# AI in Manufacturing and Maintenance: Use Cases and Implementation

Author: Marin Niehues, Niehues Consulting · Published: 2025-08-04 · Updated: 2026-11-16 · AI Adoption & Operations
Source: https://www.niehuesconsulting.com/blog/ai-in-manufacturing-and-maintenance/

AI pays off in manufacturing and maintenance where machine and process data already exist and a concrete decision on the shift depends on them. The most widespread applications are predictive maintenance and camera-based quality inspection. Many plants test such applications on individual assets, yet the value only materializes once maintenance and production adopt the results in their daily routines. This article sorts the fields of application, shows the documented effect and a calculation method, and describes what an organization must settle to reach stable operations.

## Four fields of application with demonstrable value

Applications on the shop floor can be ordered by the decision they support. This also determines which metric measures success.

- Predictive maintenance: models detect deviations from an asset's normal behavior in sensor and operating data and report them before a failure occurs. The value shows in fewer unplanned stoppages and in interventions that fall into maintenance windows. The metrics are availability as a factor of overall equipment effectiveness (OEE) and the mean time between failures.
- Visual quality inspection: camera-based models check surfaces, dimensions or assembly states on the line. They learn from images of good and defective parts and flag deviations that an inspector confirms or rejects. The value lies in scrap detected earlier and in an inspection that works the same way on every shift. The metric is the rate of scrap and rework.
- Knowledge access for maintenance technicians: language models with access to manuals and fault reports answer questions at the machine. The value lies in faster troubleshooting, especially for faults that occur rarely. The metric is the mean time to repair.
- Planning and scheduling: forecasts for demand and sequencing support production control, as does the planning of spare parts. The value shows in delivery reliability and tied-up capital.

**Value arises at the point where a decision on the shift turns out differently. A model without a link to a work order and an accountable person remains an analysis.**

## What is the difference between preventive and predictive maintenance?

Preventive maintenance services assets at fixed intervals, while predictive maintenance intervenes when data signal an impending failure. The European standard on maintenance terminology distinguishes several strategies, which usually coexist in one plant.

**Maintenance strategies at a glance**
- Corrective maintenance: repair after failure. Suitable for assets with low downtime costs and redundancy.
- Preventive maintenance at fixed intervals: servicing by time, operating hours or output. Plannable, but parts are often replaced before the end of their useful life.
- Condition-based maintenance: intervention when a measured value such as vibration or temperature crosses a threshold.
- Predictive maintenance: a model analyses the trend of several measurements, forecasts the failure and schedules the intervention into a maintenance window.

The strategy is chosen per asset. Predictive maintenance pays off for assets whose stoppage halts a line, whose failures announce themselves in measurements and whose interval maintenance is expensive. For all other assets, fixed intervals or simple thresholds remain the more economical choice.

Data requirements follow a clear gradation. A model detects deviations from normal behavior from operating data alone. It names the cause and the remaining useful life only once faults and interventions are documented. A plant without a clean fault history therefore starts with anomaly detection and builds up documentation in parallel.

## How much does predictive maintenance reduce downtime and costs?

A guide published by the US Department of Energy cites 35 to 45 percent less downtime and 25 to 30 percent lower maintenance costs for functioning predictive maintenance programs. Compared with preventive maintenance at fixed intervals, it puts the savings at 8 to 12 percent, and compared with purely reactive maintenance at 30 to 40 percent.

For planning, these values are an upper reference. They derive from industrial data that predate today's learning models and describe programs already in regular operation. A plant starting with its first asset therefore calculates with its own downtime data and proves the effect in the prototype. Above all, the range shows where the value lies: in avoided unplanned stoppages rather than in saved maintenance work.

**Effect of predictive maintenance**
35 to 45 % | less downtime in functioning programs
25 to 30 % | lower maintenance costs on industrial average
8 to 12 % | savings compared with preventive maintenance at fixed intervals
30 to 40 % | savings compared with reactive maintenance

## When does predictive maintenance pay off?

Predictive maintenance pays off when an asset's avoided downtime costs exceed the costs of sensors, data connection and ongoing model operation. The calculation works with a few figures every plant already knows.

**Business case calculation per asset**
- Step 1: determine unplanned downtime hours over the last twelve months from the maintenance history.
- Step 2: set the cost per hour of downtime, meaning lost contribution margin, overtime, express freight and scrap at restart.
- Step 3: estimate the avoidable share, only for failure types that announce themselves in measurements.
- Step 4: annual benefit equals downtime hours times cost per hour times avoidable share, plus saved interval maintenance.
- Step 5: subtract the annual costs: sensors and connection spread over their useful life, platform, model maintenance and the working time of the accountable role.

A worked example shows the order of magnitude. An asset with 60 unplanned downtime hours a year and costs of 5,000 euros per hour causes 300,000 euros in downtime costs. If a third of the failures are predictable, the benefit is around 100,000 euros a year. If the annual costs stay well below that, the asset is a candidate. For assets with few downtime hours or low consequential costs the calculation tips over, and interval maintenance remains more economical.

The cost of external support for analysis and prototype comes on top. External AI consulting in Germany is typically billed at around 1,000 to 2,500 euros per consultant day. How to set costs against benefits robustly is described in the article on the [AI business case](/blog/ai-business-case-benefits-and-costs/).

## Selection: four criteria for the sequence

A plant has more starting points than can be worked on at the same time. A robust sequence results from four questions that the business function and the executive team answer together. The approach follows the logic described in [prioritizing AI use cases](/blog/ai-use-case-prioritisation/), applied here to manufacturing.

**The four test questions**
- Decision relevance: which decision changes, and who takes it? If there is no answer, there is no use case.
- Data situation: are measurements available at the required resolution, and are faults and interventions documented traceably?
- Process link: where does the output go? A notification that triggers no maintenance order and no inspection instruction changes no process.
- Effect: which metric moves, and what is that movement worth for the asset or the line?

The test questions deliberately come before the choice of technology. Assets with high downtime costs and good documentation are the first candidates, because effect and data situation can be proven there at the same time.

## Why initiatives stall on the shop floor

The causes rarely lie in the model. They lie in the transitions between operational technology, IT and the business function. Machine data often sit in controllers that were never designed for analysis. Timestamps from different systems do not match. Fault reports contain free text that can only be analyzed with effort. Then there is the question of who monitors a model day to day when wear or the material supplier changes. How strongly the data basis determines success is shown in the article [Data quality: why it decides AI success](/blog/data-quality-for-ai-success/).

Acceptance at the machine adds to this. A system that produces false alarms loses the trust of maintenance technicians within a few weeks. The threshold for alerts therefore belongs to the introduction and is set together with the users. An alert must be rare and accurate enough to be taken seriously.

**Routes into application**
Criterion | Single pilot on one asset | Central data project | With Niehues Consulting
Time to first value | short | long | first value proven on one asset, with an expansion path
Link to operations | often open | initially open | settled from the start
Transferability to further assets | low | high, but late | built into the design
Ownership after the pilot | unclear | with IT | role and process defined

## From selection to regular operations

The route from prototype to operations follows three steps that repeat in every plant. The overarching framework for all business areas is set out in the article [From pilot to scale](/blog/from-pilot-to-scale-ai-operations/).

- Understand: the decision, the data flow and the metric are described for a specific asset or line. The result is a business case that expresses the value in availability or scrap.
- Build: a prototype runs in the real workflow, with shift supervisors and maintenance as users. It proves the value under the conditions that will apply later.
- Introduce: responsibilities and processes are defined, users are enabled, and operations pass to an internal or external team. The transition is complete only when a named person is accountable for model and metric.

The third step determines whether one asset becomes a standard for several sites.

## Niehues Consulting as a partner for manufacturing and maintenance

[Niehues Consulting](/) brings AI into operations from a delivery perspective. The work starts with business analysis and a business case for each application. A prototype follows that proves the value in a plant's real workflow, and matures until production and maintenance adopt it. Introduction includes clear responsibilities and enablement of users, as well as the transfer into operations with internal or external teams. Marin Niehues steers such initiatives as delivery manager and knows the requirements of large industrial organizations as well as those of growing companies.

The approach follows the Niehues Delivery Model with its phases of understanding, building and introducing. Three of its review questions matter most for initiatives on the shop floor. What is the baseline, meaning today's number of unplanned downtime hours, and how do we measure success? Does the prototype prove the value in the real workflow with real asset data? Who is accountable for running the model from day one? The entry point is a free initial consultation of 30 minutes, which can be booked via [contact](/#kontakt).

**What executives should take away**
Select applications by the decision they change and by the metric in which the value shows. Calculate predictive maintenance per asset with your own downtime data, and settle data access and the process link before the prototype starts. Name the accountable role so that a test on one asset becomes a method a plant uses reliably.

**Sources**
- Bundesverband Deutscher Unternehmensberatungen (BDU): Honorare in der Unternehmensberatung, fee study 2022
- Sullivan, Pugh, Melendez, Hunt: Operations and Maintenance Best Practices, Release 3.0, U.S. Department of Energy, Federal Energy Management Program, 2010

## Frequently asked questions

### What is predictive maintenance?

Predictive maintenance is a maintenance strategy in which a model detects an impending failure from sensor and operating data and schedules the intervention into a maintenance window. Unlike fixed intervals, maintenance takes place only when the data call for it. The value lies above all in avoided unplanned downtime.

### Where should AI in maintenance start?

AI in maintenance should start with assets whose failure causes high costs and whose faults are documented. There, data situation and effect can be checked at the same time, and a first benefit is measurable within a manageable period.

### What data does predictive maintenance need?

Predictive maintenance needs measurements from the asset at sufficient resolution, such as vibration, temperature or current draw. These operating data are enough to detect deviations. Diagnosing causes and forecasting remaining useful life additionally require a traceable history of faults and interventions.

### How does AI visual inspection with cameras work?

Cameras capture every part on the line, and a model compares the image with what it has learned from good and defective parts. Deviations such as scratches or missing components are flagged, and an inspector confirms or rejects the finding. Every confirmation enlarges the data basis, and inspection applies the same standard on every shift.

### What is a digital twin in manufacturing?

A digital twin is a digital representation of an asset or line that is continuously updated with its operating data. It allows changes to parameters or processes to be simulated before they are made on the real asset. For predictive maintenance it is helpful but not a prerequisite.

### How do you measure the value of AI in manufacturing?

The value of AI in manufacturing is measured through metrics the business function already tracks: availability and OEE, scrap rate, mean time to repair or delivery reliability. The business case sets baseline and target before the prototype, so that the comparison after deployment is unambiguous.

### Who is accountable for an AI model in the plant?

A named role in the business function, usually in maintenance or production management, is accountable for outcome and metric. IT and operational technology provide the platform and data access. This division is set down in writing before regular operations.
