Can AI Lead Scoring Learn From the Customers It Ranks Last?
Thryv says businesses using AI Lead Insights converted leads at 1.5 times the rate and generated 40% more revenue per client. The company says those figures come from an internal analysis of roughly 6,000 potential customers, not a randomized trial. That is a useful reason to test the score, not hand it the phone book. The harder problem is that the ranking changes the data used to judge it. If the top leads get fast callbacks while the bottom waits, the score can help create the outcome it later claims to predict. A bad score can look accurate if it is allowed to starve the customer. For the first month, I’d call a small random sample from the bottom as quickly as the top, then compare bookings, job value and why the score missed. If nobody calls the low-ranked leads, what evidence could ever prove the model wrong?
Comments
When a low-ranked lead books, keep the first message beside what the staff learned later. “Need quote” may look useless until the caller explains that the furnace is out or the new number belongs to a regular customer. The lesson is that the shop was missing context—not that a person in a mess should write a better form to earn a callback.
Once the call connects, put the score away. A “low intent” badge can turn into a colder conversation, so the ranking starts manufacturing its own proof even when someone answers. Use it to order the queue, not tell staff how seriously to take the person.