This is the third article in a series focused on creating a valuable patent portfolio for AI technologies, building on earlier discussions of why to invest in patents for AI and how to overcome the biggest challenges in patenting AI.  This article will focus on how to match up your patent portfolio with the most valuable aspects of your business, setting the stage for future articles on frameworks for how to prepare AI patent applications and how to effectively move them from filing to grant.

Building a patent portfolio is inherently a question of resource allocation, and this is no less true when protecting AI technologies. These resources include the costs of preparing and prosecuting patent applications, as well as the time and attention that your inventors and other team members may need to provide to help your patent counsel create a valuable portfolio (ideally you can create a workflow that keeps the patent process lightweight to allow your inventors to focus on innovation).

It may be helpful to prioritize the aspects of your technology that you patent to ensure effective protection while maintaining efficient resource allocation. For example, the following factors can be considered when deciding how to devote efforts to different aspects:

What is the value-add of the AI solution to your business: the part of your business where the AI solution comes into play can guide where to direct patent protection. For example, if you sell devices or services that use AI to make real-time decisions, it may be most valuable to build protection around these decisions. This may be the case across a variety of industries and technology sectors, from computer vision technologies that classify real-world objects in order to make decisions based on the classification, to natural language processing technologies that interpret spoken language in real time to communicate with users, to recommendation engines that recommend products or services to customers.

If you sell products that are manufactured using AI systems, such that the product implicitly incorporates the AI value-add before it is installed or operated by an end user, it may be useful to file applications with claims focused on model training (ideally this can be accompanied by other applications that are directed to inventive features of the product itself that do not rely on the AI solution). For example, this may be the case if you sell components in which temperature and pressure of the manufacturing process for the components are controlled using an AI algorithm.

If you currently have access to unique data that allows you to train smarter, more accurate machine learning models, consider pursuing claims focused on the generation and management of training data or model training, which can act as a roadblock in case your competitors gain access to similar data in the future.

A related consideration is whether your business is the AI vs. whether you are implementing AI to improve your products and services. If the former, it may make sense to devote more resources to protecting the model and data aspects of your technology, since it is likely that the value add of the AI solution can be applied to a very broad range of actions or insights. If the latter, it may be worthwhile to devote more resources to protecting the actions that are performing or the insights that are delivered using the AI solution that you are incorporating into your overall commercial strategy.

Who are your competitors and customers: a basic rule of patent strategy is to make sure that claims can be infringed by a single actor. Therefore, AI-related patents will be more useful when the solution being claimed is something that a competitor would perform without actions by other parties. 

For example, if other companies are using customer data to generate insights that are then delivered back to the customers, and you have a reasonable expectation that they are applying AI algorithms to generate the insights using the customer data, protection can be effective in many ways — go after the insights themselves, a few examples or a category of the machine learning models that might be used to generate the insights, and how the customer data might be filtered or scrubbed prior to use by the AI algorithm to result in a more accurate or efficient algorithm.

As another example, if other companies are training the machine learning model in-house, and supplying software or firmware to end user devices with a fully-trained model (e.g., a neural network in which the weights and biases of the neurons have been set during the in-house training), consider focusing on the data used to train the model and the process of training the model, so that the actions performed by the end user’s device are not required for infringement.

These considerations can be used to establish frameworks for preparing patent applications, which will be discussed in the next article in this series.