Key Factors for the Success and Failure of AI Implementation

Artificial intelligence technologies are rapidly developing and, in theory, can offer unprecedented opportunities for automation, analytics optimization, and innovation. However, successful AI implementation requires addressing a number of complex challenges, deep understanding, strategic planning, and cross-functional collaboration. Correct AI development and implementation are impossible without considering technical, organizational, and financial aspects.

According to international research on AI implementation, over 80% of AI projects fail. This is twice the rate for the development and implementation of IT projects unrelated to AI/ML.

What is Successful Implementation?

First and foremost, it's solving specific user problems, not pushing half-baked software solutions that sometimes even complicate things. The reality is that, in the pursuit of leadership in the AI industry, companies are often launching completely immature products. This is often fueled by the desire to justify the financial costs of implementing a project, as admitting failure would cause colossal reputational damage. Haste and the desire to quickly meet expectations are the main reasons why AI solutions fail to deliver the desired results.

International experts identify the following reasons for AI project failure:

• Business representatives often don't understand or don't articulate the problem that needs to be solved with AI/ML. Incorrect preliminary analysis of business processes and poor communication between participants in an AI implementation project at the outset will likely lead to failure. The organization lacks sufficient data to adequately train AI models, or the data quality is very low, which is the result of inadequate analysis during the preliminary preparation stage of AI solution development.

• The company focuses more on using the latest technologies than on solving the real problems of its business users, contributing to the wrong solution choice.

• Organizations may lack the appropriate infrastructure and qualified staff to manage their data, develop, and deploy complete AI models.

• The technology is being applied to a problem that is too complex for artificial intelligence to solve. AI is not a magic wand that can make every complex problem disappear; in some cases, even the most advanced AI models cannot automate complex processes.

AI Project Planning: The Foundation for Successful Development

How can you successfully overcome these challenges? The first thing to understand is that business processes are not exactly the same across organizations. There are similar business processes with their own specifics, and these specifics are unique. Therefore, a solution often needs to be selected and developed from scratch, and there's no 100% guarantee that it will be found right the first time. However, there are standard architectural solutions that can reduce the time spent searching for the right tools, aid in their development, and simplify subsequent support.

Any AI model operates and trains on data from a unique business process and must be integrated into it. This is an ongoing, cyclical process: during implementation, the data fed to the AI model will begin to change, as the business process is influenced by an external factor related to the implementation of the model itself. The model hasn't "seen" the new data; it's trained on the old data. For the solution to work again, the model needs to be further trained, and this is an ongoing, cyclical process. Unfortunately, unlike standard IT development, this specific cycle of changes and implementations doesn't allow for a ready-made ML solution out of the box that doesn't require support and refinement. And it's necessary to prepare not only for consistently successful implementations, but also for financial costs, which can only be minimized through an iterative approach and proper planning of the entire AI/ML project. The specific nature of AI development precludes the full application of the conventional IT development cycle, requiring a different approach, which is beginning to emerge with the advent of MLOps concepts and tools.

Stages of AI Implementation: How to Minimize Risks

Stage 0: Idea

The "maturation" stage. At this stage, the problem and idea transition to the search for solutions and the recognition of the need for data collection. This is an important stage in the transition from the business objective to the technical implementation stage, essentially laying the groundwork.

The foundation of the entire solution is laid. This stage should result in a preliminary roadmap based on business process analysis and a hypothesis for pilot testing.

Stage 1: Research

This stage helps determine whether the data necessary for implementing an AI solution is available and whether it is suitable for successful implementation. During this stage, target hypotheses should be formulated that align with the business goal. Then, the search for the necessary AI models begins. This stage should result in an understanding of the company's readiness to implement this solution, whether such a solution is even possible, and whether further steps are needed, such as collecting additional data, setting up new business processes, etc.

After research, you can return to business process analysis again if the company's goals cannot be achieved with the available data. It is also possible that a ready-made AI tool for solving a specific problem does not yet exist on the market. Another common problem is poor data quality. Consequently, it is necessary to return to the previous stage to bring the infrastructure and data up to standard.

Stage 2: Prototyping

This is the stage of testing the target hypothesis, developing the base of your AI/ML model based on the previous analysis, with minimal effort required to visualize the model's results on a limited data set. This stage allows you to test whether the problem can be solved, how effectively it can be solved, and whether the investment in the AI/ML solution will pay off. If successful, the prototype helps understand the possible architectural solution and develop a plan for the project's MVP (Minimum Viable Product).

At this stage, it is possible to return to either Stage 0 or Stage 1. However, a common mistake is often made here. Either A/B testing of the solution is not conducted at all, or it is conducted incorrectly. As a result, a raw solution with absolutely no proven potential impact is sent into development.

Stage 3: Piloting

This is the stage where most AI projects fail due to analytical and evaluation errors, as they approach it either with the wrong business goal, unclear AI/ML solutions, or underestimated impacts and false expectations.

At this stage, the solution is implemented into the business process with the goal of testing it in practice. There's a high risk of false expectations of potential wow effects being dashed, as a significant portion of AI solutions are currently not evaluated for their potential impact; as a result, the solution may fail to deliver any results. However, all of this could have been avoided by minimizing costs through stages 0-2, eliminating unnecessary risks at each step through iterative work planning, and developing a methodology for mathematically assessing the potential effectiveness of the result.

Stage 4: Scaling, Adaptation, and Maintenance

Here, we ensure the resilient and effective operation of the AI solution in changing conditions after the initial implementation. The solution itself also begins to significantly influence these conditions, as it triggers the restructuring of many business processes and adaptation to new operating conditions. This stage should ensure that the AI solution will deliver long-term benefits to the organization and enable further development of its potential.

Without adequate support and methodological assistance for interacting with the new tool, AI can lead to a loss of user trust. Adaptation is an important stage in establishing new business processes and obtaining new data for model training; only then can scaling be achieved.

And please remember that consistently completing all stages of AI implementation in a company and carefully considering all the implementation details will help you achieve your business objectives.