Two jobs, one handover
Candidate sourcing is the step that finds people who match an open role, drops anyone already in the pipeline, and prepares a first message a recruiter can send. It runs on people who are employed and not applying. The work is search plus a pipeline check on people who will never see the job posting.
Recruiting is everything after someone answers: the screen, the schedule, the panel, the offer, the close. It belongs to whoever owns the requisition. It runs for weeks, across more people than the sourcing step ever touched.
The two jobs meet at one artifact, a reviewed first touch. Sourcing is done when a recruiter has read the draft, fixed what is wrong, and is ready to send. Recruiting starts when the reply arrives. A team that cannot point at that artifact is running one blurry job, and it will buy tooling for the wrong half of it.
Sourcing ends at the reviewed first touch
A sent-email count is the easiest number a sourcing tool can show you. A hiring process needs a later line, the moment a message is good enough that the recruiter who will answer the reply will put their name on it.
The candidate-sourcing job Kortix runs states the boundary in its own header: "Sources and drafts only · recruiter reviews and sends." Its guardrail is narrower still: "The agent never sends outreach itself and never adds a candidate to Greenhouse."
That boundary exists for ownership. A recruiter who edits a draft has decided what the company says to this person, and the reply lands in a conversation they started. A recruiter who reads a send report has inherited a conversation someone else started, with a candidate who expects an answer.
What a sourcing handover has to contain
A batch of drafts is not a handover unless it carries four things:
- who the person is: name, current role, and where the match was found;
- why they matched: one or two details from their background, set against the role's real bar;
- the draft itself, written from those details and nothing else;
- the pipeline state: whether they have applied, been contacted, or been passed on in an earlier round.
The last two decide whether the batch is worth anyone's morning. A draft built from a template reads the same to every recipient, which is the part a reader spots first. A shortlist that skipped the pipeline check sends the recruiter back into the ATS before a single email goes out.
The worked example on the Kortix page carries its own label: "Illustrative example. Your counts depend on the role and the sourcing profile." In it, 31 profiles were reviewed, 13 were already in the pipeline, and 18 drafts went to the review inbox. The pipeline check ran before anyone wrote a line, which is the part of the handover worth copying.
What changes when an agent takes the sourcing half
Run on a schedule, the sourcing half stops depending on whether anyone has the hours. A cron trigger gives one open role one complete sweep each morning. It searches LinkedIn against the role's profile, reads the Greenhouse pipeline, drops the matches already there, writes up to 20 drafts, and stops.
Each run is scoped to one role and gets its own machine. One isolated sandbox per session. Each session has its own isolated machine and branch. The ATS, LinkedIn and the review inbox arrive as connectors. Connector credentials are brokered server-side and never enter the machine.
What moves is the recruiter's judgment, from searching to editing. Reviewing a batch of drafts and sending the ones worth sending is a different morning from hunting for each of those people by hand. The reviewer sees why each match was made before committing to a send.
The rules behind all of this are files: the profile the search runs against, the order of the dedupe check, the 20-draft cap, and the instruction to stop before sending. Changing them follows the same path as any other file in the repo. Session work reaches main through a change request. Merge is default-deny for agents.
The list nobody owns after the reply
Both roles fail in the same place, and it is a day after the send. Replies land in an inbox while the recruiter is in interviews, and nobody has been assigned the second message. The candidate who answered waits, then takes another offer. The pipeline then reports a sourcing problem when it has a follow-up problem.
A sourcing step can prevent that by ending where it does. The batch arrives in a review inbox instead of a sent folder. The person who sends is the person who will answer, and the handover shows up in their calendar instead of being implied by a tool.
Two numbers tell you whether the boundary is holding. The first is how many names the pipeline check removes before drafting, and a rising count means the ATS is doing its job. The second is how many drafts you send without rewriting. A falling count means the drafting rules have learned your voice.
Where to go next
The full sequence, including where the pipeline check belongs, is in how to source candidates. If you are weighing a tool that ranks profiles against a job that reads your pipeline first, AI candidate sourcing covers what a ranked list leaves to you. For the definition itself, the candidate sourcing guide starts there and works down to the pipeline check.