI investments aren’t producing the returns you expected, the diffusion lag framework gives you a much more accurate mental model than the popular narrative of rapid AI transformation.

1. The first implication is that your timeline expectations have probably been wrong. While the popular narrative suggested that companies investing heavily in AI in 2023 and 2024 would see transformative returns by 2025 and 2026, the historical patterns suggest that companies investing heavily now should expect significant value to emerge over five years, not eighteen to twenty-four months. This is because the human and organizational adaptation required to extract value from the technology takes time, and that time has always been measured in years rather than months.

This reframing matters for everything from how you budget to how you communicate with boards. If you are reporting AI ROI on quarterly cycles and treating each quarter without dramatic improvement as failure, you are misunderstanding what is happening. The investments you are making now are creating capability that will produce returns years from now, assuming you sustain them through the long period when results look worse than the investment.

1. The second implication is that the work required to capture AI value is dramatically different from the work most organizations are doing. Most companies are focused on technology deployment, training programs, and adoption metrics. These activities make sense if you believe the technology itself will produce the value. They don’t make sense if you understand that the value comes from human and organizational adaptation that the technology only catalyzes.

The work that actually produces returns looks different. It involves deep workflow analysis, mapping how work actually flows in your organization today, not how org charts claim it flows. It involves systematic redesign of processes around AI capabilities, not just adding AI to existing processes. It involves change management of a scope most organizations underestimate, addressing not just skill development but identity transformation as people’s roles evolve. It involves governance development, quality control system creation, performance metric redesign, and career path reconsideration. While this work is unglamorous, slow, and often invisible, it is where AI value actually gets created.

1. The third implication is that the organizations that will eventually dominate their industries through AI are the ones doing the unglamorous work now, while their competitors are still focused on technology and tools. This is the same pattern that played out with every previous transformative technology. The companies that ultimately captured the most value from electrification were the ones that rebuilt their factories around the logic of production lines, a process that took years of systematic work and didn’t produce impressive results along the way. The same will be true with AI.

What does this mean for Leaders?

Understanding diffusion lag changes several strategic questions in important ways.

The question changes from ‘what AI technology should we adopt?’ to ‘what organizational changes do we need to make for AI to eventually produce value?’ While the first question leads to technology evaluation and vendor selection. The second question prompts an honest assessment of your workflows, culture, governance, training systems, and willingness to invest in change that won’t show results for years.

The question changes from ‘how quickly can we deploy AI?’ to ‘how do we sustain investment through the productivity lag period?’ Most AI initiatives die not because they’re poorly conceived but because organizations don’t sustain commitment through the period when results look bad. Understanding that this period is normal, even necessary, is what allows organizations to push through it.

The question changes from ‘what’s our AI ROI?’ to ‘what capabilities are we building that will eventually create economic value?’ The honest answer in most organizations right now is that AI ROI in the short term is genuinely poor. The right response is not to abandon AI investment but to reframe what you’re measuring.

The question changes from ‘how do we use AI?’ to ‘how do we redesign our organization around AI possibilities?’ These sound similar but lead to very different work. The first question results in adding AI to existing workflows. The second question requires a fundamental rethinking of how work gets done in your organization.

The question changes from ‘what’s our AI strategy?’ to ‘what’s our human and organizational transformation strategy that AI enables?’ A pure AI strategy treats the technology as the point, but a more accurate framing recognizes that the technology is a catalyst for organizational transformation, and the strategy work is really about that transformation.

The Honest Message Your People Need

If you are a CEO or senior executive who has been promising rapid AI ROI to your people or investors, you may need to have an uncomfortable conversation. The honest message is that the timeline for value capture is longer than commonly understood, and the investment required is more about organizational change than technology deployment.

Here’s a version of that conversation that gets the framing right: ‘Our AI investments are positioning us to capture transformative value over the next three to five years. We are following the same pattern that determined which companies dominated their industries through previous waves of general-purpose technology adoption: the companies that did the hard organizational work while competitors focused on the technology itself. In the short term, our results will look modest because the human and institutional changes required to extract value from AI take time to develop. This is not a failure of strategy. It is how transformative technologies have always produced economic value, and the companies that maintain investment through this lag period are the ones that ultimately win.’

Without this reframing, you face a different and worse conversation: explaining why your competitors are seeing similar results to yours, why nobody seems to be capturing the value that was promised, and why you should continue investing in something that isn’t working. This conversation has predictable outcomes: budget cuts, project terminations, leadership changes, and, ultimately, failure to develop the capability your competitors will eventually exploit.

The Opportunity Hidden in the Frustration

The widespread frustration with AI ROI is creating an opportunity for leaders who understand the diffusion lag pattern. While most organizations are pulling back from AI investment because they are not seeing immediate returns, organizations that maintain or increase investment now are positioning themselves to capture disproportionate value when the diffusion eventually catches up.

The pattern from electrification comes in here. The factory owners who maintained investment in electrification during the early decades, when productivity gains weren’t materializing, were the ones who eventually dominated their industries when the production line revolution finally arrived. The factory owners who abandoned electrification because it wasn’t producing immediate returns missed the productivity explosion entirely. They had to scramble to catch up, and many never did.

The same pattern is likely with AI.

The organizations doing the unglamorous work of organizational transformation are now building capability that will be very difficult for competitors to replicate quickly, even though it’s not producing immediate ROI. When the broader institutional and organizational changes finally make AI value capture more straightforward, these organizations will be positioned to capture disproportionate gains.

The Practical Implications

If you’re persuaded by this framing, several practical implications should follow.

1. Stop measuring AI ROI primarily through short-term productivity metrics. These metrics are misleading during the diffusion lag period. Rather, develop better metrics that track capability building, organizational adaptation, and learning. Track metrics such as completed workflow redesigns, employees demonstrating advanced AI fluency, established governance frameworks, and achieved process integrations. These leading indicators predict future value capture even when current ROI metrics look poor.
2. Invest disproportionately in the organizational work that creates AI value. Process redesign, change management, governance development, training infrastructure, and cultural transformation should receive much more investment than they typically do in AI initiatives. The technology investment should be smaller than most companies make it, and the organizational investment should be larger.
3. Communicate the timeline honestly to all stakeholders. Set expectations that significant returns will emerge over the years rather than quarters. This reduces the pressure to abandon initiatives prematurely and protects investments during the difficult middle period when results look their worst.
4. Develop strategic patience as an organizational capability. The companies that succeeded with previous general-purpose technologies had the discipline and conviction to maintain investment through the long diffusion lag.
5. Watch competitor behavior carefully. Many will pull back from AI investment as ROI disappointments accumulate. This is the moment when relative position is most easily improved. Maintaining investment while competitors retreat creates compounding advantages that become harder to reverse over time.

If you have been frustrated by your AI investments and tempted to pull back, the historical pattern suggests this is exactly the wrong response. If you have been wondering whether you should improve investment despite poor short-term returns, the historical pattern suggests this is exactly the right intuition.

Will your organization be positioned to get the real ROI on AI INVESTMENTS? The answer depends on choices being made right now, during the frustrating period when the data looks worst, and the temptation to retreat is strongest.

Choose wisely.

Originally published in The AI Catalyst Report ↗