There he goes. One of God’s own prototypes. A high-powered mutant of some kind never even considered for mass production. Too weird to live, and too rare to die.
Hunter S. Thompson wrote that in Fear and Loathing in Las Vegas. You could be mistaken reading it in a modern context that he was actually talking about AI we contemplate whether it’s conscious (he wasn’t).
At HR Tech, a show about human resources and technology, we were heavy on the AI and light on anything traditional tech or human resources.
Too weird to live. Too rare to die.
Thompson ended his trip with about 36 hours alone in a hotel room, writing in a notebook. I spent my 36 hours in meetings and at social events, looking for the places where the AI talk had turned into work.
Thanks to Nick, Dylan and Nicole from 3Sixty Insight, who were on site with me for this bender. Thanks also to our sponsors, Insperity, Paylocity, Tavio and Lytiqs.
Before the trip I wrote down four questions I wanted answered. Where are AI agents doing real HR work? Are companies building talent or just buying and cutting it? How do you know a candidate, and the data about them, is real? Who owns pay now?
I got answers to the first three.
Nobody in the meetings I took brought up who owns pay. I also came home with four answers to questions I hadn’t asked.
Where the Work Is Real
What’s happening today? Some early action is happening but it is still very early.
1. Agents are in production at a handful of companies, with more to come
SAP told us it’s putting agents into production with a small group of marquee customers it announced at SAP Connect earlier this month. ERP Today names IBM and John Deere among them. SAP said every analyst and every customer is asking whether the agents are in production. SAP is also its own customer for workforce planning, across roughly 100,000 employees.
Like my experience with many of Workday’s announcements, Alight‘s assistant isn’t live yet. Its new employee experience launches in April and replaces site search with a conversation that can complete transactions. Alight told us 80% of people pick their benefits in the last weeks of enrollment and come back about seven times a year, usually when they need something.
An employee who hurts a knee gets one suggested action instead of twelve. In Alight’s example, that’s a visit to Hinge Health and a look at how much of the deductible they’ve used. Alight said HR can’t see the employee’s condition.
2. One company is testing its own people before it decides who to hire
Vervoe told us about a Fortune 500 transportation company whose support jobs are shifting from handling chats to orchestrating agents. It has put 8,000 employees through an assessment for the new role. It’s now extending the assessment to its outsourced support teams and to hiring, because hiring people for today’s version of the job makes little sense if it will be gone within a year. The company hasn’t decided what to do with the results. Vervoe sees a redeployment opportunity in them.
Cornerstone‘s labor market data shows demand for what it calls human skills running two to three times demand for technical skills. Its example was natural language processing, which is in demand and which marketers tend to be good at. Cornerstone’s argument is that once you break jobs into tasks, you can see which parts to automate and which parts set the company apart.
3. Candidate fraud gets flagged, and the customer decides
Vervoe also flags plagiarism, LLM use and location mismatches, like a candidate who says Texas and takes the test from India. The customer decides what to do with each flag. Vervoe couldn’t say how many flagged candidates are actually cheating, because the customer resolves them.
Its advice is to tell candidates up front whether they can use AI and whether they have to disclose it. Without clear rules, any use of AI looks like cheating. It also says to test tasks rather than knowledge, since knowledge questions are easy to cheat on.
The labor market data that candidates and jobs get measured against has a similar problem. Lightcast said many job descriptions are now written by AI and answered by AI, so it’s putting more work into the quality of its data. An LLM, it said, gives a different answer every day about what a job’s tasks are, so Lightcast keeps a taxonomy of about 26,000 task statements to benchmark against. One customer pipes that data into its own system over MCP as an evidence trail.
Deployment Blockers
What keeps people from jumping onto the AI bandwagon? Maybe some doubts but also some real challenges.
4. Vendors can meet buyers where they are and still show them the future
Darwinbox told us it does both. A demo of its AI platform shows where buyers can get to. Then the buyer’s questions start. How do I budget for this? How do I work with IT? How do I train my team? Darwinbox said the most mundane things are what stall a deployment, and that HR IT needs a new kind of role to handle them.
Its example was a manager who can change the leave policy for all 20 people on a team without seeing how it changes their lives. Darwinbox said it looks for gaps like that and closes them through product changes and forward-deployed engineers. It launched Cortex in August.
Thompson’s novel has a police “drug expert” who gives a conference presentation the book plays as comically out of touch. A lot of vendor pitches on the show floor landed that way. They described a future state to buyers who are still working through the basics, and they leaned on fear of falling behind to close the gap.
5. Customers keep finding out their data is wrong.
SAP’s product leaders told us customers testing AI first blame hallucinations. Then they find the data is wrong. A second system doesn’t have the same numbers, which means a process is broken somewhere. SAP is building agents that catch anomalies and data quality problems for that reason.
I heard this from a number of buyers at both HR Tech and Workday Rising. It’s not any vendor or any person’s singular fault. One integration vendor lamented that a lot of “bad data” is actually inconsistently formatted data. In one project they used as an example, they said that they got address mismatches for every single employee. One system had two address lines and the other didn’t. And that’s really the most basic of data challenges.
Vendor Differentiation
There was a time when it was easy to tell a performance management vendor from an HR analytics provider. Now, there is more blur and overlap than ever.
6. Vendors are working to sound different from each other.
A Prism leader told us that about two years ago, the company read phrases from competitors’ websites to its sales team and asked which ones sounded like Prism. The takeaway was that none of the vendors in the space differentiate well. This year Prism brought its marketing team to the expo to see what the market already offers. The rule is that if the market already has a feature, don’t announce it. Deliver it. Prism also fed its product list and the exhibitor list to Claude to pick which booths to visit.
Tavio told us briefings with the big analyst firms were all about orchestration, so it shifted its messaging that way. Then it read its own website and didn’t understand it. It went back to describing what it does, which is reusable integrations that software vendors deploy across many customers. Tavio also said none of its customers have asked about MCP. In the last three to six months, a lot of vendors have announced it anyway.
7. Buyers need help proving the value.
Paylocity said customers call its AI approachable because it doesn’t ask them to become data scientists. It wants to show value that goes beyond time saved. Better candidates and retention are the examples it gave. It said it may not have cracked that yet.
Multiplier said an HR leader or CFO on a first call often can’t say how many contractors the company has across the six or seven countries where it employs people. A company with five or six local payroll vendors can pay high FX fees on $25 to $30 million in salaries without noticing.
Multiplier’s hardest problem now comes after the sale, in how long customers take to see value. It staggers some rollouts, starting with a customer’s top two countries. Others work better all at once. Either way, it doesn’t want a six-month project.
Reaching Buyers
We know that B2B software vendors are seeing longer sales cycles, fewer leads and more no-decision outcomes. Three things we heard show why.
The buying group changed. Cornerstone and Vanson Bourne surveyed 500 IT decision-makers, 500 HR decision-makers and 1,000 employees. They found that only 35% of HR and IT decision-makers make AI decisions jointly. Organizations where the CHRO and CIO co-own workforce readiness act 13% faster.
Darwinbox told us it now sells to both HR and IT, and said little has been published about how the two should work together. It pointed to Moderna, where the head of HR is the Chief People and Digital Technology Officer. Lightcast said the buyers for its AI impact analysis mostly sit in IT or operations, working alongside the head of HR.
Some buyers aren’t at the show. Prism told us all of its business is indirect, through resellers, and it expected to see only three, four or five of them in Vegas. It came to learn about AI pricing and product gaps.
Some buyers show up only when something breaks. Multiplier, which says it passed $100 million in annual recurring revenue in September, told us about half of its global payroll customers come because they can’t consolidate their vendors. Something has usually broken first, either a breach in their patchwork of vendors or growth in a country where the local provider couldn’t keep up.
Two Requests for Vendors
If I may be so bold, I have two ideas that could help vendors stand out better at these types of shows.
- Please start from a problem your buyer already has. Alight told us a major consulting client spent $40 million on GLP-1 drugs last year, and that many employers chose this fall to stop covering them. The Business Group on Health found that employer coverage of GLP-1s for obesity fell from 72% in 2025 to 60% in 2026. Alight’s answer is data on which programs work. One of its clients saved $2.6 million by cutting the ones the data showed weren’t effective.
- Please also ask why. A SAP customer asked its product team how to route 360 reviews from one place to another, which the product leader told us was a configuration question. The team could sort that out offline. Then the leader asked what the customer wanted from the reviews in the first place. That opened a bigger question.
Thompson’s wave passage describes a generation that looked headed for certain victory until the wave broke and rolled back. Don’t announce an agent because a competitor did. Don’t add AI to a product because you’re afraid of looking behind. Find the thing your buyer is stuck on and fix that.