Six jobs that keep working after everyone logs off.

It reads across the entire book, not the recent end of it — the quote that went quiet in March, the household with a second line nobody ever pitched. Left to themselves producers work the names they remember, which is a shrinking circle. What changes for a producer is that the worklist stops being “whoever came in yesterday” and starts being whoever is genuinely worth a touch today. What it will not do is manufacture urgency to fill a list: a lead it cannot justify surfacing stays where it is.

Every lead on the worklist carries three buttons — worked, not hot, dismiss. Those clicks are the training data. Once a week the ranking is refit against what your agency actually converted, so a shop writing final expense to 70-year-olds and a shop writing IUL to business owners end up with genuinely different lists. The adjustment is deliberately conservative: it tunes, it does not lurch, because a scorer that swings on one bad week is worse than one that never moved at all.

It reads finished conversations and records what actually happened in each — booked, objection raised, price question, went quiet, wrong number. The least glamorous job here and the one everything else stands on: the other five agents are learning from outcomes, and without this they would be learning from guesses. It is also what lets you ask “what do people actually say no to” and get an answer from your own threads instead of an opinion.

It watches the moments a human took a conversation over, and what they said instead of what was drafted. The instinct your best closer has for the one line that turns a “just looking” is the hardest thing in an agency to transfer — normally it leaves when they do. This turns those takeovers into something the rest of the system can use, without anyone booking a training session or writing a playbook nobody reads.

It reviews the AI’s own transcripts the way a manager reviews call recordings, and drafts specific edits — this opener underperforms on this product line, this objection is being answered in a way that stalls. Proposals sit in a queue for a human to approve or bin. It never rewrites itself behind your back, and that is a limit we chose rather than one we have not got to yet: an AI that edits its own instructions unsupervised is a thing you cannot audit after the fact.

It re-reads bookings against what the lead actually asked for and flags the ones that drifted — the lead said four o’clock and the calendar says four-thirty, or said Thursday and got Tuesday. A no-show that was really a mis-book is the most expensive kind, because everyone involved believes the lead flaked. Ships in flag-only mode: it tells you, and you decide whether it is ever allowed to correct one itself.