Sending leads to an AI researcher and filing the answers when they come back

Each submitter queries the CRM for records that need a particular answer, cleans names and websites so they are safe inside a JSON body, and posts them to the agent. Only when the agent accepts the request is the record stamped with the date it was sent; an error skips the stamp, so the record comes round again. The lead submitter has separate arms for title, state and website, each with its own stamp and an Ignore handler, so one lead never stops the batch.

The agent replies to a webhook in a receiver scenario. The receiver parses the reply and checks that the record id is present; a malformed answer is ignored without counting as a scenario error. A router on the kind of answer sends each one to its own write (title found or not found, state found or not found, website found), each with a dated stamp.

Companies go through two passes. A fit verdict of true is written without a date, which is exactly what the score submitter looks for, so a fitting company is scored on its next run. A verdict of false is written with a date, and the company's pending leads are detached so the CRM's own flows move them out of the working queue. Scores are parsed with their reasons and never go below ten; a disqualified company gets ten.

For companies that still need a property type, a scenario fetches each homepage and Claude writes a short description and keywords into a company-details field that the CRM's own scorer reads. A homepage that can't be fetched, or that Claude can't classify, gets a date stamp and nothing else. Each run first triggers the CRM's scorer for the companies described the run before, so description and score stay one run apart instead of racing.

Every write across the six scenarios has a Resume, Skip or Ignore handler chosen for its job: skip the stamp so the record is retried, skip the record so the batch continues, or ignore a malformed answer so it is neither written nor counted as a failure.