Workday stays the system of record. The AI layer sits next to it. In enterprises that run on Workday HCM and Recruiting, the practical 2026 workflow is: an agent (ChatGPT, Claude, Phenom, or Eightfold AI) drafts requisitions, screens applicants, and proposes interview slots; a human recruiter reviews and approves; the signed-off rationale gets written back through Workday Studio or the Recruiting REST API. The candidate record stays in Workday. The interpretive layer, what the agent thought and why, persists somewhere auditable. That last piece is where most pilots fail.
The workflow that survives an audit
1. Requisition drafting with ChatGPT or Claude. The recruiter pastes intake notes into a prompt template. The model returns a structured draft: title, level, scope, must-haves, nice-to-haves, comp band placeholder, EEO language. The recruiter edits, opens the requisition in Workday Recruiting, and saves the prompt, model version, and diff as artifacts. This matters when legal asks six months later why the requisition said what it said. See the pillar on AI hiring in 2026 for the template library.
2. Sourcing with Eightfold or Phenom. Eightfold's Talent Intelligence and Phenom's X+ platform both integrate with Workday at the requisition level and push ranked candidates back with a fit score and a one-paragraph rationale. The rationale is the part that needs to persist. The Workday candidate record holds the application; the AI's reasoning needs a separate home with a pointer back to workday_candidate_id.
3. Screening with a structured rubric. Drop the resume and the requisition into Claude or ChatGPT with a fixed rubric: domain depth, scope of past roles, ownership signals, two flags. The model returns a scored card. The recruiter approves or overrides each line. Override reasons are the most valuable artifact in the whole pipeline; they teach the next iteration of the rubric.
4. Interview scheduling. Paradox's Olivia and GoodTime handle the calendar choreography. Workday holds the interview kit. The agent proposes panel composition based on requisition skills and prior interviewer load. See AI interview scheduling for the full handoff pattern.
5. Decision and writeback. Panel feedback goes into Workday. The synthesis (why this candidate, why not the other two finalists) is the decision artifact. Workday's text fields will hold it, but they will not version it, link it to the prompts that produced it, or expose it to the next agent that picks up the requisition.
Worked example: GC search at a 4,000-person SaaS company
The General Counsel role opens in Workday Recruiting. A sourcer uses Eightfold to surface 40 candidates. Claude scores each against the rubric the search committee approved. Twelve advance to recruiter screen. Paradox schedules. Three reach final panel. The committee picks one. Six weeks later, a board member asks why two specific finalists were declined. The Workday record shows the disposition code. The actual reasoning, the scored cards, the committee chair's override notes, the screening prompts, lives wherever the team put it. If that place is a shared doc folder, it is stale. If it is the recruiter's chat history with Claude, it is gone.
Where the workflow breaks: enterprise audit trail
The persistent-state gap is the entire problem. Workday holds the candidate. The ATS comparison piece on Greenhouse AI recruiting hits the same wall from the mid-market side. Compliance teams need the full chain: prompt, model, output, reviewer, decision, timestamp. One way to solve this is a workspace like Dock that holds the agent's rationale, the reviewer's edits, and the signed-off decision as structured rows with pointers back to workday_requisition_id and workday_candidate_id. Workday stays the canonical record. The interpretive layer becomes queryable. See agent audit and compliance for the chain-of-custody pattern and agent identity lifecycle for how to attribute writes to a specific agent version.
Why this matters
Enterprise hiring at Workday scale is not a tools problem, it is an accountability problem. Josh Bersin's talent research and SHRM's talent acquisition guidance both flag the same gap: AI adoption is fast, governance is not. Teams that ship treat the agent's reasoning as a first-class artifact, not a chat transcript.
If you operate Workday today and want the rationale layer in one place, Dock for HR shows the row model.
FAQ
Does Workday have native AI for recruiting? Yes. Workday Recruiting includes AI-assisted candidate matching and skills inference, and Workday acquired HiredScore in 2024. For deeper agent workflows (requisition drafting, structured screening, panel synthesis), most enterprises layer ChatGPT, Claude, Phenom, or Eightfold on top of Workday rather than replacing it.
Can AI agents write back to Workday directly? Through Workday Studio integrations and the Recruiting REST API, yes. The pattern that passes review is agent-drafts, human-approves, then the approved payload writes back. Direct agent writes to a system of record without a human gate fail most enterprise risk reviews.
What gets logged for SOX or EEOC defensibility? At minimum: the prompt template version, the model and model version, the raw output, the reviewer identity, any overrides with reasons, and the final decision with timestamp. Workday holds the decision. The full chain needs a separate, queryable store.
How is this different from Greenhouse plus AI? Greenhouse is the ATS for many mid-market and growth-stage companies; Workday is the HRIS plus Recruiting suite that anchors most enterprises. The agent workflow is similar in shape. The compliance surface area at Workday-scale companies is larger, which is why the audit-trail piece is non-negotiable.