Walk through a typical termination from a manager's perspective. Calculate severance. Find the policy document. Look up the correct termination code. Navigate to the right Workday business process. Hope the right approvers are in the right roles. Email HR when something does not work. The whole point of manager self-service was to keep this work out of HR's queue. In practice, HR is still in the queue, because the system was never designed for the complexity of the actual task.
Performance cycles are worse. The process consumes enormous time. The real value-add work happens outside the HRIS, in emails and side conversations and one-to-ones. The system captures the form, not the work. Staffing events disappear HR teams into spreadsheets for weeks. Reorganisations live in PowerPoint until somebody types them into Workday. The HRIS is the system of record. It is rarely the system of work.
What the data actually says
Industry benchmarks have been consistent for years. Self-service tools resolve fewer than half of employee HR needs. Only about 25 percent of employees initiate HR queries through official systems at all. The rest goes to direct messages, hallway conversations, and the HRBP's inbox. The cost of self-service was never that it eliminated HR work. It was that it added a layer that did not eliminate the work it was meant to displace.
What agent-assisted HR actually does
A true agent, not a chatbot, can do four things a self-service portal cannot. Reason through multi-step processes. Pull data from multiple systems. Apply policy logic and perform calculations. Execute transactions while keeping humans in control of key decisions. The combination is what changes the model.
A useful comparison. In the self-service model, an employee asking about saving PTO for December lands on a knowledge base with 14 policy articles, picks the one that looks most relevant, reads it, and tries to map it back to their own situation. They either get it right or they end up in the HRBP's inbox. In the agent-assisted model, the same question gets parsed: the agent already knows the employee's country, contract type, current balance (120 hours), accrual rate, and the entitlement rules, returns a tailored answer ("if you take five days in December you will end the year with 32 hours carry-over, which is within policy"), and offers to file the request. The interface looks similar in both cases. What is different is where the reasoning happens. In one model the employee does it. In the other the system does it and the employee approves the result.
The infrastructure problem that is underdiscussed
Roughly 80 percent of the context needed for HR processes exists outside the HRIS. It lives in Excel spreadsheets, SharePoint, email threads, and institutional knowledge that no system holds. Agents that only operate within Workday cannot deliver agent-assisted HR. They can deliver agent-assisted Workday, which is something different and much smaller.
This is the part that breaks naive deployments. A "policy agent" that retrieves from clean, structured Workday content works in a demo. The same agent in production hits a question that depends on a country-specific exception buried in a SharePoint folder that nobody has touched since 2022. The agent answers confidently. The answer is wrong. Trust erodes.
Effective agents have to operate across the landscape, not just within the system of record. The retrieval layer matters more than the model. The content cleanup matters more than the prompt design. Skip either and the agent feels confident and unreliable.
What changes across the HR service tiers
Tier 0 (self-service) becomes active resolution rather than passive knowledge. The agent does the work, with the employee in the loop, instead of pointing the employee at the work and walking away. This is the layer where the shift is most visible to employees on day one, and it is also the layer where the trust dynamic is fragile: a single bad answer in the first week reshapes how people use the agent for a year.
Tier 1 (shared services) shifts from volume processing to exception management. Transactions still flow, but they flow through agents. The shared services team handles the cases the agent cannot, and reviews the cases the agent did.
Centres of Excellence evolve into design hubs. The COE defines the decision frameworks, the policy logic, the human-in-the-loop checkpoints. The agent executes them at scale.
HRBPs potentially transition from 85 percent tactical work to something closer to genuine strategic advisory. This is the one most leaders are sceptical about, and it is the one with the biggest organisational impact if it lands. The HRBP role has been tactical for so long that a real shift is harder culturally than it is technically.
“The risk that catches most rollouts is trust. Employees who distrust or avoid the agent quietly create more shadow processes than they already had, and the agent never sees the workload it was meant to absorb.”
What predictions and warnings are credible
Gartner has predicted that by 2030, 60 percent of HR tasks will be completed through AI agents or LLM interfaces, with HR staffing potentially reducing by up to 30 percent while service quality improves. Treat the number as directional. The shape of the prediction is consistent with what we are seeing across implementations.
Gartner has also predicted that over 40 percent of agentic AI projects will be cancelled by the end of 2027. Most failures are not technical. They are change management failures. The pattern that repeats across early enterprise rollouts: an HR organisation shuts down its traditional support channels in favour of the agent before the agent is reliable enough to replace them. Employee satisfaction collapses. The agent gets blamed, the technology gets blamed, and the rollout gets reversed. The technology was working. The transition was not.
What separates the leaders
Across the early rollouts that have landed cleanly, a few habits show up consistently. The leaders treat agent deployment as work redesign rather than technology deployment; the agent is the easy part and the redesigned process around it is where the actual work lives. They run a change-management stream from week one, at the same seniority as the technical stream, instead of bolting it on toward the end. HRIS functions as the architect and translator in their operating model, owning the platform operating model and the governance while HR owns the experience and process design and IT owns the platform itself. The combination is what holds the rollout together when the first messy edge case shows up in production.
The one thing to start with
If there is a single move worth making in the next month, it is to map your end-to-end workflows, including the spreadsheet and email steps. The map of what actually happens is more useful than the map of what Workday thinks happens, and most HRIS teams only have the second. The map is what surfaces where an agent is plausible, where it is premature, and where it cannot help until the underlying process is fixed. Without it, the agent conversation drifts into demos. With it, the conversation gets uncomfortable, which is the sign it is the right one.
Two pieces of work follow naturally from the map. Own the operating model for agent deployment yourself, rather than waiting for Legal to write a policy first (teams that wait usually wait forever). And start the AI literacy work across the HR function now: not deep ML knowledge, just practical literacy on the difference between a chatbot and an agent, where retrieval matters, and where human-in-the-loop is non-negotiable. The HR function that does not understand the technology will not own the operating model around it, and somebody else will.
