The Operations Bottleneck That Often Goes Unmeasured

Walk through a well managed manufacturing plant and there is usually no shortage of operational data. Teams monitor machine uptime, cycle times, scrap rates and throughput, then use that information to identify constraints and improve performance.

However, some processes that influence overall capacity receive much less attention. In particular, activities that depend on people reading, counting and extracting information from documents are not always measured as operational processes.

This matters because a small team responsible for reviewing documents can sometimes determine how quickly the rest of an organisation can move.

A process that may not appear on the dashboard

Similar patterns appear across many industries. Underwriters review submissions, procurement teams analyse tenders, claims specialists read reports and legal operations teams examine contracts.

These activities are often treated primarily as expert work rather than as processes that can be measured and improved. Human judgement is certainly important, but not every minute spent on a document involves judgement. A significant proportion of the work may consist of locating information, counting items and transferring data between systems.

Commercial casework manufacturing provides a useful example because document review can directly affect the number of projects a manufacturer is able to quote for.

The estimating process

Before manufacturing begins, commercial casework projects typically go through a detailed estimating process. An estimator may need to review architectural drawing sets containing hundreds of sheets and specification documents that run to hundreds of pages.

The drawings are examined to identify and count items such as cabinets, countertops, elevations and other relevant units. Specifications then need to be reviewed for requirements that may affect materials, hardware or other aspects of the final price.

Manufacturers may invest substantially in production equipment such as CNC routers, edgebanders and automated finishing systems. The earlier estimating process, however, has often depended on document viewers, spreadsheets and manual marking.

An estimator may review drawings sheet by sheet, identify units and record quantities by type. Depending on the complexity of the project, this can take several hours. Errors can also have consequences. If an item is missed or counted incorrectly, the resulting estimate may not accurately reflect the requirements of the project.

As a result, the document processing capacity of a relatively small estimating team can influence how many opportunities the wider business can pursue.

Technologies for document based workflows

Recent advances in computer vision and language models have made it possible to automate some of the more repetitive parts of these workflows.

One application is object detection for technical drawings. A computer vision system trained on construction documents can identify and classify particular objects on a drawing. This requires specialised training because architectural conventions vary between firms and projects. Symbols representing similar objects may differ significantly between drawing sets.

Another application is the use of language models to search and extract information from specification documents. Rather than manually searching through a long document for a particular requirement, a system can identify relevant passages and provide the associated page reference.

In both cases, the objective is not necessarily to remove human expertise from the process. Instead, automation can take on repetitive activities such as searching and counting, while people remain responsible for reviewing the results and applying judgement.

Results from a production deployment

A commercial casework manufacturer has used this approach in daily production since 2025. The system processes architectural drawings, detects and classifies cabinet units and produces structured takeoff data that can be used in the quoting workflow.

Reported results from the deployment include:

• Drawing review time reduced from two to six hours to approximately ten minutes.
• Capacity of approximately 20 drawing sets per estimator per day, compared with two to four previously.
• Specification lookup reduced from more than 30 minutes to less than 30 seconds, with page references provided.
• A drawing set containing 470 sheets has been processed.
• The system reached break even within one month.

Further details about the production implementation are available in this case study.

These results illustrate how changes to an upstream information processing task can affect operational capacity. In a bidding environment, increasing the speed of estimating may allow a business to evaluate and respond to more opportunities.

The important operational question is therefore not only how efficiently a factory produces work, but also how efficiently the organisation processes the information required before production can begin.

Accuracy and review time

Accuracy is an important measure for document processing systems, but it is not the only one that matters.

In one production implementation, automated detection accuracy improved from 76% to approximately 90% through systematic analysis of errors. Further improvements were technically possible, but there is an operational trade off to consider.

At around 90% accuracy, an estimator can review the generated output, identify errors and make corrections in approximately ten minutes. Even at a higher level of automated accuracy, a human review step may still be appropriate before a commercial quotation is submitted.

This means that the value of additional accuracy depends partly on whether it materially reduces review time. Improving a model from 90% to 97%, for example, may have less operational impact if the output still requires a similar review process.

The implementation process and accuracy improvements in the system provide an example of this trade off.

When evaluating this type of technology, organisations may therefore benefit from measuring both automated accuracy and the time required for a qualified person to review the result. Accuracy indicates how reliable the automated output is, while review time has a more direct relationship with throughput.

Common limitations

The limitations of these systems are relatively predictable and should be considered during implementation.

Computer vision models can sometimes merge adjacent objects into a single detection or fail to identify individual units. Image quality and symbol density can both affect performance. Drawings containing many closely positioned objects, particularly when available only at low resolution, can be more difficult to process accurately.

Differences between architectural drawing conventions also present a challenge. A system trained on one collection of projects may encounter unfamiliar symbols or layouts when a new architect or client is introduced. Performance may then decline until the system is updated with representative examples.

For this reason, model maintenance and retraining should be treated as part of an ongoing operational process rather than as a one time implementation activity.

Another consideration is the workflow surrounding the model itself. Image preprocessing, resolution management and the way individual sheets are prepared before analysis can significantly affect results. Selecting a strong model is important, but model selection alone does not determine the performance of the overall system.

Identifying similar constraints

The broader lesson is applicable beyond manufacturing.

An organisation can begin by identifying processes where output depends on how quickly a relatively small team can review documents. These constraints may not be visible in standard operational reporting because they are often considered part of professional or administrative work rather than part of the production process.

Measuring the activity can provide a useful starting point. Recording the time required for representative tasks over a period of time can help distinguish work that genuinely requires expert judgement from work that consists primarily of searching, extracting and counting information.

Where repetitive document processing is the primary constraint, automation may now be a practical option. Examples of manufacturing deployments using these approaches demonstrate how computer vision and language based systems can be integrated with operational workflows involving real project documents.

The technology itself is only one part of the opportunity. The first step is recognising that a document based activity may be functioning as an operational bottleneck.

Manufacturing teams have long used measurement to identify constraints on the production floor. Applying the same approach to information intensive processes can reveal constraints that have previously remained outside the usual operational dashboards.