A good AI agent should pay for itself many times over. The problem is that the value hides in three different ledgers, and most teams count only one: usually the smallest.
Measure an agent on three lines: revenue generated (deals the agent touched before closing), revenue recovered (carts and stalled deals it brought back), and tickets deflected (questions resolved without a human). Counting only the first understates the return by a wide margin, because deflection is usually the steadiest of the three.
1. Revenue generated
Conversations that touched the agent before a purchase. This is the most visible line and usually the one leadership asks for first. It is also the one most likely to be overstated, because "touched" is doing a lot of work in that sentence.
Be strict about attribution. A useful rule is to count only conversations where the agent answered a substantive question: not sessions where the widget was opened and closed. An honest smaller number survives scrutiny; an inflated one gets the whole programme questioned the first time someone audits it.
2. Revenue recovered
Abandoned carts and stalled deals the agent brought back. Easy to overlook, often enough to justify the platform on its own. This line is where proactive triggers earn their keep, and it is the cleanest to measure because there is a clear counterfactual: the session was leaving.
3. Tickets deflected
Questions resolved without a human. Multiply the volume by your cost per ticket and the saving stacks up fast. This is also the most defensible line internally, because support cost per ticket is a number the business already tracks.
500 conversations × 30% resolved without a human × $15 per human ticket = real money saved, every month.
24/7
The fourth line nobody counts
The fourth return is research. Every conversation is a record of what a prospect wanted to know before buying, in their own words. That corpus tells you which page is unclear, which objection repeats, and which feature people assume you have: insight no analytics tool produces, because analytics records clicks, not questions.
Teams that read their conversation logs monthly tend to change the site more, and more accurately, than teams that only read the dashboards. It is the cheapest research budget you will ever have.
When the numbers say stop
An honest framework has to be able to say no. If, after a full quarter, deflection is under 15% and recovered revenue is not covering the subscription, the problem is usually one of two things: the knowledge base is too thin for the questions being asked, or the hand-off is broken and conversations die at the edge. Both are fixable, and both are visible in the logs.
Add the three lines up and the question stops being "can we afford an agent?" and becomes "can we afford not to have one?". The pricing page has the cost side of that equation; the features page has what the agent has to do to earn it.
