At a glance
- Company
- Herges Stahl- und BlechbauFamily-owned maker of welded steel and sheet-metal constructions, mostly one-off
- Industry
- Manufacturing
- Where and when
- Germany. Published March 2026.
- Who produced the figures
- Official recordThe figures were measured and published by the Mittelstand-Digital Zentrum Saarbrücken, a publicly funded centre that ran the project with the firm. They cover one process, the rolling of sheet metal.
What they did
Herges is a family firm that has made welded steel and sheet-metal constructions for more than 80 years. Most of its products are one-off pieces built to a customer’s order. Each one needs a quote, and the quote depends on how long the work will take. A quote that is too low loses money. One that is too high loses the order.
Production times were estimated from experience. The firm had tried writing times down by hand, but the records for identical parts differed by up to 30 minutes and were of little use.
The project had two steps, and was limited to one process: rolling sheet metal.
- Record real times. The centre and the firm broke the process into steps and built a simple app for workers to log each one against the job. A prototype was adjusted step by step together with the workers, who were trained on it.
- Train a model on them. An AI model was trained to predict the time for a new job from the dimensions of the sheet. It was tested on jobs it had not seen.
The numbers
How far the estimate was from the real time, on average:
| Whole process, including moving material | Rolling only | |
|---|---|---|
| Experienced staff | 24.3 minutes (72.6%) | 8.6 minutes (39.0%) |
| AI model | 4.5 minutes (15.2%) | 3.8 minutes (16.5%) |
- About five times closer than the experts on the whole process.
- About twice as close on the rolling alone.
The centre’s explanation is that the experts overestimated the side tasks, such as moving material, on the basis of personal experience.
What happened next
The model was built into a tool for preparing quotes, which the firm is trying out. Staff enter the dimensions and get the expected worker and machine time. Three things were built in:
- Evidence beside each estimate. The tool shows the ten most similar past jobs with their measured times. Staff can see at once if the estimate looks wrong and change it.
- The firm can retrain it. Herges wanted to be able to train the model itself. New jobs keep being recorded, and the model is updated from them at regular intervals.
- Room to grow. The app can be set up for other steps, such as welding or assembly, without programming.
What the case shows
- Experienced people can be consistently wrong about time. Here they overestimated the parts of the job nobody watches closely.
- The data did not exist until someone collected it. The first half of the project was a way to measure.
- The comparison was made before the tool was trusted. AI and experts estimated the same jobs, and both were checked against what really happened.
- The tool shows its evidence. Similar past jobs sit next to every prediction, which lets a person check it in seconds.
What this says about the practices
Practices the company applied
- For each kind of task, test whether a person with AI beats either one alone Before the model went into use, its predictions were compared with the estimates of experienced staff on the same jobs. Do next · Effort: medium (weeks, a few people)
- Measure time saved per task against a baseline The firm first built a way to record real production times. Its earlier hand-kept records differed by up to 30 minutes for identical parts. Do first · Effort: low (days, one person)
- Keep a person checking anything that goes to a customer or carries legal weight With each prediction the tool shows the ten most similar past jobs and their measured times, so that staff can judge the estimate and change it. Do first · Effort: medium (weeks, a few people)
- Start with one costly process, not a company-wide rollout The project was limited to one process, rolling sheet metal, with a design that can be extended to welding and assembly. Do first · Effort: medium (weeks, a few people)
- Put your own records in order before you ask AI to predict from them The first step of the project was a simple app for workers to log real production times, because the hand-kept records differed by up to 30 minutes for identical parts. Do next · Effort: medium (weeks, a few people)
- Make the AI show what its answer is based on With each prediction the tool shows the ten most similar past jobs from the database, with their measured times. Do next · Effort: medium (weeks, a few people)
Sources
- Mittelstand-Digital Zentrum Saarbrücken (March 2026). KI-gestützte Datenanalyse zur Prozesszeitenermittlung. Project report, in German. digitalzentrum-saarbruecken.de