One of the biggest unsolved challenges in manufacturing is not automation, AI implementation, or even supply chain management. It is something much more fundamental: how to capture and transfer operational know-how.
When experienced workers retire, change jobs, or are temporarily absent, a large share of their knowledge disappears with them. Newcomers or stand-ins often have to develop their own know-how through trial and error. Small variations in work erode productivity.
Manufacturers seek to reduce this problem in several ways, including training workers, guiding them during work, or automating the task. For any of these tactics, they first need to describe how to carry out the tasks. The classic approach has been to develop work instructions.1
The need for better work instructions
Leaders agree that clear work instructions are essential for stable processes. They are needed for training new employees, checking that tasks are performed consistently, and providing a foundation for continuous improvement.
Yet not all processes are properly documented in factories. Even in industries like automotive, which have obsessed over work instructions, not every process step is described. Many workstations still rely heavily on verbal explanations or informal coaching.
This gap is not due to a lack of awareness of the importance of documentation. Rather, it reflects the reality of how work instructions are typically created. A skilled eye needs to observe the process, take photos, write descriptions, and assemble everything into slides or documents. The process is time-consuming, and updating the instructions later requires repeating much of the effort.
Because the task is effortful and without prestige, documentation often falls behind operational reality. Many steps remain undocumented, and organizations continue to rely heavily on tacit knowledge that is shared informally by experienced operators. These challenges are particularly visible in environments with complex manual assembly or in high-variety, low-volume factories.
AI-assisted work instruction in Stadler
One example of high-variety, low-volume manufacturing is the train-building industry. Consider the Swiss rolling stock company Stadler. At its U.S. manufacturing site in Salt Lake City, Utah, Stadler assembles trains using mostly manual operations. Tasks range from relatively simple installations to complex work assemblies that require experience and careful coordination. To meet customer requirements, almost every new train set has variations in assembly. Creating work instructions for every repeated assembly step can feel exhaustive, yet it helps ensure high quality and productivity. What if this task could be (semi-)automated?
Stadler worked with Rimon Technologies to pilot a robust AI tool for generating work instructions. It works as follows. A skilled operator performs the task wearing lightweight recording equipment, including a 360-degree bodycam and a microphone. While performing the work, the operator simply explains the key steps they are doing in their own language.
The recording is then processed by AI software that helps convert the video and narration into a structured work instruction. First-person-view photos can be extracted from the footage, the spoken explanation automatically turned into written descriptions, and the steps organized into a clear sequence. A human-in-the-loop can cross-check, edit, and approve the work instructions. The result can then be deployed physically or digitally and easily accessed via QR codes at the workstation.
In a first pilot test led by ETH student Andrey Struffi, Stadler documented a conservative 50% slash in the time required to create the work instruction compared to the traditional manual method.2 The potential is much higher. Yet, technology capabilities do not equal technology use. Stadler, like every other company, must build routines and organizational ownership to take advantage of what such technologies can do.
It is not only Rimon Technologies that builds such solutions. Offerings range from simple video capture with YouTube-style how-to databases to sophisticated movement capture and work simulations. Some known vendors include Hapster, Poka, Augmentir, Tulip, Dozuki, Redzone, Parsable, and Augmented Industries, among others. Such companies display the role of AI in connected workforce software, including work instructions.
Help people succeed at work
The most interesting implication is not simply that documentation can become faster. The real opportunity is that more processes can finally be documented. If creating work instructions becomes easier, companies can capture and update much more of their operational knowledge. Tasks that previously remained undocumented due to time constraints can now be documented and structured.
Over time, this can make organizations less dependent on individual experts and more capable of systematically transferring knowledge. In a period where many experienced workers are retiring in manufacturing, manufacturers need to train new employees quickly.
While we wait for humanoid AI agents (we may have to wait a while), AI-assisted tools will not replace skilled operators or process experts when documenting work instructions. Instead, they lower one of the most persistent barriers in factory management: Turning practical shop-floor experience into clear work instructions.
- This post avoids a discussion on semantic differences between Work Standards, Standardized Work, Standard Operating Procedures (SOP), Work Instructions, One-point-lessons (OPLs), and operating instructions for the simple reason that companies use some of these terms interchangeably, and because the discussion already exists: See, Baudin, M. (2013) Perspectives on Standard Work, Blog post; Roser, C. (2024) All About Work Standards. 235 pages: AllAboutLean.com Publishing; Baudin, M. & Netland, T. (2022) Introduction to Manufacturing, Routledge ↩︎
- Struffi, A. (2025) AI-driven Solutions for Automating Work Instructions, Master’s Thesis, Chair of Production and Operations Management, D-MTEC, ETH Zurich (unpublished). ↩︎
What's on your mind?