The HR agent nobody prompted: what changes when Workday's agents work in the background
Workday's Autopilot agents watch recruiting, payroll and accounting and act before anyone asks. Only recruiting is live, for early adopters. The question: how often they look and what they may do.
Listen to this article · 5 min
AI-generated narration of the full article.
Thursday evening, two days before payroll closes. Nobody in the payroll team has asked anything. An agent has been checking retroactive adjustments all week, has found eleven outside the defined parameters, and has drafted the messages to the affected employees.
That is the operating model Workday announced on October 13 with its Autopilot agents. Most enterprise AI so far has been request-driven: a person asks, the assistant answers or acts. Autopilot agents are persistent. They monitor, detect and prepare work on a schedule, and bring people in when review or judgment is needed. Sana, Workday’s orchestration engine, coordinates them.
What exists today and what is planned
Only one of the four agents is in customers’ hands, and only for early adopters:
| Agent | Works in the background | Waits for a person | Status |
|---|---|---|---|
| Talent Acquisition | Watches open roles, sources candidates | Any outreach to a candidate | Early adopters now; GA early 2027 |
| Payroll | Audits adjustments before the run | Drafted employee messages | Early adopters, early 2027 |
| Accounting | Reads contracts, tests transactions | Entries, revenue recognition, audit calls | Early adopters, early 2027 |
| Deployment | Finds unused Workday features | Configuration and adoption | Early adopters, early 2027 |
Workday’s description of the controls is consistent: customers choose which tasks run automatically, which need review and when a person makes the final call; they set how often agents run; and “every action an agent takes is logged.” The release gives no default frequencies. Its claims that time to fill “moves from weeks to days” and that deployment moves “from months of planning to weeks of execution” are vendor statements without published baselines.
Three settings that did not exist before
A request-driven assistant has one main governance question: what may it do when asked? A background agent adds three, and they belong to the business owner, not the IT team.
How often it looks. Frequency is not just a cost setting. An accounting agent that tests transactions hourly will surface issues the audit team used to find at month end; that changes the audit team’s workload and the evidence it holds. Decide the cadence the process can absorb, not the fastest one the platform allows.
What it may do with what it finds. Detecting a problem, drafting a fix, notifying a person and executing the fix are four different permissions. Workday’s examples stop at drafting and notifying for consequential work, which is the right starting point. Promotion to execution should follow the evidence, one action class at a time, as we argued in autonomy is earned, not configured.
What it may not watch. A recruiting agent monitoring pipelines processes candidate data continuously, not on request. That changes the privacy analysis: purpose, retention and who can see what the agent noticed. Write the scope down before the first run.
Recruiting deserves its own review
The only agent available today is the one that touches hiring. Workday keeps a person in the loop for candidate outreach, and outreach “still requires recruiter approval before anything is sent.” That gate addresses the visible action. The less visible one is sourcing and ranking: which candidates the agent surfaces and which it does not.
In the United States, New York City’s law on automated employment decision tools already requires bias audits and candidate notices when such tools substantially assist hiring decisions for roles in the city. Whether a given configuration falls under it is a legal question for each employer, but the question has to be asked before launch, not after. More broadly, approval of the message is not oversight of the shortlist. Recruiters should be able to see why a candidate was surfaced, and the team should periodically compare the agent’s shortlists with its own.
A checklist for the first background agent
- Name the owner of the cadence, separate from the owner of the platform.
- Separate detect, draft, notify and act as distinct permissions, and start every agent at draft.
- Log what the agent looked at, not only what it did, so a finding can be traced to the data that produced it.
- Set a noise budget. An agent that flags too much trains people to ignore it; measure how many flags lead to action.
- Plan the off switch per process, so payroll can pause its agent during a close without affecting recruiting.
Background agents will be one of the most useful forms of enterprise AI, because the best time to find a payroll error is before payroll runs. They also do their work when nobody is watching. That is the point, and it is why the limits have to be set in advance.