Computer Vision in Digital Transformation: The Dynamo Lesson

Most digital transformation programs share a blind spot that rarely appears on the roadmap. Over the past two decades, enterprises have digitized almost everything that passes through a keyboard,
including orders, invoices, customer records, approvals, and supply plans.

The physical work those records describe is a different story. Think of a pallet waiting on a dock, a line running below rate, or an operator stepping into a restricted zone. Those moments are still captured largely by people with clipboards and good memories.

That gap has a historical parallel that is more useful than any vendor framework. When factories first adopted electric power, most of them changed the engine and kept everything else. The productivity gains arrived only after managers redesigned the factory around the new technology, and that redesign took a generation.

This article uses that story to examine the role of computer vision in digital transformation. It also explains why agentic vision is the operational equivalent of redesigning the factory floor, and where transformation leaders can start.

The Blind Spot in Digital Transformation

For most enterprises, digital transformation has been a transformation of information work. Most digital transformation strategies centered on ERP, CRM, and workflow platforms, which made transactions visible, searchable, and auditable. The system of record now knows what was ordered, what was invoiced, and what was scheduled. It rarely knows what happened on the floor between those events.

Physical operations still report through a narrow and delayed channel. A shift supervisor fills in a downtime log at the end of the day. A safety audit samples a few hours each month. A quality check inspects a fraction of units.

Everything else goes unrecorded, even though most sites already have cameras pointed at the work.

The CCTV footage exists, but it is archived instead of read. Leaders are left with limited operational visibility into the processes they are accountable for. Customer-facing sites have the same gap, since queue lengths and wait times shape the ability to improve customer experience but are rarely measured as they happen.

This is where computer vision in digital transformation becomes a practical concern for directors and VPs. Cameras are often the only sensor already installed across the physical estate. Computer vision is the technology that turns their pixels into structured events. Seen this way, computer vision supplies the missing input to transformation programs that are already underway.

What the Dynamo Taught Industry About Transformation

In 1990, the economic historian Paul A. David published a short paper in the American Economic Review titled The Dynamo and the Computer. His subject was a puzzle economists were struggling with at the time: why the spread of computers had not yet produced a visible rise in productivity.

David looked for an answer in the history of electrification. The first factories to adopt electric power mostly replaced their central steam engine with a large electric motor. They kept the overhead line shafts, belts, and pulleys that carried power across the building, along with the layouts those shafts required. The new technology was installed, but the factory still ran on the logic of the old one.

The gains came later, once engineers began to attach a smaller motor to each machine. Power no longer had to travel along a shaft, so machines could be arranged by the flow of work instead of the geometry of the drivetrain. Entire plants were laid out again around the product. Erik Brynjolfsson, Daniel Rock, and Chad Syverson summarize the lag in their research on general-purpose technologies:

“In the later GPT case of electrification, it took a generation for the nature of factory layouts to be re-invented in order to fully harness the new technology’s benefits.”

Brynjolfsson, Rock, and Syverson, The Productivity J-Curve

The lesson for transformation leaders is an uncomfortable one. Installing a general-purpose technology is the easy part. The value sits in the redesign that follows, and most organizations underinvest in it.

The Productivity Paradox Has a Physical Edition

Robert Solow captured the original puzzle in 1987 when he wrote that “you can see the computer age everywhere but in the productivity statistics.” The line became known as the productivity paradox, and it has resurfaced with nearly every wave of enterprise technology since.

Research on the paradox points to the same explanation David found in the dynamo. In a widely cited review of firm-level evidence, Erik Brynjolfsson and Lorin Hitt concluded that “organizational ‘investments’ have a large influence on the value of IT investments.” They were also specific about where the constraint lies:

“As computers become cheaper and more powerful, the business value of computers is limited less by computational capability, and more by the ability of managers to invent new processes, procedures and organizational structures.”

Brynjolfsson and Hitt, Beyond Computation

Computer vision has reached the same point. Detection accuracy, inference cost, and deployment tooling have improved enough that models now hold up in the real world. The constraint has moved to the process the model is plugged into. This is why many computer vision projects fail after a technically successful pilot, and why scaling AI vision across sites is largely an organizational challenge.

Three Signs Computer Vision Is Running on the Old Line Shaft

Most stalled deployments share a pattern. The camera and the model are new, but the surrounding process is unchanged, so the output has nowhere useful to go. Three versions of this pattern appear again and again in operations:

1. The dashboard that nobody opens: people counting, utilization, and downtime logging data flow into a business intelligence tile that is reviewed at a monthly meeting, long after anyone could have acted on it
2. The alert that joins the inbox: a detection triggers a message to a supervisor who already receives dozens each shift, so the alert becomes one more item to clear
3. The pilot that automates the clipboard: a PPE detection system repeats the monthly audit at a higher frequency, producing a compliance percentage without changing who responds, how fast, or what happens next

In each case, the organization has digitized the observation and left the decision untouched. That is the computer vision equivalent of bolting a dynamo onto a line shaft.

The Technology Advances Behind Agentic Vision

At a high level, computer vision work follows the same path in every deployment. Cameras capture images or videos, computer vision algorithms extract visual information from each frame, and the results go to the people and systems that act on them.

For most of the past decade, that extraction relied on convolutional neural networks (CNNs). They brought dependable pattern recognition to tasks such as detection, counting, and inspection. Vision transformers and the same technology advances that produced generative AI have since added models that can describe a scene and answer questions about it.

Hardware has moved just as quickly. Edge processors now deliver high performance at low power, so computer vision systems can analyze images and videos in real time and at large scale without sending footage to the cloud. The same perception stack also supports augmented reality, which overlays guidance on what a technician sees.

With perception this capable, the remaining work sits in the process around it. That is where agentic vision comes in.

Agentic Vision Is the Unit Drive for Operations

Unit drive puts power at the machine. Agentic vision puts perception and judgment at the process step.

Agentic computer vision describes systems that combine visual perception with reasoning and the ability to take bounded actions. A conventional video analytics pipeline detects an object and emits an event. An agentic system can interpret that event in context and gather more evidence from other cameras or systems. It can then decide whether a response is needed and carry it out through the tools it has been given.

Several developments make this practical in operations today:

• Vision language models that can reason about a scene described in plain language, in addition to classifying it against fixed labels
• Natural language access to footage, which lets operators query cameras much as they query a database, a shift explored in when cameras become queryable
• Edge computing for computer vision, which keeps latency low and sensitive footage on site, in the same way the unit motor placed power where it was used
• Integration with systems of record, so an agent can open a work order, update a schedule, or escalate to a named person

The result is a new division of labor. Cameras become participants in the workflow, able to notice, interpret, and hand off work while it still matters. This is what physical AI looks like when it is applied to the operations a company already runs.

Line Shaft Thinking vs Unit Drive Thinking

The difference between a bolted-on deployment and a redesigned one shows up in the questions a team asks at the start. The comparison below is a quick way to test which kind of program is taking shape.

Agentic Vision in Operations: Where to Start

For a director or VP responsible for transformation, the practical question is where agentic vision in operations earns its place first. A useful sequence starts from decisions instead of a catalog of use cases:

1. Find a decision that is made too late: look for processes where the physical record is manual, sampled, or reconstructed after the fact, such as bottleneck identification, changeovers, or dock scheduling
2. Map the decision, then the detection: write down who should act, within what time, and through which system, before choosing a model
3. Set the boundaries of autonomy: define which actions the agent may take alone and which require approval, with human-in-the-loop review for anything with safety, quality, or employment consequences
4. Connect to the system of record: an agent that cannot write to maintenance, quality, or scheduling systems will fall back to sending alerts
5. Budget for the redesign: set aside time and money for process changes, training, and change management, in addition to cameras and compute
6. Measure the process: report cycle time, response time, and exceptions resolved, and treat model metrics as an engineering concern

Governance belongs in this sequence from the first step. Regulations such as the EU AI Act set requirements for how certain AI systems are classified, documented, and overseen. Systems that observe workers deserve particular care. A computer vision readiness assessment is far more useful before a pilot than after it.

Plan for the J-Curve

The J-curve research carries one more lesson. Brynjolfsson, Rock, and Syverson argue that general-purpose technologies depend on complementary investments that are “often intangible and poorly measured,” which is why measured productivity can dip before it rises.

In practice, the first months of an agentic vision program will show costs more clearly than benefits. Process mapping, integration work, and training do not appear on a camera dashboard. Leaders who expect this can protect the program through the dip. They can also set milestones around decisions improved rather than models deployed, and build the case for computer vision business value based on process outcomes.

The factories that gained most from electrification were the ones that redesigned the floor. The organizations that gain most from computer vision will be the ones that redesign the decisions around it.

Computer Vision in Digital Transformation: Designing Around Perception

Digital transformation made information work visible. For computer vision in digital transformation, the next phase is to make physical work visible and, where appropriate, actionable while it is happening. That requires treating computer vision as infrastructure for decision-making, an idea further developed in AI as an infrastructure.

Viso Suite is an enterprise platform for building, deploying, and managing computer vision applications on live camera feeds across sites. Teams that want to explore ideas on recorded footage first can start with Viso Now, which analyzes uploaded video without any model training. To discuss where agentic vision fits in your operations, book a demo with our team.