Today, the door to AI is open for all businesses. It is not a futuristic technology reserved for the big tech giants. It is a mature technology and it does not take much to develop a well-functioning Machine Learning model.
Therefore, one might ask why the majority of AI projects developed never leave the sandbox and remain PoCs or hobby projects that never create any real business value. In fact, surprisingly few companies in Denmark are able to use AI for what it actually is: an all-around tool in the data toolbox that can be used to optimize or improve most corners of your business - just like other common tools for creating value with data such as BI, RPA, IoT, and Master Data Management.
We see three primary reasons why AI remains a toy in the sandbox in many companies - rather than being a transformative technology that drives value across the business and differentiates a company from its competitors.
1. Lack of strategic and business anchoring of AI projects
We experience that several companies fall into the trap of starting AI projects or hiring data scientists without first uncovering what the technology should be used for. But unfortunately, AI does not create value in itself. Often, companies start with a pilot project in a less significant area to test the technology and later conclude that disproportionate resources were invested in creating limited value.
2. Lack of organizational ability to realize AI opportunities
Even for companies with a clear strategic and business anchoring, it can be difficult to realize identified business opportunities with AI. Typically, it is not the technical development of a well-functioning ML model that creates challenges. Rather, we often see that a lack of ability to work agilely and the management's lack of acceptance of uncertainty as a basic premise mean that companies do not manage to realize significant value with AI.
3. Lack of ability to deploy and operate AI solutions at scale
On the technical side, we often encounter companies that can develop well-functioning ML models but run into challenges when it comes to 'powering up' the models. Not to mention the deployment, ongoing operation, and further development of thousands of AI solutions, which places entirely different demands on both the solution, the underlying data architecture and infrastructure, and processes for Machine Learning Life Cycle Management or MLOps.
Below is Kapacity's recipe for addressing the above challenges so that your business can get the full potential of AI.
1. Think holistically and clarify your need for AI - every day
Strategic and business anchoring of AI initiatives is an important prerequisite for developing AI solutions that create value every day.
Therefore, it is necessary to develop the ability to identify opportunities in your company and to qualify whether a given problem or business opportunity can be realized through AI.
It is not possible for the newly hired data scientist or consultants alone to provide answers about where the business value of AI lies specifically for you. Instead, domain insight, concrete business needs from end-users, and a strategic direction from top management are needed to determine this. The task for your data scientists or consultants is to teach the business to understand what is possible and to qualify whether a challenge can be solved with AI and what it will require.
To get AI out of the sandbox, it is not enough to identify potentials and opportunities once. To achieve value with AI every day, there is a need for processes that continually involve relevant stakeholders in the identification, further development, and realization of business opportunities. In the end, it is also the end-users' use of an AI solution that determines whether the solution creates value or not.
2. Accept agility and uncertainty as fundamental premises when working with AI
When it comes to the project delivery model, it is important to accept that AI and agility go hand in hand. Therefore, it can quickly become a stumbling block for companies that want to know all the details of a project process before a final solution is delivered. This simply is not feasible.
One must be prepared for the fact that not all AI PoCs (Proof of Concept) will be successful. It is instead about "failing fast" - identifying sticking points and reducing uncertainty quickly. If there is no opportunity to obtain financing for agile sprints and PoCs - and therefore waiting for heavy budget or procurement processes, the company will never get out of the AI sandbox.
Therefore, there must be a management acceptance of uncertainty as a fundamental premise when working with AI. Although the technology is mature, a machine learning model is never perfect - it is, by definition, an approximation. The question is whether the model can be good enough to be valuable. This depends on how well an ML model can convert your data into an answer to your business problem - whether the question is 'which employees are at risk of quitting?', 'which products have errors?', 'how much should we dilute the wastewater to comply with threshold values?' or something completely different. The answer to these questions must be found iteratively in collaboration between Data Scientists and the business - and can never be found before a PoC. A PoC is precisely designed to use a minimum of resources to clarify whether a concept is likely to work. Then a pilot project is necessary to prove that the solution actually works for the user before it can be refined. Not all AI PoCs should be put into production. Therefore, it requires risk-taking to achieve success with AI.
3. Establish a suitable infrastructure and standards for working with AI
Many believe that the solution to scalable production and operation of AI solutions is to buy an off-the-shelf AI platform. However, an AI platform alone is not sufficient for scaling the production and operation of AI solutions. For some companies, it may be a relevant part of the solution, but it represents just one piece of the overall solution. In reality, it is about ensuring a suitable data architecture and infrastructure - a data platform - as well as establishing fixed standards and processes for working on the development and operation of AI solutions (MLOps). It is essential to ensure coherence between working with AI and the company's overall work with data.
Data Platform: The foundation for scalable AI is a sensible data foundation. Ideally, this foundation is located in the Cloud, as it provides access to scalable infrastructure, relevant services, tools, and models on-demand. AI is a field in rapid development. The cloud makes it easy to stand on the shoulders of a global open source community and leading tech companies when developing and implementing AI solutions.
MLOps: It is crucial to establish best practices for the full life cycle of AI solutions. This involves standards for how to scale and operate ML models, ensuring that scaling AI does not require a corresponding scaling of the number of people working with AI - and that the AI solution actually does what it was expected to do. This primarily involves considerations of reusability across solutions and data teams - such as deployment pipelines, role and responsibility distribution in scaling and operation, as well as structured processes around monitoring, validation, and retraining of ML models.
Under the moniker 'artificial intelligence', the expectations for the business potential in AI have always been sky-high. And true, we can still debate whether artificial intelligence even exists. However, the reason why AI remains a toy in the sandbox for many companies, rather than a transformative technology, does not depend on whether the ideal for artificial intelligence has been achieved and what is technically possible to train an ML model to learn on its own.
The 'AI revolution' is here and available to any Danish company. The challenge is to get the technology into use where it can actually create value and in a way that makes it possible to efficiently identify, develop, deploy and operate the large number of AI solutions that are needed to transform a company's way of operating or the services it delivers.
The path to get there depends on the specific company. In this blog post, however, we have tried to highlight the biggest barriers we typically see for realizing the potential in AI and how we recommend solving them.