Make or buy? That is the question

The great enthusiasm of recent months regarding generative artificial intelligence (for example ChatGPT) and large linguistic models (such as GPT3 and now GPT4) has made even more evident a paradox that has been underway for several years. On the one hand, the increase in interest from companies and the general public for this new technology. On the other, a certain slowness in integrating complete, scalable and functional AI systems into companies, so to produce concrete impacts on business organizaitons. There are many decisions to be taken in the implementation phase. Let's try to tackle one. Remaining on the example of ChatGPT and GPT3, it is easy to see how OpenAI, the company that develops and markets ChatGPT, not only provides the ready-to-use conversational systems but also access via a series of paid cloud services (API) to the family of linguistic models underlying the solution (GPT).

In introducing these tools into the life of companies, one should therefore ask oneself: is it better to use ChatGPT as it was designed and marketed by OpenAI or is it better to start from the model that enables it (GPT3), to create customization specific to the context and types of use we want to work on? A similar question applies to virtually every area of AI, from predictive maintenance systems to quality monitoring systems, passing through forecasting and recommendation systems. On close inspection, however, this make-or-buy choice is not so different from what companies already had to do while developing previous technologies. In the current context of AI, however, some peculiarities make the problem somewhat different compared to other more consolidated technologies. At least three factors influence this dynamic.

The first is related to the very characteristics of AI. At present, in fact, the use of systems based on "machine learning" and the related "training" implies that the elements of the AI system to be customized are substantially different from other technologies. In addition to the enabling infrastructures (software and hardware), in fact, customization can involve data, extend to models or even touch the internal functioning mechanisms of algorithms.
The second factor is linked precisely to the peculiar structure of an AI system. The need for training data with certain characteristics, the computational power (and related costs) necessary to train large models, as well as the technical skills to work beyond the configuration of already existing and tested solutions, represents a barrier to entry for a large percentage of companies in terms of achieving customization.

The third factor refers to the peculiarities of the market which has structured itself in response to these characteristics. Solutions range from finished products and standardized APIs that offer little or no configuration capability, to solutions and development environments that support technically skilled professionals in developing solutions that extend more established approaches.

Although there is no single approach to determining the meeting point between these three factors, and therefore the type of product and level of customization to focus on, it is possible to trace three dimensions that can support the choice.
First, design features. The availability of adequate and distinctive data to carry out the training, the resources (in terms of money and time) available, the costs related to the training, and the level of differentiation and strategic nature of the business problem addressed with AI are all considerations when thinking about the search for the balance between customization and standardization of solutions.

Second, the maturity of the organization. The availability of AI skills in the organization or the presence of adequate resources to supervise and support the work of external operators is an essential point in order to access the creation of systems with a high degree of customization.

Finally, the maturity of the market, both in terms of performance and characteristics of the AI models available to address the strategic problem at hand, and in terms of overall maturity of these systems, which include both the AI component and the connectors and software and hardware elements required for operation. To avoid reinventing the wheel, internal organizational needs must be adequately balanced wit solutions available on the market, articulated on the entire spectrum of available AI options.