32 AI IN THE OPERATIONAL MANAGEMENT OF LARGE-SCALE IT SYSTEMS
organisation, as well as an important step towards a more environmentally sustainable organisation.
Over the last decade we have observed a very rapid pa ce of advancement of AI in terms of both theoretical
developments and applications and it is unlikely to slow down any time soon. This rapid development as well
as the pervasiveness of AI as a new technology also means that it has already profoundly affected organisations
and societies, with implications for many of our spheres of life. Like many new technologies, regulation is
lagging behind and therefore often cannot address many of the pressing concerns, inclu ding the ethical
implications of AI, which are much cited in particular when AI is used for automated decision making. These
concerns are valid and need to be appropriately addressed. However, they should be addressed in a way that
does not hinder the development of AI capabilities in the EU. It is therefore paramount to adopt a strategic
approach to regulating AI in the EU. This appro ach should allow for sufficient flexibility for the development of
new technology and its application across the economy, while at the same tim e ensuring that whenever AI is
deployed where it may have direct implications for human beings, it must be transparent, auditable and
compliant with all applicable legal and ethical requirements.
For eu-LISA, as for any other organisation providing IT services implementation of AI is not a question of if
but w hen and to what extent The EES and ETIAS both foresee a certain level of artificial intelligence or
automation and will therefor e have an immediate effect on individuals. Considering that the precision of AI
systems based on machine learning algorithms depends to a significant extent on data sets used for training,
and the quality of data sets used for the training of biometric recognition systems to be used in the EES will to
a significant extent determine the quality of the AI systems. With this in mind, a careful evaluation of the trade-
off between privacy prote ction and system performance will need to be made, in particular focusing on the
benefits and disadvantages of using real data sets for ML system training.
There are, however, many applications where AI can be deployed without any implications for human beings.
This report has therefore largely focused on applications where ethical and legal c onsiderations are either not
relevant or are secondary; namely on those applications where AI can improve the performance of eu-LISA
either by improving the quality of service (e.g. chatbots or virtual as sistants) or the performance of the
infrastructure (e.g. automated failure detection), or by reducing costs and improving performance by taking
over some of the repetitive routine tasks (e.g. AI on IT service desks).
Although AI has great transformative potential, it does not come without challenges and risks. Therefore,
organisations considering the depl oyment of AI in either internal operations or services should take th e
following factors in consideration:
Development of machine -learning based systems will require large data sets in order to train the ML
algorithm. In order to make sure that the AI system performs with a relatively low le vel of errors, the
training data sets need to be of very high quality. Creation of high quality data sets will require significant
Development and deployme nt of AI based on machine learning (in particular deep learning neural
networks) algorithms require substantial computational resources, which are normally not available within
one organisation. Therefore, the development and deployment of AI external service providers, such as
Amazon Web Services or Microsoft Azure, or shared infrastructure, such as the EuroHPC, will be necessary.
AI systems both rules-based and ML-based are often not static, as they operate in a continuously
changing environment. In order to take into account the changes in the environment, AI systems must be
continuously update d to r emain accurate. This will require either updating the code of the r ules-based
systems, or retraining the ML algorithms with new da ta sets or using the more recent continuous training
approaches. This has direct implications for the procurement of such systems, as well as creating a