Intention-to-treat measures the effect of making something available, across everyone it was made available to. If a firm gives an AI assistant to a thousand people and eight hundred use it, the intention-to-treat effect is averaged over all thousand. The effect on the eight hundred users alone is a separate and larger figure.
For a business case the intention-to-treat figure is usually the one that matters, because the organisation pays for every licence and cannot pick its users in advance. Vendor case studies tend to quote the effect on active users, which overstates the return on the licences bought.
The same logic explains why comparing heavy users with light users is not a measure of benefit. Heavy users chose to use the tool, and they often differ from light users in ways that also affect the outcome. Use that comparison to plan training, and book the intention-to-treat effect.