The useful answer to "can AI prep an interviewer" is yes, but only when the agent has the resume, the prior round notes, and the role rubric in front of it. Generic question generators produce questions any candidate could have rehearsed. A real workflow stitches Greenhouse or Ashby data, Pillar.app transcripts, and a model like ChatGPT or Claude into a brief that names the signal each interviewer should chase in the next forty-five minutes.
The workflow, step by step
1. Pull the canonical record. The agent reads the candidate row from Greenhouse or Ashby: resume, recruiter screen notes, the structured scorecard for the role. This is the same record the screening pass was written against, so the agent already knows which competencies were rated and which were marked unknown.
2. Read what previous interviewers actually heard. If the loop uses Pillar.app or Metaview for interview recording, the agent ingests the prior round transcripts and pulls out specific claims the candidate made: "led the platform team," "rewrote billing in six weeks," "owned the on-call rotation." Those claims become the things the next interviewer probes.
3. Map gaps to the rubric. The role rubric lists the competencies the panel must collectively cover. The agent checks which ones are still rated "no signal" after rounds one and two, then assigns the open competencies to the right slot in the loop. The hiring manager gets technical depth, the cross-functional partner gets collaboration, the bar raiser gets the failure mode.
4. Draft the brief. For each upcoming interviewer the agent writes a one-page brief: candidate summary, two or three claims from prior rounds worth probing, three to five targeted questions tied to the rubric, and a short note on what good and bad answers sound like. The interviewer edits the brief inside Greenhouse or Ashby, or in a doc the agent can read back from.
5. Hand the debrief back the same way. After the interview the scorecard goes into the ATS, and the agent files the rationale (which question got probed, what the candidate said, why the rating moved) alongside the prior rounds so the next interviewer inherits it.
Worked example: a Series B platform engineering hire
A 50-person infra team is hiring a senior platform engineer. Recruiter screen and technical screen are done in Greenhouse. The Pillar.app transcript shows the candidate said she "owned the migration off Kubernetes" but the screener did not press on which workloads or what the rollback plan was. The agent flags this as the top probe for the hiring manager round. It also notes that "incident response under load" is still unrated, and routes that competency to the staff engineer doing round three. Each interviewer opens a one-page brief twenty minutes before the call. After debrief the loop has full rubric coverage and the hire/no-hire meeting runs in twenty minutes instead of ninety.
Where this breaks today
The brief is only useful if the next interviewer can see it. Most loops fail here. Round one notes sit in Greenhouse, round two in a Notion doc, round three in someone's head. The agent has no persistent place to keep the running picture of the candidate, so it starts from scratch each session and the panel ends up asking the same question three times. One way to solve this is a workspace like Dock that holds the agent-readable interview brief, the rationale per round, and the open rubric gaps as live rows, with pointers back to the greenhouse_application_id so the canonical scorecard data stays in the ATS. Pair that with agent identity and audit trail so you can see which agent wrote which brief, and the compliance posture survives a later EEOC or adverse action request.
Why it matters
Structured, evidence-based interviewing is the single biggest lever on hiring quality, and the research on selection method validity has been consistent for thirty years. The constraint was never the method, it was the prep cost. Agents collapse that cost, which means structured interviewing finally gets used the way the structured interview literature said it should.
For the full pillar, see how to do hiring with AI in 2026. Related: AI interview scheduling and Dock for HR.
FAQ
Q: Does the AI replace the interviewer? A: No. The agent prepares the brief, the human runs the conversation and writes the scorecard. The model is bad at judging humans in real time and good at reading documents and proposing targeted probes.
Q: What tools do I actually need to start? A: An ATS that exposes candidate data through an API (Greenhouse and Ashby both do), an interview transcription tool like Pillar.app or Metaview, and a model with a workspace context. ChatGPT or Claude both work for the drafting step.
Q: Won't candidates feel surveilled if interviews are transcribed? A: They have to consent. Pillar.app and Metaview handle the notice and consent flow. Without consent the workflow falls back to interviewer-written notes, which the agent can still ingest.
Q: How do I keep the brief from leaking bias from earlier rounds? A: Surface only claims and rubric gaps, not ratings. The next interviewer should know what the candidate said, not what the previous interviewer thought of her. This is the same discipline that makes structured interviewing work without AI.