Why It is Hard to Find AI in SMEs: A Survey from the Practice

and How to Promote It

Andreas Bunte

, Frank Richter

and Rosanna Diovisalvi

Institute for Industrial IT, Langenbruch 6, 32657 Lemgo, Germany

Swiss Global Investment Group AG, H

unenberg, Switzerland

Keywords:

Artiﬁcial Intelligence (AI), AutoML, SME, Manufacturing, Survey.

Abstract:

AI seems to be an important aspect of Industry 4.0, which was introduced about 10 years ago. The main results

of interviews about AI with 411 people from 68 companies have been summarized in this paper. Most of those

companies were SMEs. Main challenges for the application of AI have been identiﬁed. Concrete solutions

that can support the implementation and application of AI are presented. The need to adequately support AI

in SMEs is underlined and speciﬁed.

1 INTRODUCTION

The term Industry 4.0 shall promise a new era of in-

dustrial production (Kagermann et al., 2013). Pro-

duction plants should be more ﬂexible to react faster

to market demands, more user-friendly, and more pre-

dictable (Kagermann et al., 2013), e.g. through condi-

tion monitoring. A huge potential is predicted, which

is in a range from 10 to 35 % for additional produc-

tivity on conversion costs (R

ußmann et al., 2015),

whereas other studies identify savings even up to

70 % (Bauernhansl et al., 2016). This potential shall

partly rely on Artiﬁcial Intelligence (AI) (Wahlster,

2017).

Since the term Industry 4.0 is known for about 10

years (Wahlster, 2017), we want to review the state

of the utilization of the term AI in the industrial envi-

ronment. Do people exactly know what AI is in the

context of manufacturing? How far goes the imple-

mentation and success of AI in manufacturing? What

are the concrete beneﬁts of using AI? If the phrases

of huge potential are not just good for marketing pur-

poses, it is expectable that AI is well known in com-

panies.

The paper is structured as follows: At ﬁrst, we

summarized the key ﬁndings out of the conducted

interviews about AI in the production area compa-

nies, with a focus on Small- and Medium Enterprises

(SMEs). Secondly, we lined out requirements and

potential strategies to support the usage of AI in the

companies. Afterwards, two best practice solutions

are introduced. Finally, we draw a conclusion and

give an outlook for potential future steps. The contri-

bution of this paper is to uncover potential challenges

with regards to AI based on conducted interviews and

to offer concrete strategies to support AI.

2 CHALLENGES AND HURDLES

OF AI IN PRODUCTION

This article is based on conducted interviews with 411

people from 68 companies located in Germany, Aus-

tria, and Switzerland. The purpose of the interviews

was to primarily ﬁnd out, (a) how common the use of

AI in production is, (b) what effects employees recog-

nize and observe by the use of AI and (c) how the eco-

nomic effects of AI are being measured in industry.

Therefore, we interviewed experienced employees in

production that are decision-maker including foremen

and plant managers, from different industry sectors,

predominantly from machine and plant construction

companies. The revenues of the respective companies

are up to almost 250 million Euros per ﬁscal year. The

minimum revenue of the respective companies was 12

million Euros, each based on the ﬁscal year 2019. We

used an interview guideline in form of a structured

questionnaire including several open questions. The

latter seemed important and adequate for the present

case since we wanted to avoid leading people in a cer-

tain direction by preparing closed questions and there-

fore get an unintentional bias by limiting the number

of possible answers. The goal of the interviews was to

614

Bunte, A., Richter, F. and Diovisalvi, R.

Why It is Hard to Find AI in SMEs: A Survey from the Practice and How to Promote It.

DOI: 10.5220/0010204106140620

In Proceedings of the 13th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2021) - Volume 2, pages 614-620

ISBN: 978-989-758-484-8

get sound feedback from the practice. Therefore, the

interview partners and the companies were not chosen

representative, we did not follow a strict, narrow SME

deﬁnition, whereas most of the companies ﬁt into the

established ones. Additionally, we did not distinguish

the different industry sectors with regards to the pre-

sented results of our interviews, since there was no

signiﬁcant difference identiﬁed.

2.1 Basic Challenges with Regards to

Intelligence and AI

To thoroughly analyze the potential impact of using

AI it is – amongst others – ﬁrst of all, necessary to

understand what intelligence means and what it might

be or might not be. To begin with, AI is an expres-

sion, that is little concrete. Even though people some-

times love unspeciﬁc and generic expressions and un-

substantial but good sounding phrases, it is usually

not helpful for making a business successful. Further-

more, there is no common deﬁnition of what intelli-

gence is – at least up to now. Not having a common

deﬁnition of intelligence, how can we deﬁne artiﬁcial

intelligence? Dealing with catchwords, not really be-

ing clear about the concrete meaning, might not be

efﬁcient. Let’s consider human beings and the human

intelligence ﬁrst: What makes a human being intelli-

gent? If he or she is analytical in the way he or she is

thinking? If he or she can solve problems sufﬁciently

and in a timely manner? If he or she is creative? If

he or she is empathic? If he or she is able to draw

logical conclusions to problems? If the speed of pro-

cessing information is fast? The ability to concentrate

on issues and their solutions? The ability to learn new

things and to unlearn unnecessary or less successful

strategies? The ability to memorize things? All of it

might make sense, but is it the full truth? We don’t

know.

Let’s look at some of the attempts to deﬁne in-

telligence. The psychologist William Stern deﬁned

intelligence as the ability of an individuum to adjust

his or her thinking intentionally and goal-oriented to

new demands (Stern, 1912). Intelligence might be

understood as the common adaptability to new tasks

and conditions of life. The social psychologist Peter

Hofst

atter deﬁned intelligence as the ability to deter-

mine the structure and the ability to discover and re-

veal redundancies (Hofst

atter, 1966). The psycholo-

gist Robert Sternberg assumed, that intelligence com-

prises analytical, practical as well as skills related to

experience. In this context, intelligence is mainly re-

lated to the interaction of individuals with their envi-

ronment. The psychologist Linda Gottfredson says,

“intelligence is a very general mental capability that,

among other things, involves the ability to reason,

plan, solve problems, think abstractly, comprehend

complex ideas, learn quickly, and learn from expe-

rience. It is not merely book learning, a narrow aca-

demic skill, or test-taking smarts. Rather, it reﬂects a

broader and deeper capability for comprehending our

surroundings-’catching on’, ’making sense’ of things,

or ’ﬁguring out’ what to do.” (Gottfredson, 1994)

As you see, the existing deﬁnitions of the word

“intelligence” are often widely generic themselves,

but also different. As stated above, generic deﬁnitions

don’t really help to be successful in the long-run. Re-

gardless of all the different deﬁnitions of what intel-

ligence could mean, we can state justiﬁably so that

human intelligence is hard to squeeze in a certain def-

inition and that it probably comprises a broad range of

different aspects, such as learning experience, ability

to goal-oriented problem-solving, foresighted think-

ing and acting, but probably also moral aspects, em-

pathy, and creativity. As we saw before, those aspects

are hard to really specify and to sufﬁciently concretize

in a generally accepted deﬁnition. The human brain

is still one of the mysteries in our world. It is com-

plex and has been explored only very little compared

to its potential and ability. Hence, we might want to

be careful with using the word “intelligence”.

Let’s now move on to the expression AI. It might

be helpful to understand AI not as an attempt to copy

human intelligence or human behavior and try to op-

timize it. It might be more useful to understand AI

as the simple and reasonable attempt to further im-

prove processes with regards to effectiveness and ef-

ﬁciency by using “modern”, state-of-the-art technolo-

gies. In industry, we see lots of attempts to improve

processes under or disguise of AI. In fact, AI seems

to be nothing else than stated above: Improving pro-

cesses with regards to effectiveness and efﬁciency by

using state-of-the-art technologies, that are able to re-

act to changes in their environment within a certain

scope.

One of the results of the survey we conducted un-

derlines the issue with the expression “artiﬁcial intel-

ligence”. The answers are of a wide variety. Some

of the interviewees see AI simply as the ability of

machines to adequately react to certain events in a

faster way than human beings can do – whatever ’fast’

means. Others even consider chatbots as intelligent

systems. Obviously, these people had not much to do

with chatbots so far, otherwise, they knew, that chat-

bots are everything else but intelligent. Some inter-

viewees named machine learning (ML) as part of AI.

A majority of the latter were not able to name con-

crete new business models behind their attempt to use

ML in their respective company. Only a handful of in-

Why It is Hard to Find AI in SMEs: A Survey from the Practice and How to Promote It

615

terviewees stated, that AI has nothing to do with intel-

ligence, but can solve some problems more efﬁciently

and effectively than human beings can do.

2.2 Usage of AI in Manufacturing

We will now show a summary of selected results, ex-

plain them brieﬂy, and share some implications.

The ﬁrst noticeable fact based on the given an-

swers is, that only 4% of the interviewees stated

that they have personal experience with AI. Only 5%

stated that they are currently using AI in manufactur-

ing. On the other hand, 67% of the interviewees be-

lieve that AI has - respectively would have - a positive

economic impact (see Fig. 1).

In this context, it seems to be of interest to ﬁgure

out the reasons, why so little companies are using AI

in manufacturing. 24% of the interviewees stated, that

their company is too small for AI. 5% stated that there

is not enough potential for improvement by using AI.

Since several companies have revenues between 30

and 50 million Euro, it is surprising that many inter-

viewees think their company is too small for using AI

or has not enough potential for improvement.

One of the challenges in particular in SMEs seems

to be a lack of expertise in the ﬁeld of AI. On the other

hand, there is more than one-ﬁfth that saying that their

company is currently evaluating the potential use of

AI. At least SMEs seem to be sensitized and try to ﬁg-

ure out if the implementation of AI in manufacturing

makes sense. Overall, it seems that SMEs are slowly

catching up but have still a long way to go due to a

lack of resources and experience in AI.

2.3 Beneﬁts of AI: Is It Measurable?

Most of us would hopefully agree, that it only makes

sense to implement and use new technologies and

methods, if the economic outcome can be objectively

measured and if it helps you staying competitive. To

our surprise, only 5 % of the interviewees

stated, that

they measured the economic impact of AI in their

company.

This is a critical point here seems to be the fact,

that only very few companies measure the economic

success of AI. But where is the sense in implementing

AI, not knowing if there is a concrete positive impact

and how big this impact is – in terms of savings or

increased proﬁts?

Let’s look deeper into the challenges of measur-

ing the success of AI in manufacturing by using pre-

dictive maintenance as an example. Predictive main-

Comprises only people working in companies, that im-

plemented AI solutions in manufacturing.

tenance is supposed to prevent unplanned downtime

of machines in manufacturing processes by using ma-

chine learning. The overall goal is to alert as to what

is needed to avoid stoppages of production in the fu-

ture.

Predicting the future seems to be a pretty am-

bitious goal. How many of you think that anyone

or any algorithm or any software solution is able to

predict the future? Generally, it is not possible, of

course. The more degrees of freedom a system has,

the worse is the predictability, because it is impossi-

ble to capture “all” data that have or might have an

inﬂuence. For example, the temperature, machine vi-

brations, humidity, quality of raw material, utilized

capacity, dust, the age of machines, and the quality of

spare parts. This list would be almost endless if you

tried to capture “all” data. Not using “all” data would

presume, that you know in advance, what data to col-

lect, and what data are unnecessary. Furthermore, the

more data are captured, the more data are needed for

a precise model. Since the system is dynamic and the

behavior might change over time, in real-world sys-

tems it is impossible to get enough observations to

learn a model in all stages. However, for speciﬁc sys-

tems, predictions can be made within a certain range,

as in (von Birgelen et al., 2018) the condition of a sin-

gle blade of a machine can be identiﬁed and a better

point in time for maintenance can be predicted, com-

pared to what humans can do. But this is not a free

ride for applying predictive maintenance. A holistic

evaluation has to be done for every use case.

Based on the interviews, it seems nothing else

than sort of a self-fulﬁlling prophecy that AI has a

positive economic impact. However, it is difﬁcult

to evaluate whether or not predictive maintenance

reduces machine downtimes and at the same time

saves cost for maintenance. Let us explain that a bit

more accurate: Overall, predictive maintenance usu-

ally reschedules the point of maintaining machines,

which might be shifted to an earlier point in time to

prevent machines from unintentional stoppages. If

you maintain a machine based on the suggestion of

an algorithm earlier then you would have done it in

the past by using regular maintenance schedules, how

can you say that this particular machine would have

really had an unexpected downtime without the use of

predictive maintenance? How can you deﬁnitely say

that maintenance based on predictive maintenance is

better than regular maintenance programs? How can

you deﬁnitely say that changing e.g. wear parts based

on suggestions of predictive maintenance solutions is

better than regular maintenance programs? And fur-

thermore, how can you be sure that it is more cost-

effective to rely on predictive maintenance? The an-

ICAART 2021 - 13th International Conference on Agents and Artiﬁcial Intelligence

616

Figure 1: Results of the interviews.

swer is: In many cases, you simply can’t tell. And that

might be one reason that most of the companies have

no calculations, or at least no proven about savings

or proﬁt increases with regards to predictive mainte-

nance.

One conclusion might be, that it is good to have

adequate support from AI, but that it is still helpful

to scrutinize the result and use your own (real) intel-

ligence, before making any decisions with regards to

maintenance in manufacturing. Not fully giving up

our position in the driver seat in favor of AI might be

a good advice in the long-run. So use the beneﬁts of

AI, but be at least a bit cautious and alert of how far

you (blindly) rely on AI in manufacturing.

3 SOLUTIONS TO SUPPORT AI

IN SMEs

In this section, we introduce strategies to support and

foster AI applications especially in SMEs. Therefore,

we identiﬁed three main aspects: Usage of the term

’AI’, easy application of AI, and concrete determina-

tion of its beneﬁt.

3.1 Differentiated Usage of the Term AI

As outlined in section 2, there is no common and gen-

erally accepted deﬁnition of the term ’AI’. Further-

more, AI and ML are often seen synonymous. Addi-

tionally, the terms are just frequently used for market-

ing purposes to promote a product or a whole com-

pany as being innovative and up-to-date. All this

makes it difﬁcult to use the term ’AI’ correctly and

in the right context, especially for people that do not

have deep knowledge about AI. Furthermore, many

people are not aware of the difference between use

cases (’the problems’ to be solved) and the solution

itself. AI is not the use case, it is just one possible so-

lution. But with the hype of AI, it seems that the use

case is not important. Important is that the solutions

include AI.

However, as far as the community is not able to

provide a common deﬁnition, it is not possible to al-

ways use the term in the right context. The contri-

bution of the scientiﬁc community should be a more

stable and functional deﬁnition of AI for the indus-

try sector. Nevertheless, if the deﬁnition is available,

the community has to apply it properly. This includes

companies that want to promote their products as well

as the research institutes that promote their work. Ev-

erybody in the community should clearly distinguish

between the ’underlying problem’ and ’solution’. If

the tackled problem has been solved, the quality of

the result is the crucial point, not if AI is part of the

solution.

3.2 Easy Application of AI

Hurdles for the application of AI are, according to the

interviews, the missing expertise, costs/amortization

time, and the size of the company, which might be

interpreted as missing infrastructure. Certain tools

for AI and especially machine learning, such as

R (Ghatak, 2017), the Python library TensorFlow

(Abadi et al., 2016), or ML.NET (Ahmed et al.,

2019), are available and basically relatively easy to

use. But these tools just provide algorithms. Some

important and time-consuming steps, such as data ac-

quisition, pre-processing, and the algorithm conﬁgu-

ration has to be implemented or adapted by the users.

Even if the algorithms are available, their application

requires a much deeper knowledge.

Some approaches try to automate machine learn-

ing (Auto-ML), e.g. TPot (Olson et al., 2016) or auto-

sklearn (Feurer et al., 2015). Such methods can be

helpful by selecting and optimizing the selected algo-

rithms and their conﬁguration, but there is still an ini-

tial effort to get the data and combine them with the

used methods. There are ﬁrst research projects that

Why It is Hard to Find AI in SMEs: A Survey from the Practice and How to Promote It

617

address these issues, such as ManuBrain (Burggr

af,

2020) or KOARCH (Bunte, 2020). In KOARCH, a

Cognitive Architecture for Artiﬁcial Intelligence in

Cyber-physical Production Systems (CAAI) has been

introduced (Fischbach et al., 2020). The CAAI does

not only automate the selection process of algorithms,

but it also selects the whole pipeline of data prepro-

cessing, modeling, and model usage. Furthermore,

if there is sufﬁcient knowledge about the respective

production plant, it is possible to automatically adapt

production plant parameters and create a closed-loop

application, e.g. for optimization. If the production

plant provides OPC UA server, the CAAI has simply

to be connected to these servers and a declarative aim

has to be selected, e.g. ’energy optimization’. Due

to the usage of virtualization, it is easy to set up the

system and prepare its application. A graphical user

interface can be integrated, which enables the usage

of CAAI without programming skills. Nevertheless,

the CAAI does not aim to replace AI experts. It is

seen as a reasonable extension since it is likely that

AI experts might ﬁnd better solutions. Nevertheless,

such approaches are a good starting point for SMEs,

which can be used to get in touch with AI technology

and evaluate the beneﬁts without unreasonable costs.

To conclude, SMEs do not need algorithms that

have a slightly better run time or outperform another

algorithm on an artiﬁcial data set. From their per-

spective, new algorithms only make sense if there

are problems identiﬁed that cannot be solved with

an existing one. SMEs do often not have the basics

for the application of AI, e.g. infrastructure and ex-

perts. They might need external support during the

ﬁrst steps that have to be enabled by solutions with

relatively low effort and low preconditions regarding

the expertise and infrastructure.

3.3 Determine Beneﬁt

Based on the results of the interviews, we could iden-

tify two groups with regards to the determination of

AI beneﬁts. The ﬁrst group is convinced of the beneﬁt

of AI, its simple presence is obviously enough to sat-

isfy them, since they have a strong belief in AI tech-

nologies. The other group is per se skeptical and does

not believe that AI can be beneﬁcial for their business.

Somehow or other, it is improvident not to evaluate

the concrete beneﬁts of AI. If you don’t measure ben-

eﬁts, how can you ’sell’ your solution as being suc-

cessful?

As mentioned above, it is not always easy to eval-

uate the beneﬁt. So, it is comprehensible that compa-

nies do not create costly studies before they integrate

AI into their machines, just to have a reference value

for evaluation. However, this is seen as a precondition

to evaluate the beneﬁt. As a result, it is mostly un-

known which processes can be supported by AI and

which are not predestined for it so far. It would create

a huge beneﬁt if evaluation data were available to con-

vert from try-and-error to a comprehensible decision

based on reliable knowledge and data.

From a scientiﬁc point of view, it is necessary

to have a broad knowledge base to evaluate the po-

tential beneﬁt of the usage of AI upfront. This re-

quires that especially companies, research institutes,

and universities need to work closely together, gather

relevant data, and publish the concrete improvements

being achieved with AI, but also the hurdles and chal-

lenges they are faced. But even if companies might

have such data available, they are mostly reluctant

to share them with a broader community because of

conﬁdentiality and with regard to their competitors.

Furthermore, the beneﬁt might be company-speciﬁc

and thus must be described in detail to gain leverage.

For example, optimization of a production module in

a sequential production line provides a large bene-

ﬁt if it is the bottleneck of the production line. The

same optimization in another company can be with

less or without beneﬁt if the module is not the bottle-

neck. Another hurdle might be, that people are not

willing to share failures with regards to implementing

AI solutions. But from a scientiﬁc perspective, it is

also important to know what doesn‘t work well and

especially why. One contribution to solve this issue

could be a clear guideline, how an evaluation should

be done, and how to balance the need for conﬁdential-

ity of gathered data and the need to share those data

with the AI community. To gain more efﬁciency and

effectiveness companies can cooperate e.g. with uni-

versities and set up joint research projects related to

concrete business cases. Those joint projects have a

good chance to be funded e.g. by governmental or-

ganizations. Furthermore, companies can also beneﬁt

from the knowledge of researchers and experts at uni-

versities. In the end, it is a win-win situation.

4 BEST PRACTICES

There are several examples of the successful AI ap-

plications in industries, that we introduce in this sec-

tion. We chose anomaly detection and optimization

as best practice examples. Besides these areas of ap-

plication, there are many more use cases that can be

solved with AI, such as decision support, extending

the pay per use to pay per stress, which considers the

wear of a machine, or use AI for the generation of test

cases. All presented use cases have been implemented

ICAART 2021 - 13th International Conference on Agents and Artiﬁcial Intelligence

618

by members of the authors’ institutions.

4.1 Use Case 1: Anomaly Detection

Anomaly detection is a good ﬁrst use case to apply

AI. The goal of anomaly detection is to observe the

plants’ behavior and detect changes and variations in

that behavior. However, the results are usually shown

to the operators, but no further processing steps, such

as scheduling maintenance or diagnose the root cause

of the anomaly are performed. This further process-

ing requires much additional effort for e.g. data label-

ing and modeling. Experienced operators can be well

supported by an anomaly detection algorithm, since

it is sometimes not obvious to identify the assignable

cause of an anomaly. By knowing the anomalous sig-

nal, operators can mostly identify the cause of the

anomaly by themselves. Overall, anomaly detection

can support operators to identify undesired states of

the machine, detect performance drifts, and ensure the

product quality.

We implemented anomaly detection functionali-

ties to a production plant in the food industries. Since

food is a sensitive product, it is important to identify

malfunctions in production as early as possible to not

risk contamination. For each module in the produc-

tion plant, a model of the normal behavior was learned

by the application. Hybrid automaton were used for it,

as introduced in (Niggemann et al., 2012). The model

enables the differentiation between discrete, contin-

uous, and time failures and thus covers all relevant

issues. Since the underlying process contains fast-

changing signals, it was a challenge to acquire the

data with an adequate sampling rate. However, we

were able to monitor the production plant and identify

unexpected behavior, so that the operator can be ade-

quately supported. The used approach was perfectly

matching the companies’ need for additional observa-

tions of their plant to maximize operating safety. The

key for the project was a well-deﬁned use case, where

the beneﬁt of the AI technology can be veriﬁed eas-

ily, as well as a bunch of algorithms that could be eas-

ily tested to identify the most suitable one. However,

the easy application of AI, as described in section 3.2,

would have reduced the effort of experts signiﬁcantly.

4.2 Use Case 2: Optimization

Optimization is a broad ﬁeld. In this respective use

case, we focus on the optimization of industrial equip-

ment. Typically, optimization is used to improve

product quality, reduce resource consumption, or in-

crease throughput. In many cases, the optimization

has a direct impact on proﬁtability, which makes it

very interesting for companies.

In our speciﬁc example, the resources of an indus-

trial cleaning process were optimized. This process

needs water, energy, and an operator. Depending on

the local parameters, e.g. type of pipe, amount of dirt,

and type of dirt, the operator has many parameters

which he can adapt. Differences between operators

have been identiﬁed, some clean with more water and

high pressure only once, where others use less water

and less pressure but clean multiple times. The aim

of this project was to support the user to identify the

best matching parameter for the current cleaning task.

Since it is an individual process, the challenge was to

get reliable data, such as the size of the pipe or the

accumulation of dirt. Because these values are given

by the operator, the data contains uncertainties. To

deal with this, a Bayesian network was used to per-

form the optimization. Therefore, the basic structure

was modeled by experts and the probabilities between

the notes are learned from the collected data. Further-

more, a rule-based system was implemented to sup-

port the decision process. Overall, the system can

now support the operator by identifying the best pa-

rameters for the respective cleaning process. Details

of the solution can be found in (Shrestha and Nigge-

mann, 2014) and (Shrestha and Niggemann, 2015).

The key to this project was again a well-deﬁned use

case with a clearly planned beneﬁt. The Bayesian

network was created by experts. Automated methods

suggested in section 3.2 would probably not reach the

same good results.

5 CONCLUSION AND FUTURE

WORK

The main results of the 411 interviews are that there

are not many applications of AI in SMEs and the AI

expertise in SMEs is still low. The beneﬁts of AI are

not often really measured, so most people just blindly

trust that AI is beneﬁcial. To support the usage of

AI, the term should be deﬁned properly in the con-

text of automation systems, the usage of AI has to

be simpliﬁed, and the beneﬁt has to be measured, to

have a solid basis for decision-making. However, AI

in SMEs can not only be successful, it is also man-

ageable as shown by the two best practices. The best

practices also show that the suggested solutions can

help, but they are not panacea.

Due to the method of the interviews, there are

some limitations: The interviewees were mainly from

SMEs and working within production. The compa-

nies were predominantly machine and plant construc-

tion companies. So, the introduced results in section 2

Why It is Hard to Find AI in SMEs: A Survey from the Practice and How to Promote It

619

can be understood as a snapshot of the practice that

might differ in different industry sectors.

There are additional points that should be ad-

dressed in future work. Especially the deﬁnition of

AI and the measurement of its’ beneﬁt has to be done

by a broader community in a joint effort. But not only

the deﬁnition is important; everyone can start right

now to use the term AI more carefully and question it

as needed. To ease the usage of AI, it can be supported

through projects such as ManuBrain or KOARCH.

Even if these projects follow a general idea, they fo-

cus on speciﬁc aspects of the (semi-)automatic appli-

cation of AI.

Properly used, AI is beneﬁcial to companies and

can actively support problem solving in different ar-

eas of manufacturing.

ACKNOWLEDGEMENT

The work was supported by the German Federal Min-

istry of Education and Research (BMBF) under the

projects ”KOARCH” (funding code: 13FH007IA6).

REFERENCES

Abadi, M., Barham, P., and Chen et al., J. (2016). Tensor-

ﬂow: A system for large-scale machine learning. In

12th USENIX Symposium on Operating Systems De-

sign and Implementation (OSDI 16), pages 265–283,

Savannah, GA. USENIX Association.

Ahmed, Z., Amizadeh, S., and Bilenko et al., M. (2019).

Machine learning at microsoft with ML .net. CoRR,

abs/1905.05715.

Bauernhansl, T., Kr

uger, J., Reinhart, G., and Schuh, G.

(2016). WGP-Standpunkt Industrie 4.0. Technical re-

port, Wissenschaftliche Gesellschaft f

ur Produktion-

stechnik WGP e. V., Darmstadt.

Bunte, A. (2020). Koarch website.

Burggr

af, P. (2020). Manubrain website.

Feurer, M., Klein, A., Eggensperger, K., Springenberg, J.,

Blum, M., and Hutter, F. (2015). Efﬁcient and Robust

Automated Machine Learning. In Advances in Neu-

ral Information Processing Systems 28, pages 2962–

2970. Curran Associates, Inc.

Fischbach, A., Strohschein, J., Bunte, A., Stork, J.,

Faeskorn-Woyke, H., Moriz, N., and Bartz-Beielstein,

T. (2020). CAAI – A Cognitive Architecture to Intro-

duce Artiﬁcial Intelligence in Cyber-Physical Produc-

tion Systems. The International Journal of Advanced

Manufacturing Technology.

Ghatak, A. (2017). Introduction to Machine Learning,

pages 57–78. Springer Singapore, Singapore.

Gottfredson, L. S. (1994). Mainstream science on intelli-

gence: An editorial with 52 signatories, history, and

bibliography. Wall Street Journal.

Hofst

atter, P. (1966). Einf

uhrung in die Sozialpsychologie,

volume 4. Kr

oner Stuttgart.

Kagermann, H., Wahlster, W., and Helbig, J. (2013). Rec-

ommendations for Implementing the Strategic Ini-

tiative Industrie 4.0 – Securing the Future of Ger-

man Manufacturing Industry. Technical report, acat-

ech – National Academy of Science and Engineering,

unchen.

Niggemann, O., Stein, B., Voden

carevi

c, A., Maier, A., and

Kleine Buning, H. (2012). Learning Behavior Models

for Hybrid Timed Systems. In Twenty-Sixth Confer-

ence on Artiﬁcial Intelligence (AAAI-12).

Olson, R. S., Bartley, N., Urbanowicz, R. J., and Moore,

J. H. (2016). Evaluation of a tree-based pipeline op-

timization tool for automating data science. In Pro-

ceedings of the Genetic and Evolutionary Computa-

tion Conference 2016, GECCO ’16, New York, NY,

USA. ACM.

ußmann, M., Lorenz, M., and et al. (2015). Industry 4.0:

The Future of Productivity and Growth in Manufac-

turing Industries. Technical report, The Boston Con-

sulting Group (BCG), M

unchen.

Shrestha, G. M. and Niggemann, O. (2014). A bayesian

predictive assistance system for resource optimization

— a case study in industrial cleaning process. In Pro-

ceedings of the 2014 IEEE Emerging Technology and

Factory Automation (ETFA), pages 1–6.

Shrestha, G. M. and Niggemann, O. (2015). Hybrid ap-

proach combining bayesian network and rule-based

systems for resource optimization in industrial clean-

ing processes. In 2015 IEEE 20th Conference on

Emerging Technologies Factory Automation (ETFA),

pages 1–4.

Stern, W. (1912). Die psychologischen Methoden der Intel-

ligenzpr

ufung und deren Anwendung an Schulkindern.

Kongress f

ur experimentelle Psychologie. Bericht.

J.A. Barth.

von Birgelen, A., Buratti, D., Mager, J., and Niggemann,

O. (2018). Self-organizing maps for anomaly local-

ization and predictive maintenance in cyber-physical

production systems. Procedia CIRP, 72:480 – 485.

51st CIRP Conference on Manufacturing Systems.

Wahlster, W. (2017). Artiﬁcial Intelligence for Industrie 4.0.

In Proceedings of the KI 2017: Advances in Artiﬁcial

Intelligence, 40th Annual German Conference on AI,

Heidelberg, Germany. Springer.

ICAART 2021 - 13th International Conference on Agents and Artiﬁcial Intelligence

620