Model Behavior: Harnessing the Benefit of a Vendor’s AI Technology

“AI will most likely lead to the end of the world, but in the meantime, there’ll be great companies.” –Sam Altman



Artificial intelligence (AI) is a controversial topic. It is easy to imagine a near future where AI solves some of our greatest problems and a relatively more distant future where AI becomes our greatest problem. For now, AI has yet to rebel against us and is proving to be a valuable tool in our everyday lives. AI is being deployed to help companies improve productivity, reduce costs, streamline processes, and unlock analytics and insights that weren’t previously available. Like past disruptive technologies, AI presents new issues under familiar areas of concern. Every company needs to know how their data is being used. AI technology adds a new layer of complexity to that all-too-familiar issue.

Any time companies share data with vendors, regardless of whether AI technology is involved, they should pay attention to how their data will be handled, even if only to comply with any applicable laws and regulations. Putting legal requirements aside, companies should also pay close attention to how their data is being handled to ensure that their business interests are being protected. In a previous post on Pillsbury’s SourcingSpeak blog, we discussed the benefit of contracting for full ownership rights over every output arising from a relationship with an AI technology vendor. By doing so, companies can protect themselves against any unintended consequences that could arise from giving an AI technology vendor any type of usage allowance over the data processed by the AI technology, the trained AI model or the knowledge gained from training the model. Even though taking full ownership over every output generated from a contractual relationship provides a high degree of protection, budgetary restrictions often don’t allow for that type of arrangement, so it is important to understand the type of contribution that may be required when engaging an AI technology vendor.

Non-AI technologies, compared to AI technologies, are simpler to protect against when dealing with data usage concerns. Companies can address many such concerns solely by placing restrictions on how the data is used. On the other hand, AI technologies require companies to consider how the trained AI model will be used, and, furthermore, how future AI models will be trained as a result. This concept may be better illustrated with an example.

Vendor X versus Vendor AI
Imagine an online retailer is looking for a technology vendor to provide a display widget that allows the retailer to prominently feature products on its website. The retailer has two options, Vendor X (non-AI technology) and Vendor AI (AI included).

Vendor X provides a user-friendly display widget that allows users to upload product inventory information, track product popularity based on sales converted, and allows users to make selections on the products to feature at any time. Vendor X provides this technology at scale to hundreds of clients and wants to use what it learns from clients to make product improvements to its user interface.

Vendor AI provides the same display functionality as Vendor X, but Vendor AI’s widget tracks consumer behavior on the site, learns shopper habits, and then dynamically decides what products to feature every time the page is loaded. Like Vendor X, Vendor AI provides this technology at scale and wants to use what it learns from clients to make product improvements. Unlike Vendor X’s product that is valued for its user interface, Vendor AI’s product, like most AI products, is measured on the AI technology’s ability to learn and act.

The retailer is in a highly competitive market and is sensitive about sharing any information that may be used to help a competitor.

Taking the retailer’s concerns into consideration and Vendor X’s need to make product improvements, a simple compromise could be made between the parties that all data uploaded to the widget (such as the inventory or product popularity data) will be treated as confidential and deleted upon termination. However, Vendor X may use aggregated or anonymized usage statistics to monitor and improve the product. By differentiating the types of data and placing simple restrictions on their use, both sides can arrive at a fair resolution.

If the retailer and Vendor AI agree to the same restrictions that were agreed to above with Vendor X, the retailer may still be exposed to the risk of inadvertently helping a competitor. When dealing with AI technology, special consideration must be made regarding the trained AI model and how future models are trained. Applied to this example, if the retailer’s use of the AI technology trains an AI model to identify that every time it rains, people buy more hot chocolate, so therefore, on rainy days, hot chocolate should be a featured product, then that trained AI model or specialized knowledge acquired should also be treated as confidential information and deleted upon termination. But the protections should not stop there. The next question that should be asked is: can the vendor teach future untrained versions of the AI model to look for relationships between weather patterns and comfort foods? If the retailer is uncomfortable with that level of transfer learning, could the vendor take more attenuated findings such as identifying the number of interactions it needs to determine reliable buying patterns? Navigating through these types of contributions to an AI technology can become highly nuanced and complex, because depending on the quality and quantity of data being contributed, it may be difficult to measure the downstream impact of allowing any transfer learning to occur.

Remember the Basics
When engaging with AI technology vendors, companies should be mindful about the data they are contributing and the allowable transfer learning from their use, and they should always ask themselves the following questions:

  • What is the value of their data and what are the possible ways it could be used to train future AI models?
  • Could the use of such future AI models by competitors impact their business?

The AI technology space is rapidly evolving, and even as it does, companies are finding creative ways to compromise around these issues. Company decision makers would do well to employ a more traditional form of learning—constantly updating their own understanding of the latest shifts in the landscape, especially prior to any major transaction involving AI. After all, you want contracted AI to boost your own bottom line, not that of your competitors.