The AI HR Helpdesk Test: Five Questions That Separate Answering From Resolving
Employees raise 3.6 HR needs a month and only 6% get an instant answer. Five design questions decide whether an AI helpdesk resolves those requests or only answers them.
In the last days of July, one of the big service-desk platforms opened early access to employee-service AI agents that read across seventy-plus enterprise sources and answer people in Slack and Teams (UC Today). It is not an isolated launch. The ITSM suites and the HCM vendors are shipping their own versions this year, and the trade press has already started asking whether the employee portal survives the shift. By your next renewal, an AI HR helpdesk will be a line item on most quotes you receive.
The buying question is smaller than a choice between logos: when someone asks how parental leave works for their contract, in their country, what happens after the agent produces a paragraph?
The queue is bigger than your ticket count
Most HR teams size the problem by what lands in the ticketing system. That count is wrong, because most requests never become tickets.
Research Applaud published with Censuswide in January 2026, covering 1,000 employees at UK organizations with more than 2,000 staff, found that 79% ask HR for help at least once a month, averaging 3.6 needs per person per month. Only 6% get an instant answer through AI or chat. 36% wait a full day or longer. Employees sort out fewer than half of their own HR needs, 47%, without a person getting involved. Applaud puts the drag on a typical 1,000-person company at roughly 12,800 hours a year, about $385,000 (Applaud/Censuswide, January 2026).
Those hours don’t sit in one place. They sit in a colleague’s DM, a manager’s inbox, a Slack channel where someone guesses at the travel policy, and the four minutes an HR business partner loses to a payroll question during a succession conversation.
Meanwhile HR has plenty of AI. Culture Amp’s study of 264 HR professionals, released 22 July, found 86% using it for content creation and 83% for brainstorming, while fewer than 40% of HR operations have moved into any kind of AI automation and a third of the professionals surveyed run agentic workflows (HR Dive). HR is using AI to write. Two thirds are not running agentic workflows at all.
Why another portal won’t close the gap
The instinct is to buy a better front door: a cleaner portal, a smarter search box, a chatbot on top of the handbook. Employees have already given their verdict on the portal by not visiting it, the pattern our Agentic HR Academy calls the portal nobody visits.
Our COO Shlomo Weiss put the same idea on stage at Irresistible this June: “We do not want to be system number 146.” A helpdesk that requires the employee to know it exists, log in, and phrase the question in the words the index expects has already lost most of the volume it was bought to absorb.
Five questions to ask before you buy an AI HR helpdesk
1. Does it know who is asking?
“How much notice do I give?” has a different answer for a salaried employee in Germany, a shift worker in Ohio, and a contractor in Singapore. A retrieval system that ranks handbook paragraphs by keyword will hand all three the same text and get two of them wrong.
You need an agent that resolves the person first and the policy second: their entity, contract type, jurisdiction, benefits plan, manager, tenure. Ask any vendor to demo the same question asked by three people on three different contracts, and to show you where each answer came from.
2. Can it act, or only answer?
This is the line most tools have not crossed yet. Look closely at what actually ships in the early-access programs opening now: conversational search and answers come first, while the layer that completes an HR or IT task follows later in the rollout (UC Today).
Answering “you have 11 days left and here is how to request them” still leaves the employee with a form to find. An answer moves information, while a resolution moves the work off someone’s desk: the agent files the request into Workday or SuccessFactors and confirms it back to the person who asked. The academy piece on autonomous action sets out where that boundary sits. In our own product it lands here: Gloat’s HR Helpdesk Agent indexes the handbook, benefits plans and jurisdiction rules, resolves who is asking through the Knowledge Graph, then writes PTO requests, benefits changes and address updates back into the HCM, citing the source document each time.
3. Does it live where the question gets asked?
People ask HR questions in the place they are already typing. The Applaud research found 26% of employees go to an HR system or portal first, and another 24% reach for email, phone or messaging. If your agent answers in Teams, Slack, Copilot and email, you capture the DMs and the manager pings that never reached the queue. If it answers in a portal, you capture the fraction of people who were going to file a ticket anyway, which is the volume you could already see.
4. What is it not allowed to do on its own?
Some requests should never resolve without a person involved. A pay discrepancy, a leave request tangled up with a medical condition, an accommodation question, anything that smells like a grievance: the correct behavior is a fast, well-briefed handoff to a human with the full thread attached, not a confident paragraph.
Write that list before procurement, not after go-live, and decide where each item sits on the spectrum from suggestion to action that our academy piece on governed autonomy lays out. Then check three things in the product: the escalation carries context so the employee never repeats themselves, every action leaves an audit trail, and permissions are enforced at retrieval so the agent cannot surface a document the person was not entitled to see.
5. How will you know it worked?
Deflection rate is the metric every vendor leads with, including ours, and it is the easiest one to flatter. A conversation that ends because the employee gave up counts as deflected.
Track four numbers instead, with a baseline captured before launch:
- Resolution without human touch, measured against tickets that stayed closed for 14 days.
- Reopen and re-ask rate, which is where a plausible wrong answer shows up.
- Time to answer, split by question type, so a fast benefits answer doesn’t hide a slow payroll one.
- Hours returned to the HR team, sampled the honest way: ask the team what they stopped doing.
Measuring agent outcomes goes further into the baseline question. The Josh Bersin Company’s HR 2030 blueprint, published 8 June, is worth reading alongside it: it projects the strategic share of HR’s work rising from around 30% to as much as 75%, and cautions against agent sprawl as systems scale (The Josh Bersin Company).
The point of taking the queue away
Heather Yurko, an HR leader speaking at Irresistible this June, described what changes when the routine work stops arriving in their queue: “I am not actively digging. I am having data and recommendations brought to me, and as a people person, I can do people work.”
The business case is the restructuring conversation your HRBP had time to prepare for, and the manager who got coached instead of ignored. Be deliberate about what the reclaimed hours become, because an HR team that automates tier-1 and fills the gap with more reporting has bought a faster version of the same job. Neither the restructuring prep nor the coaching shows up in a deflection report.
Start with your top ten question types. Ask a vendor to resolve them end to end, for three employees in three countries, in front of you.
