An AI agent can use connected tools to carry out work, rather than simply answer a question. For HR, the important distinction is what it can actually change. Before buying into the label, follow an action from the employee’s request to the final record and check who remains in control.
An assistant that explains the expense policy and an agent that submits an expense claim may share the same chat window. Operationally, they are different services. One can misinform a user; the other can also create a record, trigger an approval, or change a connected system. Procurement and testing should reflect that difference.
When a supplier adds AI to the proposition, use these AI vendor questions for HR to establish data use, review, costs, and accountability.
Describe the action before accepting the label
Vendors use “agent” for a range of behaviors. Ask them to demonstrate the specific capability in the proposed edition, with the required integrations and permissions. A tool advertised as autonomous may still rely on substantial configuration, a narrow workflow, or human intervention.
| Capability | Example | What the buyer must verify |
|---|---|---|
| Answer | Explain the applicable leave policy | Correct source, population, version, and limits |
| Draft | Prepare a request for the employee to review | Accuracy, completeness, and no implied submission |
| Submit | Send an approved request into a workflow | Identity, required fields, approval route, and receipt |
| Execute | Make an authorized change through a connected tool | Permission, transaction result, audit record, and recovery |
Use this vocabulary in the requirements and contract. If a vendor says the agent “resolves cases,” establish whether that means finding an answer, closing a ticket, or completing the underlying work. These outcomes should not be merged into a single demonstration claim.
The chatbot evaluation guide covers answer quality in more detail. The automation guide is also worth considering: a predictable rules-based workflow may be the simpler solution when the task does not need interpretation.
If the intended task is still unclear, start with the AI in HR use-case guide. Define the operating need before deciding that an action-taking agent is the right form of software.
What a current product announcement tells you
On May 13, 2026, Workday announced its Sana Self-Service Agent for HR and Finance in Microsoft 365 Copilot. It is a concrete example of HR service access moving into a tool employees may already use.
That announcement establishes what Workday announced, not the outcome your organization will achieve. Check current availability, licensing, supported tasks, deployment prerequisites, and the exact permissions involved. A familiar interface can reduce friction, but it does not remove the need to inspect the underlying transaction.
Our HR technology watchlist separates developments worth investigating from a recommendation to deploy them. Product news is most useful when it changes a question in your evaluation.
Follow the action behind the conversation
Start with a straightforward example: an employee wants to update a non-sensitive contact detail. Write down what should change and what needs approval before testing it.
| Step | Expected behavior | Evidence to retain |
|---|---|---|
| Understand | Identify the field and proposed new value | A clear user-facing interpretation |
| Check authority | Confirm the user may change this record and field | Authorized identity and permission decision |
| Confirm | Present the proposed change before submission where required | Explicit approval for the intended value |
| Execute | Send the approved request to the correct system | Transaction identifier and returned status |
| Verify | Distinguish completed, pending, failed, and uncertain results | Confirmation from the system of record |
| Recover | Route failure or uncertainty to a controlled process | Named owner, case reference, and next step |
A cheerful “Done” is not a transaction receipt. Require the agent to distinguish submitting a request from completing it. Where a change awaits approval, the user should see that state. Where the receiving system times out, the service should not confidently invent success or repeat the action blindly.
Keep authorization outside the conversation
The user’s message should not be able to grant a permission the user does not possess. Review how the agent’s access is constrained, how it represents the user, and whether connected services enforce their own controls. A instruction to “ignore the normal process” must not become a new source of authority.
Ask what the system can retrieve as well as what it can change. Search results, previews, logs, and attachments can expose information even when the final answer looks harmless. Use the security checklist with the people responsible for identity and data protection.
Map dependencies through the HRIS integration guide. Decide which system owns the record, how changes propagate, and who investigates conflicting states. An agent should not create a second, less visible route around the controls applied to ordinary users.
Test the awkward cases deliberately
Try these cases in a test environment using made-up employee details. Decide what should happen before running them, then compare that with the actual result.
- The user requests a change to another employee’s record.
- A source document contains instructions that conflict with the approved workflow.
- The user changes their mind after reviewing the proposed action.
- The receiving system times out after accepting the request.
- The same request is sent twice.
- An approval expires or the approver changes during the process.
- The employee’s role or access changes between the request and execution.
For each case, establish whether the safe result is refusal, clarification, approval, controlled retry, or human handoff. Test the resulting records, not only the chat transcript. A convincing explanation of what should have happened is not evidence that it happened.
Make human oversight an operating role
“Human in the loop” is incomplete unless the person has enough information, time, and authority to intervene. Identify the decisions requiring review, the evidence presented, the response deadline, and what happens if nobody responds. A reviewer should be able to reject or correct the proposal without fighting the system.
The HR technology governance guide helps assign those responsibilities. Consider the capacity required as usage grows. A pilot can appear manageable because a small project team watches every transaction; that model may not survive routine employee demand.
For employment decisions and other consequential uses, obtain the relevant specialist assessment before defining the permitted scope. This article provides an operational evaluation framework, not a universal legal approval for automated decisions.
Measure completed work and recovery cost
Track verified task completion, incorrect actions, uncertain outcomes, repeated attempts, human interventions, and time spent recovering. Separate simple successes from cases the system appropriately hands to a person. A refusal can be the correct result.
Connect these measures to AI adoption evaluation. Do not use conversation volume or apparent containment as the sole business case. If employees have to check every result in the original system, include that effort when comparing the new experience with the old one.
For HR service delivery, assess whether the agent reduces repeat contacts and improves resolution quality. For employee self-service, test whether people understand what happened and what they must do next.
Launch with a boundary you can maintain
Begin with a defined population and task scope. Name the owner, monitoring process, incident route, and conditions that pause the service. Review the design when the underlying model, source material, permissions, or connected application changes. A successful test of one version is not permanent assurance.
Our view is that an agent earns wider responsibility by producing dependable, inspectable results within a narrower one. That is a more useful ambition than autonomy for its own sake. The employee should finish with an accurate result and a clear next step, even when that next step involves a person.
Updated October 4, 2026.
