How to Speak Machine: Overcoming Challenges in Providing AI Context

In my last article, I emphasized that capturing data context is a crucial step companies must take to prepare for and maximize their AI investments. The key is to ensure that the data your company feeds to an AI service includes as much accurate and detailed context as possible. Context, which is metadata describing the data, can include a physical description of what the data represents, any categories or taxonomies the data belongs to, the data’s relation to other data, the source of the data, and more. This context provides information that AI engines can use to more accurately and efficiently determine which patterns to apply to the data, resulting in more meaningful outcomes.

Humans create and use context all the time. Sharing context is the reason we communicate. Language is the most conspicuous example of this, but we also share context through a look in our eyes, our body language, and the tone of our voice. Context can even be conveyed by what is left unsaid. Our shared memories and experiences provide context, enabling highly complex and detailed concepts to be communicated with a single word. It’s not an exaggeration to say that one of humanity’s "superpowers" is the efficient communication of context.

When using AI, it is critical to remember that we aren’t communicating with another human; we are interacting with a machine. Machines have very specific and limited ways of sharing context. They lack our shared memories and experiences and have none of our non-verbal means of sharing context. For machines to properly interpret the data they receive, the context must be specific and unambiguous. The best way to communicate specific and unambiguous context is to keep concepts simple, describe them in complete detail, and use consistent terminology.

Describing context this way has other benefits.

Maximize Data Reuse: By breaking context into its component parts, data with similar but not identical contexts can be identified and combined in new and innovative ways.

Interpretable Results: The simpler the data, the easier it is to understand how an AI algorithm arrived at a result. This simplicity leads to greater confidence in the output and provides insights into how to modify the algorithm for better performance.

Mitigation of Bias: Using shortcuts to communicate complex concepts can lead to the assumption of inappropriate and inaccurate contexts, reinforcing and amplifying negative outcomes. Describing context as specifically as possible ensures that the right factors are utilized or omitted, reducing bias.

Most companies do not systematically capture context in a simple, complete, and consistent manner. Before digital technologies, most business processes relied on human-to-human communication. Sharing context in such detailed ways would have slowed things down and created confusion, as our brains aren't wired to handle that level of detail.

But computers are different.

Computers, and the AI tools that run on them, are designed to access, process, and communicate millions of pieces of data per second. For them, too much data is not a problem. So, how can companies ensure their AI solutions receive the required context without overwhelming the people who operate them?

Bring the machines along for the ride.

Every day, individuals and teams make thousands of small decisions that drive business results. Each of these decisions (e.g., using red instead of blue, choosing one segment attribute over another, asking a specific team for assistance, modifying a tool to a certain extent) is a point of context that impacts business outcomes. If companies can record these decisions as they are made, using consistent terminology that is understandable by all, they can capture accurate, granular, machine-interpretable context with little or no additional effort.

Three things are required to make this happen:

1. A Simple, Shared, and Easily Understood Data Framework: This framework should describe context, including the data’s definition and relationship to other data.
2. Tools and Services: These should be designed to capture and share context so it is accessible to the applications and processes that need it.
3. A Rollout Plan: This plan should deliver both immediate value and long-term transformation, minimizing the impact of organizational change.

Next time: What this model looks like for business teams that engage with customers.