Artificial intelligence – From Strategy to Practice

How To Implement Sophisticated AI Projects

Artificial intelligence (AI) is reshaping the value chain across many industries. While automation has been a major innovation topic for many years, the focus is now increasingly on autonomy. This means that complex systems and processes no longer follow a limited set of deterministic rules but, among other things, “self-learn” to adapt to different situations. However, only a few companies in Germany and Austria know how to exploit this enormous innovation potential for themselves.

For companies that have had little contact with artificial intelligence so far, it should be said that AI must become an integral part of almost all digital products and software solutions by 2025 at the latest, and will thus become an important asset and driver for digital value creation.

But how are successful AI strategies planned and put into action?

My colleague Florian Eisermann took an in-depth look at a successful AI strategy. In his article “8 points that no AI strategy should be without”, he explains the building blocks of a successful AI strategy for the years 2022-2025 – Success Factors, Technology, Skill and Scale are the keywords here.

This article builds on the previous in terms of content and is intended to clarify the question, particularly for IT, digital and data managers, of how ambitious AI projects can be implemented in practice beyond the strategy.

Potential Value – The Key Hygiene Factors of Successful AI Project 

Before companies want to implement the introduction of AI into production on a large scale, or implement large-scale AI systems, a major learning process about their own company with all its facets and processes is upstream. Therefore, corporate decision-makers should be aware of the following hygiene factors when starting the implementation of AI projects:

Start small – do not seek to immediately network and automate the entire company, but rather pick out a case where there could be high potential through automation and autonomy.

Nothing is perfect – It takes a while to train and optimize algorithms and AI-based systems to achieve the desired results – successful AI takes time!

Know the varieties of AI – Speech and Text Processing = Natural Language Processing (NLP), Image and Video Processing = Computer Vision (CV), Data Analysis = Time Series Analytics, Self-learning Strategies = Reinforcement Learning

AI in practice is interdisciplinary – Successful implementation often requires not only data science expertise but also experience in solution design and operations and, of course, the technical and process know-how!

AI Value and Impact – Economic efficiency of AI projects can be calculated well! And the “RoA” (Return-of-AI) is an important factor especially with regard to sustainability, climate, and species protection.

In order to be able to take these hygiene factors into account and, to a certain extent, prevent the failure of AI projects in the initial phase, it is particularly important to obtain clarity regarding the expectations within the company with regard to AI, which team of AI experts should work on the projects and what data quality is available. For this purpose, the following checklist serves as an introduction to the implementation:


  • Do all stakeholders have a realistic idea of the big picture?
  • Are the executives committed to the project?
  • Are the technical and business experts sufficiently aligned?
  • Are the project risks clear to all decision makers?
  • Is the planned budget large enough to give AI technology a fair shot?
  • Is the realistic effort small enough for a positive ROI?
  • Is it clear to everyone by what standards project success will be evaluated and what is “good enough”?


  • Does your company have sufficient expertise within the team?
  • Does your company have sufficient AI expertise within the team?
  • Is the nature of collaboration between AI development and traditional software development clear?
  • Does your company follow a professionalized development process?
  • Has your company involved the subsequent users of the AI system and addressed potential concerns about the AI transformation?

Data Quality:

  • Does your company have any data at all?
  • Is the existing data sufficient to answer relevant questions of the AI project?
  • Are the data sets representative and do they cover all relevant degrees of freedom and possible variations?
  • Is the quality of the data sufficient?
  • Is the quantity of data sufficient?
  • Do you have a data strategy that governs how you obtain, manage, and maintain your data?
  • Can you obtain data annotations with reasonable effort?

From the idea to the use case 

After possible directions have been explored, the next step is to define the use case from an idea and, beyond that, to move from a PoC (proof of concept) to a market-ready product.

AI - from strategy to practice - EN

To ensure a smooth course of an AI project, the basis is an initial project phase in which, in particular, a uniform vision and mission should be defined with all participants. In addition, regular coordination meetings should be held in a defined period of a maximum of four to six weeks in an initial phase in order to discuss the planned project progress, goals and non-goals on the one hand, but above all to define any work packages on the other. Overall, progress should be made with the highest possible transparency and simultaneous ownership. As part of the “co-creation” approach, all stakeholders should be involved in shaping the strategic and technological decisions.

Nevertheless, defined expert teams should also be able to make decisions independently within the scope of their capabilities and authority, which are in the best interests of the project and do not require separate clarification:

  • Concretization of the work packages
  • Definition of cooperation and role allocation
  • Interfaces and integration concept
  • Schedule of work packages and development

A brief overview of relevant AI use cases in the industry

The majority of companies today rely on AI functionalities aimed at classic process optimization, for example by networking equipment in production. In this way, savings can be achieved in the processing and analysis steps. For example, by means of so-called asset efficiency analyses – here, it is primarily a matter of monitoring the plant and production lines within a company. Asset efficiency analyses enable the easy location and traceability of important assets, also along the supply chain (e.g. raw materials, end products and containers), in order to optimize logistics, maintain inventory levels, avoid quality problems and detect theft.

But machine learning applications are also being used for digital products. Here, connected vehicles in particular are very much in vogue. In the simplest sense, these are computerized vehicles that automate many normal driving tasks – in some cases even driving themselves. Current systems scan lane lines as one of several detection methods. This is an important advance in the context of smart mobility. There are a variety of benefits implied by self-driving cars. These include accident reduction, as the autonomous driving car can act faster than a human. In addition, these cars are equipped with cameras, radar and lasers as sensors that feed information into the differential GPS. Computer vision applications allow the cars to “see” and process what is in the environment. The radar lets the vehicle see up to 100 meters away in darkness, rain or snow. The lasers continuously scan the world around the car, providing the vehicle with a continuous, omnidirectional 3D view of its surroundings.

In addition, a large number of companies already have autonomous robots along the process chain. The machines or the computers are usually based on artificial intelligence and make use of cognitive abilities that resemble human behavior. Robotics is also frequently used in manufacturing. Companies also use machine learning to optimize the automation of their maintenance and service operations.

From the use case to concretization

After completion of the initial project phase, the main task is to demonstrate the technical feasibility of the outlined use cases and the objective of a selected use case – thus an “ongoing” design and requirements phase is indispensable.

The purpose of such a requirement analysis is above all to form a foundation for future decisions. For this purpose, the view of users and customers should be considered in addition to a project position in order to be able to formulate requirements of machine learning applications in a practical way. However, the requirement analysis is not only needed for the pure formulation of the requirements. It is also important to find out whether the established requirements (and ultimately the entire project) are economically viable in addition to being technically feasible.

The results are used in the creation of the subsequent software implementation and the start of a proof-of-concept implementation phase to enable more detailed planning:

  • Validation of data, goals and results
  • Definition of epics
  • Transfer to software backlog
  • Prioritization of epics

A good start is half the journey

A good preparation and definition are helpful in order to concentrate later on the clean implementation. The focus is on being able to react to changes in an agile manner, to train the models cleanly, but also to question the established plan from time to time or to adapt it “ongoing” and in close coordination. Validation of goals is important, as are clear test criteria and integration into the surrounding systems and the IT landscape.

However, currently around 85% of AI projects “get bogged down” in the PoC stage and miss an enterprise-wide rollout. This is mainly due to the fact that no clear goals and defined processes have been established and often no common goal is being pursued. In addition, AI projects are running behind expectations that are too high – continuing with a lack of specialists, silo thinking, and even rejection within internal structures, which also lead to possible failure. Only if these challenges are all overcome can positive contributions to value and innovation be expected.

A successful initial project phase and, in the next step, a successful requirement phase are therefore the key to success for subsequent technical implementation – a good start is half the work.


Do you have any topics to discuss with our experts?

Get in touch

Get in touch