AI candidate screening in 2026 means an agent reads applications against a written rubric, returns a per-criterion score with cited evidence, and writes the decision back to the ATS. It does not mean a black-box "fit score." The workflows hiring managers trust have three properties: the rubric is published before screening starts, every score links to the resume text supporting it, and every reviewer override is logged. The model choice is replaceable.
The workflow
1. Write the rubric in the JD, not in the model. Before any AI touches a pipeline, the hiring manager writes 4 to 7 criteria with weights and what "evidence" looks like for each. "5+ years Postgres at scale" is a criterion. "Strong technical background" is not. Most teams keep this in the requisition record in Greenhouse or Ashby so the rubric travels with the role.
2. Score with citations, not vibes. Hand the rubric plus the resume to Claude or ChatGPT through your ATS integration. The prompt asks for one score per criterion, one sentence of rationale, and a direct quote from the resume as evidence. Ashby's native AI screening and Eightfold both expose this pattern. If you are rolling your own, the structured-output mode in either model gives you JSON you can write straight to a Greenhouse scorecard.
3. Normalize at the recruiter step. The recruiter spot-checks the top 20 and the bottom 20 against the rubric. They are not re-scoring every candidate. They are looking for systematic drift: the model overweighting brand-name employers, missing non-US degree equivalencies, penalizing career gaps. When drift shows up, the rubric gets edited and the batch is re-run. This is the same pattern as a dangerous-ops contract for screening: the agent proposes, the human gates the batch.
4. Hiring manager reviews rationale, not rankings. The manager opens the top 15 and reads the per-criterion notes. They are checking whether the model understood the role, not whether they agree with the ranking. A bad rationale on the number-three candidate is more diagnostic than the order itself.
5. Log every override. When the manager moves a candidate up or down, the reason gets captured next to the original score. This is the artifact you need for agent audit and compliance, and it is what an EEOC adverse-impact review will ask for first.
Worked example: SDR role at a Series B SaaS company
400 applicants in Greenhouse. Rubric is 5 criteria: outbound experience, ICP fit, written communication, ramp signal, location. Claude scores each application in about 4 seconds and writes a structured scorecard back. The recruiter pulls a stratified sample. She notices the model is treating "BDR" and "SDR" as different and penalizing BDR titles. She edits the rubric, reruns, and the band reshuffles. The hiring manager opens the top 12, reads rationale, advances 8 to phone screen, and flags 1 to revisit. Total elapsed time: under a day. Every decision has a written reason.
Where the workflow breaks today
The agent forgets. The rubric edit the recruiter made on Tuesday does not survive to Friday's batch. The hiring manager's reason for advancing candidate #7 lives in a Slack thread the model cannot read next week. The override audit trail the EEOC will eventually ask for is scattered across five tools. One way to solve this is a workspace like Dock that holds the rubric, the per-candidate rationale, the reviewer overrides, and the agent identity doing the scoring as persistent rows, with pointers (greenhouse_application_id) back to the ATS record so Greenhouse stays the system of truth for candidate data. The pattern is covered in more depth in Dock for HR.
Why it matters
Screening is the highest-volume, highest-risk step in hiring. The EEOC's AI governance program and the NIST AI Risk Management Framework both push the same direction: documented criteria, traceable decisions, human review. A screening workflow that cannot produce the rubric, the score, the rationale, and the reviewer override on request is one subpoena away from being shut down.
Want the full hiring stack, not just screening? Read the pillar: How to do hiring with AI in 2026.
FAQ
Is AI candidate screening legal in 2026? Yes, in every US jurisdiction, subject to disclosure rules in NYC (Local Law 144), Illinois (AIVIA), and Colorado (SB 205). What is not legal is using a screening tool you cannot explain. The EEOC treats AI screening as a "selection procedure" under Title VII, which means adverse-impact testing applies the same way it does to a written test.
Which model should I use for screening? Claude and ChatGPT both work. The model matters less than the rubric and the structured-output contract. Pick the one your ATS integrates with cleanly. Ashby and Greenhouse both ship native integrations; Eightfold runs its own models. The replaceability is the point: if a better model ships next quarter, you swap it without rewriting the workflow.
How do I prove the screening was fair if challenged? Produce the rubric as it existed on the date of screening, the per-candidate score with cited evidence, the reviewer overrides with reasons, and the pass-through rates by protected class. If any of those four artifacts is missing, you have a problem. This is why the persistence layer matters.
Can the agent reject candidates automatically? It can, but most teams that have been through an EEOC inquiry do not let it. The pattern that holds up: the agent ranks and writes rationale, a human recruiter clicks reject. The audit trail then shows a human decision-maker on every rejection, with the AI scorecard as input rather than verdict.