India, Nov. 16 -- An AI agent may be able to answer thousands of customer questions. The figure that matters to the business, says Bob Wambach, is what each answer costs, how good it is, and how that compares with the way the work was done before. A token bill tells a company how much it spent on model usage. It says little about what the company got in return.

The stakes rise when an agent can act on a live system. Adding disk space before an application runs out is a relatively straightforward decision. Adding computing capacity after a software release is different: the extra capacity may hide a problem in the new code. Giving an agent permission to act means knowing whether it can tell those situations apart and what to do if it gets...