HR Tech 2026 closed at Mandalay Bay in Las Vegas on 22 October, and the accounts written around it agree on the main event. The AI agents on show now take actions inside HR systems, and the vendors behind them described their controls in specific terms: approvals, logs, inherited permissions. Read side by side, four published accounts from the week then place responsibility for those actions in four different spots. The vendors’ own releases leave that question to the employer, which means every HR team buying agents next year has to answer it for itself.
What the vendors actually announced
Two releases from the week show the range. On 21 October, Oracle introduced what it calls Claw-powered Fusion Agentic Applications for HR, built on Fusion Claw, a runtime the company describes as governed agentic execution. The first HR applications cover workforce planning, recruiting, compensation and payroll. Workforce Orchestration builds workforce plans by weighing development, redeployment, hiring and external talent. Payroll Command Center reviews employee data, imports, timecards, retroactive changes and historical patterns to flag issues before a run. Workforce Compensation prepares pay plans against employer budgets and pay guidelines for human review.
“The applications reason across Oracle Fusion Applications data, provide context, coordinate the work required to meet business objectives, and carry out human-approved actions,” said Lew Thompson, senior vice president, Applications Development, Oracle. The release closes with a line every HR buyer should read twice: “For employment-related workflows, organizations retain responsibility for employment decisions and can establish appropriate human oversight, approval processes, and controls consistent with their policies and applicable legal requirements.”
ServiceNow went in a different direction a week earlier. Its 15 October EmployeeWorks release, which CX Today grouped with the conference news, lets employees with no coding background build their own Productivity Routines. The company says these background agents repeat the same steps each time and deliver finished outcomes with no human intervention, inside guardrails set by administrators. The first assistants for HR Service Delivery include an HR Business Partner Assistant that handles approvals and case management. All of it is in controlled availability. “AI’s real promise is to close the gap between intention and action,” said Jacqui Canney, chief people and AI enablement officer, ServiceNow.
So one vendor routes each action to a person for approval, and the other lets an employee build an agent that finishes the job alone. Both describe their approach as governed. The coverage then had to decide what governed means.
Four accounts, four places to put responsibility
ChartHop: responsibility lives in the platform
ChartHop, an HR software vendor that exhibited at the show, published its five takeaways on 23 October under the name of Josie Sutcliffe, its VP of Marketing. “Fewer people were asking whether to use AI,” it reports from the floor. The account says governance got more specific, listing logged agent actions, agents that follow the permissions a company already has, and independent testing before an agent is trusted. It points to Workday’s Agent Passport and to UKG’s payroll approach, where AI drafts corrections and payroll professionals review them. Its conclusion for buyers is to ask vendors to show the approvals, the logs and the permissions. “Permissions are becoming the most important AI feature nobody talks about,” it says. A reader should weigh that ChartHop sells a permissions feature of its own, which the post describes.
CX Today: responsibility lives with the manager
CX Today’s Francesca Roche covered the Oracle and ServiceNow news for workforce management leaders on 26 October. Her account accepts the vendor governance story and moves the weight to people. As agents take on transactional work, she writes, “managers must be ready to translate strategy into daily behaviors.” Her clearest line draws the boundary: “a customer recommendation can be automated, but accountability for its human and service impact cannot.” In this telling the control is manager capability, and the question for leaders is whether their managers are ready.
Rep Cap: responsibility is unresolved and costs hours
Steve Smith, writing Rep Cap’s Work Tech Weekly on 19 October, framed the same question as an open one going into the show. He called accountability the elephant in the Expo Hall and attached the only cost figure in any of the four accounts: in G2 data, 31% of buyers say employees spend six-plus hours a week overseeing agents. “That’s a job design problem,” he wrote. On payroll he wrote, “A payroll error isn’t a hallucination you can apologize for.” His expectation is that incumbents keep the rules and data while outside models do the reasoning.
Paul O’Brien: responsibility needs a named owner
Paul O’Brien, a startup economist who attended, wrote his preview on 19 October from outside the HR buyer’s seat. Most of his piece argues that payroll, scheduling and HR systems already hold retention data that cities, investors and schools ignore. On agents he writes that “Someone has to own those agents, audit them, and answer for what they do, and in a startup that someone is too often a nobody,” and adds that “an agent nobody manages is an employee nobody can fire.”
Where the accounts disagree
All four agree that the vendors moved from assistants that answer to agents that act. They split on where the control sits. ChartHop treats governance as a product property a buyer can inspect in a demo. CX Today treats it as a management skill. Rep Cap treats it as a labour cost that lands on whoever reviews the agent’s work. O’Brien treats it as an ownership gap, with a person missing from the org chart.
They also differ on mood. ChartHop and CX Today write as if the controls have arrived. Smith and O’Brien, both writing before the doors opened, treat the controls as questions the show would have to answer. The two later accounts report the vendors’ design choices, and neither puts a number on the review work those choices create.
In our reading, the coverage adds up to a handoff. Vendors shipped approval steps, logs and permission models, and Oracle’s release states in plain words that employment decisions stay with the employer. No account and no release from the week names the HR role that approves an agent’s payroll correction, says how many hours that review takes, or describes what happens when an employee-built routine in a ServiceNow-style tool touches an HR case. The approval mechanics exist. Staffing them is the piece none of the four accounts covered, and Smith’s G2 figure is the only number on its cost.
We have argued before that HR should staff agent operations before the credential exists, and the week’s launches add to that case. Earlier this month Workday framed its own agents the same way, as autopilots that run within customer-set limits, and our comparison of four vendors’ cases for trusting AI agents with pay data found each one fencing the model in a different place.
What it means for the HR leader
The four accounts give an HR leader four checks, one from each, and running all four covers the gap between them.
From the platform view
Ask every vendor on a shortlist to show, in a live demo, the log entry for one agent action: what it did, what data it used and who approved it. Then ask whether the agent runs with the permissions of the person using it or with a broader system account.
From the manager view
List the agent workflows planned for 2027, such as payroll pre-run checks, position changes and compensation plans, and write next to each the manager or HR role that approves its output. A workflow with no name next to it should not go live.
From the cost view
Estimate review hours per week for each workflow before signing. If the G2 figure holds for your teams, six hours a week of oversight is about 15% of a 40-hour week per reviewer, and that cost belongs in the business case.
From the ownership view
Decide who can switch an agent off, and write it down. For tools that let employees build their own routines, set the rule for which HR data those routines may touch before the feature leaves controlled availability.
Oracle’s release puts responsibility for employment decisions on the employer. The first draft of the approver list is the document that makes that sentence workable, and it can be written this quarter, before any of these agents reaches general use.
Source: Oracle, ServiceNow
