AI candidate sourcing in 2026 is a four-stage workflow: the agent pulls prospects from LinkedIn Recruiter, GitHub, conference rosters, and prior pipelines; a sourcer triages with a structured rubric; outreach goes out through Gem or hireEZ; and the candidate notes persist into a workspace the next requisition can read. The pieces work. What breaks is memory. Most teams rebuild the same passive-candidate research three times a year because nothing holds it between roles. This post is the workflow that fixes that, with tools named at every step.
The workflow
1. Define the role in writing first. Before any tool runs, write a one-page brief: must-have skills, three "tell" signals (a specific repo, a conference talk, a prior employer), and the disqualifiers. Most sourcing waste comes from skipping this. The brief is the prompt your agent will use for the next eight weeks. For the deeper version of this step, see the AI hiring pillar.
2. Fan out across sources. Run the brief through LinkedIn Recruiter for employment history, GitHub or GitLab for engineering signal, Crunchbase for funding-stage context, and conference proceedings (KubeCon, NeurIPS, SaaStr) for speaker lists. ChatGPT and Claude are useful here as the synthesis layer: feed them a LinkedIn export plus a GitHub profile and ask for a one-paragraph match assessment against the brief. Bardeen handles the scraping glue between sources when an official API does not exist.
3. Triage with a scoring rubric, not vibes. Score every prospect 1 to 5 on the three "tell" signals and reject anything that scores below 3 on two of them. A sourcer can clear 200 prospects in an afternoon this way; without the rubric the same pile takes a week. Gem and hireEZ both have built-in scoring, but a Google Sheet works if you do not have a license yet.
4. Sequence the outreach, then hand off. Gem and hireEZ run the multi-touch sequences. Personalization is where the agent earns its keep: paste the candidate's last blog post or commit into Claude and ask for a two-sentence opener that references it without being weird. Reply rates on personalized first touches run roughly 2x the cold-template baseline in our spot checks.
5. Move warm replies into the ATS with context. When a candidate engages, push them into Lever, Greenhouse, Ashby, or Workday with the full research note attached, not just the resume. The screening step downstream will be twice as fast if the sourcing notes survive the handoff.
Worked example: a Series B staff engineer hire
Fifty-person Series B, hiring a staff backend engineer, eight-week timeline. The sourcer wrote a brief naming three signals: a Kafka or Pulsar repo with real commits, a talk at a distributed-systems conference, and prior time at a company with similar scale (10M to 100M events per day). LinkedIn Recruiter returned 340 names matching title and tenure. Bardeen pulled GitHub handles for the 180 with public profiles. Claude scored each profile against the three signals. The sourcer triaged the top 60 in one sitting, Gem sequenced 24, eleven replied, four converted to onsites, one accepted. Total sourcer time: about 14 hours, against the 40-plus this role used to consume.
Where the workflow breaks: the notes vanish
Eight weeks later the same company opens a senior backend role with overlapping criteria. The 56 prospects who scored well but were not contacted, or who replied "not now," are gone. The agent has no memory of the previous run. The sourcer rebuilds the list from scratch.
One way to solve this is a workspace like Dock that holds the candidate notes, scores, and "not now" responses as durable rows the next requisition can query, with greenhouse_application_id or lever_opportunity_id pointers back to the ATS so the canonical record stays where it belongs. The ATS owns the application, Dock holds what the agent interpreted around it. See Dock for HR and the agent identity lifecycle for how the audit trail works when the agent acts on behalf of a sourcer.
Why it matters
LinkedIn's 2025 Future of Recruiting report found recruiters using generative AI save roughly a full workday per week (LinkedIn). That savings disappears the moment you start every new role from a blank slate. Compounding only happens when the research persists. ERE's sourcing community has tracked the same pattern for years (ERE): the best sourcers keep a personal "talent map" outside the ATS. The 2026 version of that map is an agent-readable workspace.
Try the full pillar
If you are still scoping the program, the AI hiring pillar covers sourcing, screening, interviewing, and offer in one place.
FAQ
What is the single biggest mistake teams make with AI sourcing? Skipping the written brief. Without three concrete "tell" signals the agent has nothing to score against, and the output is a generic title-and-tenure list.
Do I still need a sourcer if the agent does the work? Yes. The agent fans out and scores; the sourcer triages, personalizes, and handles the human conversation. The role shifts from list-building to judgment.
Should I pick Gem or hireEZ? Gem leans toward CRM and sequencing depth; hireEZ leans toward diverse-pipeline search and contact-info enrichment. Pick based on which gap is bigger for your team this quarter.
How do I keep candidate notes from going stale? Tie every note to a date and the role it was written for, and have the agent re-verify the top signals (current employer, last public activity) at the start of any new requisition that reuses the prospect.