The Next Models Are Coming, Agents Are Scaling — and None of It Matters Until You Rebuild the Process

AI as Normal Technology — and Why It Won't Work Until You Rebuild the Process

There is a particular kind of frustration that shows up in organizations that have adopted AI early and not seen the returns they expected. The models are good. The demos are impressive. The vendors are confident. But six months in, the ROI is unclear, the workflows feel awkward, and the promised transformation has not arrived.

This is not a technology problem. It is a process problem. And we have seen it before.

The Electricity Mistake

When electric motors were introduced to factories in the late 1800s, the initial approach was direct substitution. A factory built around a steam engine had its machinery arranged in a specific way: everything clustered near the central power source, connected by a complex system of shafts and belts. When electricity arrived, most factories simply replaced the steam engine with an electric motor and left everything else exactly as it was.

It did not work well. Power was more reliable. Costs went down slightly. But the transformative productivity gains that electricity promised did not materialize — for forty years.

Economist Paul David documented this in what became one of the most cited papers in economic history: factories were "everywhere but in the productivity statistics." The reason was structural. The benefits of electricity were not in the motor itself — they were in the entirely different factory layout that electricity made possible. Distributed motors. Each machine with its own power source. Workers able to move freely. Assembly lines that followed the logic of the work, not the logic of the power shaft. Natural lighting through repositioned windows. Separate buildings for separate processes.

You could not get those gains by plugging electricity into a steam-engine-based layout. You had to redesign the factory.

Narayanan and Kapoor cite this analogy explicitly, and they are right to. AI adoption is following the same pattern. We are in the "electric motor substituted for steam engine" phase. We have plugged AI into existing processes and are measuring the output. The tools are faster. Some tasks are cheaper. Individual workers save 40–60 minutes per day. But the structural gains — the ones that transform how organizations operate — have not arrived, because the factory has not been redesigned.

The Coordination Cost Nobody Talks About

Think about where the real friction in organizational work actually lives. It is rarely in the execution of a single task. Most professionals are reasonably good at their individual work. The friction is in the space between tasks: the hand-off from analyst to manager, the alignment call between sales and product, the three-email chain before a simple decision gets made, the meeting that exists only because two systems do not talk to each other.

Coordination costs are enormous, invisible, and almost entirely unaddressed by current AI deployments. We use AI to write the email faster. We do not use AI to eliminate the reason the email needed to be written. We use AI to summarize the meeting. We do not ask whether the meeting needs to happen.

In the language of the factory analogy: we have given every worker a faster tool. We have not asked whether the workflow connecting all those workers still makes sense.

Cross-functional work is where this is most acute. Consider any process that touches more than one team — a client onboarding, a product launch, a compliance review, a hiring decision. Each step in that process was designed around two constraints: human cognitive limits and the information-sharing friction of the pre-AI environment. Approvals exist because a human reviewing the work could only catch errors if explicitly looped in. Hand-offs exist because systems did not share data. Status meetings exist because there was no other way to maintain a shared picture of progress.

AI eliminates many of those constraints at the technical level. But the org chart, the process documentation, the approval workflows, and the meeting cadence all still reflect the world those constraints created. We have changed the tools. We have not changed the factory.

What Rebuilding the Process Actually Looks Like

Narayanan and Kapoor are careful to note that AI's economic impacts will unfold over decades, not years. This is not pessimism — it is the historical pattern for every general-purpose technology. The gap between invention (a new capability exists), innovation (it becomes a product), and adoption (it changes how organizations work) is measured in years to decades. The reason is not technology adoption speed. It is the time required to rebuild the organizational logic around the new capability.

What does rebuilding look like in practice? A few principles:

Start with the hand-offs, not the tasks. Map every place where work moves from one person, team, or system to another. That is where AI eliminates coordination overhead — not by speeding up tasks, but by making the hand-off unnecessary or seamless. An AI that monitors a process end-to-end and flags exceptions removes the need for a status update meeting. An AI that drafts the next step in a workflow as soon as the prior step completes removes the gap between "done" and "started."

Redesign approvals around risk, not ritual. Most approval workflows were designed for a world where errors were hard to catch. AI changes the error-detection calculus fundamentally. An AI that audits every transaction in real time makes sampling-based human review redundant. An AI that flags anomalies before they become problems changes what "oversight" means. Approval processes that exist for ritual compliance — the three-signature rule that no one remembers the origin of — can be eliminated. Approvals that exist because humans need to exercise judgment in ambiguous situations need to be redesigned around human-AI collaboration, not replaced.

Treat the org chart as a hypothesis, not a given. The structure of most organizations reflects the information-sharing constraints of a pre-AI world. Teams exist because coordination across teams was expensive. Hierarchies exist because information had to flow through people to reach decision-makers. When AI can make information available to anyone in real time, the rationale for many structural boundaries weakens. This does not mean flat organizations are always better — Galbraith and the org design literature are clear that structure serves purpose. But the purposes served by current structures should be re-examined against new constraints.

Measure friction, not just output. Most AI ROI frameworks measure task-level improvements: time saved writing the report, accuracy of the forecast, cost per query. These are real, but they miss the larger opportunity. The harder measurement is friction: how many handoffs does this process require? How much time elapses between the completion of one step and the start of the next? How many meetings exist to coordinate work that could be coordinated by a system? Reducing friction — even slightly — across a high-volume process compounds dramatically.

The Real Question

Narayanan and Kapoor end their essay with a point about human agency: institutions, choices, and decisions shape how technology develops. AI is not a deterministic force that will transform organizations on its own schedule. It is a set of capabilities that organizations have to actively choose to apply — and apply to the right problems.

The right problem is not "how do we make our current process faster." It is "what process would we design if we were starting today, with AI as a native capability?"

Imagine an organization where coordination across teams happens through shared AI-maintained context rather than recurring meetings. Where hand-offs between functions are monitored and completed by agents that know the state of every step. Where approvals are triggered by intelligent risk signals rather than calendar-based reviews. Where the friction of cross-functional work — the emails, the alignment calls, the status decks — has been systematically eliminated because the information flow no longer requires a human intermediary at every step.

That organization will not look like anything that exists today. But the capability to build it is already here.

The question is whether leaders are willing to do what the factory owners of the 1920s eventually had to do: not just adopt the new technology, but tear down the old layout and build something new around it.

Source essay: Arvind Narayanan & Sayash Kapoor, "AI as Normal Technology", Knight First Amendment Institute, April 15, 2025. | Background: Paul David, "The Dynamo and the Computer: An Historical Perspective on the Modern Productivity Paradox" (1990) | Galbraith, Designing Organizations | Narayanan & Kapoor, AI Snake Oil (Princeton University Press, 2024)