English | EN 
The question isn’t whether to implement artificial intelligence (AI), but rather when and how to begin. According to research by Strand Partners, the adoption of AI is currently hindered mainly by the lack of qualified personnel. Only eight percent of businesses are able to find the right people[1].
This means that the economic potential of AI is not being fully reaped. AI is accompanied by investments that can only be repaid if it is used in the right way. Therefore, be careful to avoid the following mistakes when starting with AI.
Thinking after the fact
When developers start working on a project, a good idea may pop up. What initially seems interesting often turns out to be an innovative technical achievement without knowing whether it solves a real problem. In such projects, many assumptions are made without involving colleagues from the business, legal, or even end users. Don’t just focus on technical KPIs that measure the success of the project. KPIs should also provide insight into the economic value for the organization. Therefore, a clear strategy from the start is crucial, with colleagues from other disciplines actively involved.
Lack of adoption plan
The development of AI requires a clear and successful adoption plan. Otherwise, the use of AI is destined to fail. Is there a change management plan in place when the way of working for end users changes, for example, and what is needed to successfully use the AI app and in what way? IT experts sometimes fail to consider how AI can be embedded in existing processes or procedures. If there is no contact with end users, even a well-performing AI app can lead to adoption problems.
Data infrastructure and quality not ready for AI
The principle of ’garbage in, garbage out’ certainly applies to AI. However, the lack of good data infrastructure and data quality is a common mistake that prevents the success of AI. Therefore, a well-established data infrastructure increases the chances of success of AI projects. In general, the IT maturity of the organization is a determining factor. The use of the best-performing AI models also often requires GPUs. Also, critically evaluate the existing cloud infrastructure. The lack of a (public) cloud infrastructure can hinder the success of AI projects.
Betting on one proof of concept
Success is not a given in AI projects. Therefore, do not bet on one horse to prevent disappointment and demotivation to continue. Identify a use case that has everything it needs to succeed, while at the same time not raising high expectations. Therefore, manage expectations with realistic goals.
AI gets a special treatment
Often, a special innovation process is set up for the development of AI systems. This can result in AI applications being developed without applying quality, testing, and safety standards that are embedded in regular application development processes. The lack of AI testing expertise is often the cause. For example, with specific large language models, it is difficult to test whether questions are answered correctly, simply because answers can vary. A multidisciplinary team offers a solution. It includes experts who know how to use end-to-end tests in combination with the knowledge of AI experts about the validation of non-deterministic answers. Therefore, ensure that AI systems are developed with the same quality engineering mindset as regular IT applications by testing both the AI model and the system.
Performance over explainability
In the academic world, the accuracy of AI models is often the most important benchmark. This does not apply to the business world, where explainability is key. If the AI solution makes more than 95 percent good predictings without knowing how and why they were made, then the adoption process becomes a difficult story.
[1] https://www.businessinsider.nl/nederland-ai-bedrijfsleven-groei-economie-2030/
