Vendors are pitching an agent system of record to HR. Critics say it belongs to IT. Both are arguing about the wrong object — and HR has lost this exact argument once before, with contingent workers.
Workday put the question on the table in February 2025 when it announced an Agent System of Record: a place to onboard AI agents, manage their lifecycle, monitor what they do, control their access and hold them to a governance and compliance regime. The framing was explicit. Workday's chief product officer for HCM, David Somers, described digital workers as "joining human workers as one combined workforce", needing "a unified platform to manage them as one".
The objection arrived quickly, and it was not unreasonable. Writing at diginomica, Phil Wainewright argued that "AI agents need careful management and governance. But that's a job for IT, not HR," comparing the logic to putting cash machines or self-service checkouts under HR simply because they displaced jobs. ServiceNow has since pushed its own agent management strategy, Boomi has built an agent registry, and AI governance specialists such as Holistic AI and Credo AI are positioning around the same territory. The category is contested, which usually means the underlying question has not been stated precisely.
Here is the more precise version. "Should agents be in the HR system" is not one question. It is three, and they have different answers.
Three obligations, three owners
Every record in an HR system exists to discharge an obligation. Working backwards from the obligations makes the ownership question tractable.
Visual 1 — What an agent record would have to carry, and who already owns that obligation
Obligation | Applies to an AI agent? | Natural owner |
|---|---|---|
Employment law: contract, working time, right to work, protected characteristics, grievance and dismissal process | No. None of it attaches to software. | HR — but there is nothing here to administer for an agent. |
Pay, benefits, tax and social contributions | No. An agent has a cost, not a compensation package. | Finance and procurement, as with any software licence or consumption charge. |
Identity, entitlement, access review, audit trail, decommissioning | Yes, and urgently. | IT and security. These are service-management and identity-governance disciplines with mature tooling. |
Establishment: job architecture, headcount plan, span of control, capacity by function | Yes — and this is the one currently unowned. | HR, and nobody else can do it. |
How to read it: On the first three rows the critics are right — the record itself is an IT and finance artefact. The fourth row is where the argument has been losing something, because it is a workforce planning obligation that only the People function is positioned to discharge, and it does not require owning the agent's record at all.
Read that way, the debate resolves into something less dramatic than a turf war. The agent's record — its identity, permissions, activity log and retirement — is IT's, and should sit where machine identity already sits. What is not IT's, and is not currently anybody's, is the question of what an agent does to the establishment: to job architecture, to the shape of a team, to the headcount plan submitted in the autumn for the following year.
The system of record is the small question. The larger one is whether next year's workforce plan can account for capacity that has no employee number, no cost centre in the payroll, and no line in the establishment.
HR has already lost this argument once
There is a precedent, and it is uncomfortably close.
When contingent labour scaled, the same reasoning applied: contractors are procured, not employed; the record belongs in a vendor management system; HR administers employees. The logic was defensible and the outcome was a workforce a substantial part of which the People function could not see, plan for, or report on — a visibility problem that remains largely unowned two decades later, and one this publication has covered directly.
Agents are on the same trajectory, faster. They are procured by function heads, expensed as software, provisioned by IT, and deliver work that was previously in a job description. If the People function concludes from the ownership debate that agents are simply not its concern, the same gap opens — and it opens in workforce planning, which is precisely where the data is already thinnest.
What "counting" an agent actually requires
The practical difficulty is that agents resist the units workforce planning is built on. A person is one head, in one job, in one team, at one location, for a period of time. An agent is a capability that may run across four functions, scale to zero overnight, be duplicated eleven times by different teams without central knowledge, and be retired by a vendor deprecating a model.
That is a real modelling problem and not a reason to defer. The workable approach is to stop trying to express agents as headcount and instead express them where they actually bite: as capacity against a named function, sourced somewhere other than the establishment.
Questions to settle before the next planning cycle
Who signs off the deployment of an agent that performs work described in a job description? If the answer is a function head and a software budget, the establishment has been changed without a decision being recorded.
Can we produce a list of agents in production by function? Not by vendor, by function — that is the view that maps to the workforce plan, and the one nobody generates.
Where does an agent's work show up in capacity planning? If nowhere, the plan is describing a workforce that no longer matches the work.
Who decommissions an agent, and does anyone tell HR? A capability withdrawn without notice is a capacity loss against a function, and it will present as a resourcing problem months later.
Are we accepting a system of record because it answers a governance question, or because a vendor offered one? The identity and audit case is strong; the HR-ownership case has not been made, and buying the platform does not settle the establishment question in any event.
What this changes
For an HRIS owner being asked to evaluate an agent system of record, the honest reading of the evidence is that the identity, access and audit case belongs in the IT and security stack, and that adopting an HR-branded version of it does not by itself buy anything the organisation would not otherwise have. That is a procurement judgement, and it is not this publication's to make.
The judgement that does belong to the People function is separate, and it is being crowded out by the vendor argument. Someone has to decide how work performed by non-human capability appears in the workforce plan, who authorises it when it substitutes for a role, and how it is reported to a board that will ask about productivity per head. None of that requires owning the agent's record. All of it requires HR to stop treating the ownership debate as the whole question.
Sources and method. An HRHubsMedia original. Workday's Agent System of Record announcement (13 February 2025), its scope — agent onboarding, lifecycle management, monitoring, governance, compliance and access control — the quotation from Workday chief product officer David Somers, the counter-argument and quotation from Phil Wainewright, and the competing moves by ServiceNow, Boomi, Holistic AI and Credo AI, per diginomica. The three-obligation framework, the contingent-workforce comparison and the planning questions are HRHubsMedia's own analysis; readers may wish to read this alongside our earlier coverage of contingent workforce data visibility and of workforce planning inputs. Vendors are named as market context, not endorsements. Journalism, not procurement advice. Corrections will be made openly on this article.
