Artificial intelligence: successfully scaling proof of concepts
From the AI PoC to production maturity
According to Gartner, 80% of AI projects fail after the proof of concept. Find out how to avoid the PoC
Technological advances have brought companies to the point where digital processes are no longer just an option, but a necessity. In particular, the use of artificial intelligence plays a central role for companies, as it significantly supports flexibility and efficiency in scaling processes.
A recently published study (German) on the current mood of companies around the topic of ‘cloud computing’ once again clearly highlights how company representatives rate technologies and in which areas they want to invest more in the coming years. It is hardly surprising that the number of respondents who see AI as a central investment and transformation topic in 2024 has increased again.
According to the study, conducted by analyst firm techConsult, 56% stated that AI has high strategic relevance for them in the company, closely followed by cloud computing with 47%. The two topics are therefore closely interrelated.
AI paves the way for faster development and delivery of digital products and seamless customer experiences, and this realisation has now made it to the boardroom.
This change is imperative, as customers expect ever better digital customer experiences and the increasing shortage of skilled workers is forcing more and more companies to rethink their approach. Internal cost structures must be efficiently optimised while simultaneously increasing the service level for customers. AI can help here.
More and more companies are therefore moving towards transferring individual processes and applications into a proof of concept (PoC), particularly in the field of artificial intelligence, in order to validate their feasibility and efficiency gains.
Although the use of artificial intelligence holds enormous potential, the vast majority of PoCs fail after their implementation and do not make it into production.
Index
7 reasons why AI proofs of concept fail
But why is that and how can it be changed? Essentially, seven core aspects can be identified that lead to the failure of an AI proof of concept:
1. Unrealistic expectations
Measurable, usable results are expected immediately, without taking into account that collecting and processing high-quality data takes time.
It is therefore advisable to identify possible business cases at the outset and to carefully distinguish promising PoCs from those that only work well in an ‘innovation lab’. Questions such as ‘How much time do I want to spend on this project?, How much budget do I have for it?, Do I have the necessary technical resources and know-how in-house? Do I need external advice?’ etc. should be clarified in advance to avoid nasty surprises at the end.
2. Insufficient data
One of the most common reasons for failure right at the start is the lack of sufficient high-quality, segmented and labelled data to train the AI model. Data treasures remain in departmental silos and never come into play.
While a PoC usually only requires a subset of data to validate its feasibility and can work on a single Phyton notebook, scaling it often requires large, cleaned and reliable data volumes in a structured form. If these are not available, you can also use synthetically created data, depending on the business case. However, the question of data must be asked early enough in any case.
3. The wrong model
AI models have been springing up like mushrooms long before the hype surrounding generative AI. The models, their performance and their areas of application differ considerably. If you don’t know the basic concepts and differences between the models, you’ll quickly end up with the wrong solution.
It therefore makes sense to take a close look at the available models before they are deployed. A basic understanding of the models also facilitates ‘make or buy’ decisions, because not every standard solution fits every case, and there is no need to reinvent the wheel if something already on the market has proven itself.
4. Isolated teams
AI is not a technology that is detached from the rest of IT and should be seen as the sole saviour. A team that is isolated from the rest of the company does not have the necessary information to successfully implement projects and scale the PoC. Therefore, it must be checked in advance whether the PoCs can be integrated into the existing IT infrastructure and which teams need to work closely together to do so.
5. AI is not made a top priority
A lack of understanding of AI and a failure to integrate it into the overarching corporate strategy are a recipe for failure. If you don’t make artificial intelligence a top priority and take the entire company with you through the change process that goes with it, you shouldn’t start at all. AI is one of the most transformative technologies of our century, bringing changes to our daily decision-making processes. These changes must be embraced by management as well as by each individual employee. For this to succeed, management must not only be a role model, but also guide, accompany and support their entire team through these processes with sensitivity and understanding. Continuous communication, transparency, training opportunities and enablement are essential factors for the acceptance of AI in the company.
6. Setting up PoCs without the relevant department
‘The relevant departments are unfamiliar with technology and are not even involved in setting up the PoCs.’ Those who think this way overlook something essential: the relevant departments are the keepers of valuable information about their domain. They have accumulated years of subject-specific knowledge that must be incorporated into the design of the PoCs. Otherwise, in the worst case, optimisation will miss the mark. Frustration is inevitable here. Therefore, the following applies: Talk to the relevant departments from the outset and involve them in the planning and design to ensure a good and usable result.
7. ‘We can do it alone’
Production-ready AI solutions require not only good data and AI models but also solid enterprise integration and governance readiness. This is where classic software engineering comes into play, including related topics such as enterprise architecture, EAM, consolidated development processes, data protection and security, and, above all, solid continuous and automated quality assurance. These aspects are not the focus and area of expertise of the AI PoC and those responsible for it.
Taking these core aspects into account when developing the PoCs significantly increases your chances of scaling and transferring the solution to production and brings you a step closer to the goal of successfully implementing artificial intelligence.
Comments are closed.