How do I know if my AI is still working properly?
Traditional technology is deterministic: the same input always produces the same output. But AI models are not. They reason within a context, and if that context isn't carefully managed, the quality of the responses can deteriorate. What was once a correct answer may no longer be so, and an indicator like customer satisfaction begins to gradually decline.
At that point, the risk arises that no one will notice it in time, when organizations measure the results in the first few months and celebrate them, but leave the model in production without follow-up.
In this episode of our "Off the Record" vlog, we explain what AI observability is and how an AI model degrades. You'll learn that AI models are not deterministic and can degrade over time; that observability automatically measures quality, resource consumption, biases, and availability; that maintaining a model in production requires more effort than building it; and that what is observed can even change the original business case.
What is AI observability?
It is the ability to automatically measure how a model behaves in production: the quality and quantity of responses, consumption, biases, and availability. These metrics are used to define alerts and decide when to adjust or apply...“fine tuning” or retrain. It also allows for responding to regulatory bodies, which request evidence on bias, quality, and availability.
The most underestimated effort
Development is generally the least demanding phase. However, once in production, a model requires more attention than during its development, including observability, a support team, and governance. For example, a two-person team developing a solution may need four people to maintain it.
An example: when the observation changes the business case
In a project with a conversational sales agent, the initial goal was to improve lead conversion. By analyzing the conversations, the team realized that users primarily needed help using the tool. They changed the agent's pitch, expanded the reach to other users, tweaked a piece of technology to respond faster, and added a web channel. All of this was discovered because someone was watching.
