Can AI Agents Make Continuous Improvement Truly Continuous?

For more than a century, manufacturers have relied on time and motion studies, manual observations, and periodic Kaizen events to improve production processes. AI agents capable of understanding manufacturing operations through video intelligence could fundamentally change how continuous improvement is performed, making it faster, more scalable, and potentially continuous.

What if continuous improvement no longer required weeks of engineering analysis, manual time studies, and dedicated improvement events?

What if an AI agent could continuously observe manufacturing processes, identify waste, investigate problems, and recommend improvements while engineers focus on implementing the changes that matter?

That's the central question I explored with Devin Bhushan, founder and CEO of Squint, in this episode of the AI in Manufacturing podcast.

Squint is developing industrial AI agents that use video intelligence to understand manufacturing processes, capture operational knowledge, and automate many of the activities traditionally performed by industrial engineers and continuous improvement teams.

Our conversation explored how these capabilities are being applied in real manufacturing environments, including a project where an AI agent completed in a single day an analysis that had previously taken eight engineers 78 days.

We also discussed how AI agents could move beyond optimizing individual processes toward continuously improving entire production lines, factories, and eventually manufacturing enterprises.

Why Is Continuous Improvement Still So Labor-Intensive?

Manufacturers have spent decades developing increasingly sophisticated approaches to operational excellence.

Lean manufacturing, Six Sigma, the Toyota Production System, and other methodologies have provided systematic ways to identify waste, reduce variability, improve quality, and increase productivity.

Yet many of the activities required to implement these methodologies remain remarkably manual.

Time and motion studies still frequently involve engineers observing operators, recording activities, measuring cycle times, and analyzing the results using spreadsheets and charts.

Devin highlights an interesting contradiction.

Continuous improvement is fundamentally about identifying bottlenecks and eliminating unnecessary work. Yet the continuous improvement process itself has remained dependent on manual observation and engineering analysis for more than a century.

The methodologies have evolved, but the tools used to perform much of the work have changed considerably less.

This creates an important limitation.

A manufacturing facility may have hundreds of processes that could benefit from systematic analysis, but only a small team of industrial engineers or Lean specialists available to perform that work.

Consequently, improvement initiatives must be prioritized, and many opportunities remain unexplored.

Devin also observes that manufacturers experiencing rapid growth often have the greatest need for process improvements but the least time to conduct them.

Meanwhile, manufacturers facing cost pressures may be reducing the engineering resources they need to improve operational efficiency.

AI agents could help address both challenges by automating much of the observation, measurement, analysis, and documentation required for continuous improvement.

Why Is Video Intelligence Important for Understanding Manufacturing Processes?

One of the defining characteristics of Squint's approach is its use of video as a primary source of operational information.

Manufacturers already collect significant amounts of production data through MES, SCADA, historians, and other industrial systems.

These systems can tell us how long a production cycle took, how many units were produced, when equipment stopped, or whether a quality target was achieved.

But they don't necessarily tell us how the work was performed.

Consider two production shifts operating the same equipment and producing the same product.

One shift consistently achieves higher throughput.

Traditional operational data might reveal the performance difference, but identifying the underlying cause could require engineers to observe how operators perform their tasks.

Perhaps one team has developed a more efficient sequence of activities, eliminated unnecessary movements, or discovered a better way to handle materials.

These improvements may exist primarily as undocumented operational knowledge.

Devin argues that video provides a much richer representation of manufacturing processes because it allows AI agents to observe the interactions between people, machines, materials, and the physical environment.

Rather than relying exclusively on machine-generated measurements, the agent can examine how activities unfold and how different parts of a process influence one another.

Squint uses this information to reconstruct manufacturing workflows, generate process models, and identify opportunities for improvement.

This also creates an opportunity to capture the practical knowledge of experienced operators before that knowledge disappears through retirement or workforce turnover.

What Happens When an AI Agent Analyzes 144 Hours of Manufacturing Footage?

One of the most interesting examples Devin shared involved a turbine blade repair facility in Singapore.

The manufacturer had previously conducted a detailed continuous improvement exercise on a production line involving 18 operators.

That analysis had taken a team of eight engineers 78 days to complete.

The company wanted to evaluate whether Squint's AI agent could perform a comparable analysis using historical footage of the same process.

They provided 144 hours of video and gave the Squint team one day to analyze the production line and present their findings.

The agent examined the footage, reconstructed the production workflow, identified the activities performed at different workstations, and calculated relevant process measurements.

It generated time studies, charts distinguishing value-added and non-value-added activities, and an analysis of the overall production line.

The team then gave the agent a specific optimization objective: maintain the existing cycle time while reducing the number of operators required.

According to Devin, the agent recommended a configuration that would allow the line to operate with 12 operators instead of 18, potentially freeing six operators for other production activities.

It also generated revised workstation arrangements, workload-balancing charts, and simulations explaining how the proposed configuration could work.

But perhaps the most interesting findings concerned inefficiencies that the earlier engineering analysis had overlooked.

In one operation, an operator was polishing a component and repeatedly using compressed air to remove dust accumulating on the workpiece.

The agent observed that the workstation already had a downdraft table designed to remove dust.

It identified the manual dust-removal activity as potential waste and recommended investigating whether the downdraft system was operating effectively.

This illustrates an important capability.

The AI wasn't simply measuring how long an operator spent performing an activity. It was interpreting the activity in the context of the equipment and the intended manufacturing process.

Devin reports that the customer responded positively to the findings, including several observations that had not emerged during the original analysis.

The case demonstrates the potential to dramatically reduce engineering analysis time, although the proposed operator reduction was an optimization recommendation rather than a verified implementation outcome.

How Are AI Agents Different from Traditional Computer Vision Systems?

Computer vision is not new to manufacturing.

Industrial companies have used vision systems for decades to inspect products, detect defects, verify assembly operations, and automate quality control.

However, these applications have traditionally been designed around relatively specific tasks.

A vision system might inspect a component and determine whether it meets a quality specification.

Understanding an entire manufacturing process requires a different level of interpretation.

An AI agent may need to recognize operators, equipment, materials, and workstations, understand how they interact, and follow activities over time.

It must distinguish between productive activities, waiting, unnecessary movements, and other forms of process waste.

It also needs to understand the operational objectives and constraints that determine whether a particular improvement is practical.

Devin explains that Squint combines video understanding with industrial knowledge derived from Lean manufacturing, Six Sigma, and other established methodologies.

The agent can also incorporate additional context from work instructions, engineering documents, and operational systems.

A manufacturer might ask it to increase throughput by 15%, subject to a capital expenditure limit and without increasing headcount.

The agent can then analyze the process, evaluate potential improvements, and present recommendations along with supporting calculations.

Rather than simply classifying what appears in a video, the system is intended to perform a broader engineering investigation.

Importantly, it also supports interaction with human experts, allowing them to question assumptions, provide missing constraints, and refine recommendations.

Can AI Agents Investigate Manufacturing Problems Without a Predefined Solution?

Another interesting aspect of the discussion was whether AI agents must always be given a narrowly defined optimization objective.

Devin argues that they can also perform more open-ended investigations.

He shared an example involving a manufacturer of fiber-optic cable.

The company operated two production lines, but one consistently experienced a higher defect rate.

The manufacturer wanted to understand the underlying cause.

Rather than beginning with a predefined hypothesis, the investigation started with a broader question: why was one line producing more defects than the other?

According to Devin, the AI agent requested additional video footage from inside the production equipment.

The team installed a GoPro camera and recorded approximately 16 hours of operation.

The agent then proposed a design of experiments, specifying different operating conditions to test and the information needed to analyze the results.

By examining the footage alongside defect information from the MES, the agent identified small vibrations in the cable that correlated with the observed quality problems.

It was then able to narrow the investigation to specific operating conditions, timestamps, and production activities.

The manufacturing team used those findings to investigate and identify the root cause.

What's particularly interesting about this example is that the AI agent was not limited to a predefined set of rules for detecting a particular defect.

Instead, it helped structure an investigation, requested additional evidence, analyzed the results, and directed engineering attention toward a likely source of the problem.

This represents a shift from using AI primarily for detection toward using it to support industrial problem-solving.

Can AI Agents Optimize Entire Factories Rather Than Individual Processes?

Improving an individual workstation does not necessarily improve the performance of an entire production system.

A manufacturing operation consists of interconnected processes, resources, material flows, and constraints.

Increasing throughput at one workstation may have little benefit if the next workstation remains the bottleneck.

Similarly, optimizing one production line may produce unintended consequences elsewhere in the facility.

This raises a more interesting question: how can AI agents move from local process improvements toward optimizing manufacturing operations as a whole?

Devin describes this evolution in several stages.

The first stage involves automating individual tasks, such as generating work instructions or performing time studies.

The second involves giving an agent a broader objective, such as reducing production cycle time or investigating the causes of quality problems.

The next stage involves connecting the agent to information across the manufacturing facility.

With access to video, MES, ERP, and other operational systems, the agent could begin addressing problems that require a more comprehensive understanding of production.

For example, it could examine production schedules, demand, inventory, and material flows to recommend how manufacturing activities should be sequenced to reduce work in progress or improve delivery performance.

Devin's longer-term vision extends beyond individual factories.

He imagines agents comparing how similar manufacturing processes are performed across different facilities, identifying better practices, and adapting those practices for other sites.

A process improvement discovered at one factory could eventually become a recommendation for another factory operating comparable equipment under different conditions.

Much of this enterprise-wide capability remains an area of ongoing development, but the direction is significant.

Continuous improvement could evolve from a collection of relatively isolated projects into a more systematic process of learning and optimization across manufacturing operations.

Why Do AI Agents Need More Than Video Data?

Although video is central to Squint's approach, Devin acknowledges that reliable process optimization often requires additional operational context.

Video can reveal what is happening physically, but it cannot necessarily explain every constraint that determines how a manufacturing process should operate.

An agent might identify a more efficient workstation configuration without knowing that certain operators lack the qualifications required to perform particular activities.

It might recommend changing a production sequence without considering material availability, customer orders, or equipment maintenance requirements.

This is where integration with existing manufacturing systems becomes important.

ERP systems can provide information about production requirements, materials, and business constraints.

MES and historians can provide production measurements, events, and equipment operating information.

Learning management systems can provide information about operator training and qualifications.

Engineering documentation can help explain the design and intended operation of equipment and processes.

Combining these sources allows AI agents to evaluate recommendations against a more complete representation of manufacturing operations.

Devin also discussed how AI agents could simplify the integration process itself.

Rather than requiring developers to manually implement every connection to a customized manufacturing system, agents could increasingly help generate the integration logic needed to access relevant information.

The objective is to make it easier for manufacturers to connect existing systems to AI-driven workflows, particularly in environments with extensive legacy infrastructure.

How Do Manufacturers Turn AI Recommendations into Lasting Improvements?

Identifying an improvement opportunity is only part of the continuous improvement process.

The more difficult challenge is often implementing the change and ensuring that the new way of working becomes standard practice.

Manufacturers have long faced this problem.

An engineering team may conduct a detailed analysis, identify process improvements, and develop new standard operating procedures.

But translating those findings into consistent behavior across operators, shifts, and facilities requires training, communication, and sustained change management.

Devin considers this one of the most important challenges in industrial AI adoption.

Squint's approach is to connect process analysis with the activities required to implement changes.

For example, when a new process is developed, the system can compare it against the procedures operators were previously trained to follow.

It can then generate updated training material and provide interactive guidance to help operators learn the new workflow.

Devin describes using tablets and augmented reality to guide operators through procedures and demonstrate how activities should be performed.

The system can also support follow-up questions as operators become familiar with the revised process.

The broader objective is to make improvement implementation part of the same workflow as improvement identification.

Rather than producing a report that engineers must translate into separate training and implementation activities, the AI agent can help connect analysis, process redesign, documentation, and operator enablement.

This could become particularly important as the frequency of improvement recommendations increases.

After all, continuous improvement cannot become truly continuous if manufacturers lack the ability to implement and sustain the changes being recommended.

Where Should Manufacturers Start with AI-Driven Continuous Improvement?

For manufacturing leaders interested in exploring AI agents, Devin's recommendation is straightforward: start with a real operational problem.

Rather than attempting to document every process or introduce AI across an entire organization, identify an existing improvement initiative where the business value is clear.

That might involve reducing scrap, investigating a recurring quality problem, improving throughput, or eliminating a production bottleneck.

The key is to give the AI agent a meaningful problem that engineers are already trying to solve.

This creates an opportunity to evaluate the technology against an established business objective.

Can it reduce the time required to understand the problem?

Can it identify improvement opportunities that traditional analysis has overlooked?

Are its recommendations technically sound and practical to implement?

And can the resulting changes deliver measurable operational benefits?

Devin argues that demonstrating value on a specific problem is the most effective way to build confidence and encourage adoption.

Once a team experiences meaningful benefits, the approach can expand to additional processes, production lines, and facilities.

For manufacturers, it is equally important to retain appropriate engineering validation and operational controls. Recommendations involving equipment, safety, staffing, or production configurations need to be assessed against the realities of the manufacturing environment before implementation.

The objective is not to eliminate industrial engineering expertise, but to increase the amount of improvement work that engineering teams can perform.

The Future of Continuous Improvement in Manufacturing

One of the most interesting implications of this conversation is how AI agents could change the economics of continuous improvement.

Historically, manufacturers have needed to allocate dedicated engineering resources to observe processes, collect measurements, perform analysis, and develop recommendations.

Because these activities consume significant time, improvement initiatives are often performed periodically and prioritized according to their expected value.

AI agents could substantially reduce the effort required to perform much of that analytical work.

That creates the possibility of conducting more frequent investigations, identifying smaller improvement opportunities, and continuously revisiting processes as operating conditions change.

It could also change the role of industrial engineers.

Rather than spending much of their time collecting measurements and preparing analyses, engineers could increasingly focus on evaluating recommendations, implementing improvements, and addressing more complex operational challenges.

There are still important questions about reliability, integration, human oversight, and the ability to sustain improvements across different manufacturing environments.

But the potential direction is compelling.

Continuous improvement has traditionally been a methodology that manufacturers apply to their operations. AI agents could help turn it into a capability that operates continuously alongside those operations.

And that may be one of the most significant opportunities for agentic AI in manufacturing.

‍