Part 3: HR's Own Work With AI

How HR uses AI on itself: Service chatbots, onboarding, learning and policy drafting, what goes wrong and who answers for it, and how a team starts with AI literacy, a bot that says it is a bot, and one named owner. Closes the AI and HR series.

Tue, 10 Nov 2026 · 16 min

Made for you, by AI, from the sources below.

Listen to it in AI and HR

Transcript

The script the two voices were given, with the sound cues taken out. A voice may say a word a little differently, so the audio is the episode.

Mia: Welcome to Walking Points. IBM says it informed about twenty-one thousand first-line managers that HR help would only come through one chatbot. An internal score then dropped to minus thirty-five.

Adam: How bad is that, on a scale that most people haven't seen?

Mia: It's a satisfaction-style score, and the two sources that report it name the metric differently, so I won't pretend to know the exact scale. But the direction is not in doubt. Managers were forced onto a tool that didn't work yet, and they got angry.

Adam: So what is this episode actually about?

Mia: It's about HR's own work with AI, meaning the tools HR uses on itself, not on job applicants. This part was made in October 2026. It's the third and last part of our series on AI and HR, and it earns a whole episode because this is the one area where HR is both the user and the department that has to answer for it. We'll walk through three ideas in order: four jobs AI already does in HR, what goes wrong and who answers for it, and how a team can start responsibly. Then we'll look at what happens to the HR role itself.

Adam: And where did the series leave off?

Mia: Last Tuesday we covered AI and the people a company already employs, including the EU ban on emotion recognition at work.

Adam: Okay. So, four jobs. What's the first one?

Mia: The best documented one, the HR service chatbot. It's an assistant that employees and managers ask instead of calling or emailing HR. How many days of leave do I have left, how do I change my bank details, what's the parental leave policy. IBM has the most thorough public description, in its own case study of a system called AskHR. IBM started with about twenty-five separate HR chatbots and found that the more it added, the harder it was for people to get a clear answer. So it merged them into one front door.

Adam: And that front door does more than answer questions?

Mia: Yes, it carries out transactions in HR systems, like employee transfers or employment verification letters. IBM says there are close to ninety thousand transfers a year. And in 2025, IBM reports, it handled more than sixteen million user messages. That's about thirty-nine thousand a day, or roughly fifty-one interactions per employee per year. Adam, how much weight would you put on those figures?

Adam: Not much as proof, I'd think, since IBM sells this product too. They tell us what's possible, but they aren't an independent measurement.

Mia: Exactly, they're IBM's own figures about its own product, so read them that way.

Adam: Fair. Then what happened with that early score?

Mia: That's the most useful part of the story. The case study reports the early low of minus thirty-five and a score of plus seventy-four now. The fix, according to IBM, was iteration, and its own lesson is to find a small pain point, automate it, make it great, and only then scale. It's the lesson that early score taught, since managers had been pushed onto a tool that wasn't ready yet.

Adam: Right. So the second job is onboarding?

Mia: Yes, and here the evidence is thin. A search for controlled studies of generative AI in new-hire onboarding found mostly vendor and consultancy claims, like a process shortened from three hours to thirty minutes. No independent measurement of time to productivity or retention.

Adam: So there's nothing solid at all?

Mia: Nothing on onboarding itself. The closest solid evidence is from neighbouring work. Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied 5,179 customer support agents. Access to a generative AI assistant raised the number of issues resolved per hour by about fourteen percent on average. For less experienced and lower-skilled workers, it was about thirty-four percent, while experienced workers saw minimal impact.

Adam: Hold on, though. That's customer support. Why does it say anything about HR?

Mia: It doesn't, directly. That's a reasoned analogy, not a finding about onboarding. The authors suggest the tool spreads the habits of the best performers to newcomers. So there's a fair reason to think an assistant that answers a new hire's hundred small questions in the first weeks could help most the people who are newest. But nobody has measured that.

Adam: Okay, I'll take it as a hypothesis. Third job?

Mia: Learning. Vendors say AI personalises learning paths, recommends courses from role and career goals, and flags skills gaps. Independent evidence is scarce again. One recent review of the claims found figures like course completion up forty percent with no named study behind them. A more candid vendor-side observation is that today's tools mostly recommend courses based on completions, job roles and stated goals. They don't adapt to how a person learns in real time.

Adam: So what should a team do with a number like that?

Mia: Ask for the method behind it, and ask for a pilot with a comparison group, so you can see what changed against people who didn't get the tool.

Adam: And the fourth job is drafting?

Mia: Policies, job descriptions, handbooks, letters, announcements. HR is a document-heavy function. A Gartner survey of 105 HR leaders, run in June 2023, found that the most prioritised uses of generative AI were HR operations tasks like policies and document generation, and recruiting tasks like job descriptions. Only five percent of HR functions had implemented it at that point. Drafting is where the technology is strongest, and where the risk is easiest to underrate.

Adam: Why is it easy to underrate? A draft is just a draft.

Mia: Because a draft that reads smoothly gets published. That's the pattern in the two cases I want to walk through next. In both, the real question wasn't whether the bot was wrong. It was who was supposed to be checking.

Adam: Go on, then. First case?

Mia: February 2024, a tribunal in British Columbia, in Moffatt v. Air Canada, 2024 BCCRT 149. Air Canada's website chatbot told a grieving passenger he could claim a bereavement fare retroactively. The airline's actual policy didn't allow that. Air Canada argued the chatbot was a separate entity responsible for its own statements.

Adam: Wait, it argued the chatbot was its own entity? Did that work?

Mia: No. The tribunal member, Christopher Rivers, called it a remarkable submission. He found negligent misrepresentation, because the chatbot was part of the airline's website, so the airline answered for what it said. This is a Canadian tribunal decision about customers, not employees, and it isn't binding elsewhere. And to be clear, this isn't legal advice, just reporting. But it's the clearest signal of the general principle: An HR chatbot that gives an employee the wrong notice period or the wrong entitlement is the employer speaking.

Adam: And the second case?

Mia: New York City's MyCity chatbot, which The Markup investigated in the spring of that year. It told business owners they could take workers' tips, which they can't, and that landlords could refuse tenants paying with housing vouchers, which they generally can't. The city kept the bot running, called it a beta product, and added a notice that its answers shouldn't be used as legal or professional advice.

Adam: So a disclaimer is the fix?

Mia: Well, that's what the city chose. But look at what both cases share. A chatbot sounding certain about rules, and nobody checking its answers against the actual policy.

Adam: Mia, hold on. Is there any actual rule that makes HR responsible here, or is this just case law from far away?

Mia: There are two rules in the EU AI Act, and the Act is the European Union's law on artificial intelligence. The first is AI literacy, Article 4. According to the European Commission's own questions and answers, providers and deployers of AI systems must take measures to support AI literacy among their staff, and others who operate AI systems for them, like contractors. They have to take account of technical knowledge, experience, education and the context of use.

Adam: That sounds heavy. Does everyone need a certificate?

Mia: No. The Commission page says the obligation doesn't require guaranteeing a specific level for any individual, doesn't require measuring staff's knowledge, requires no certificate, and mandates no AI officer or governance board. It says relying only on instructions for use is often ineffective, and that an internal record of the training and other initiatives is enough. The article has applied since the second of February 2025, and the page says enforcement by national authorities begins in August, though it gives two different days, so check the official text.

Adam: And HR is who gets asked to deliver it.

Mia: Usually, yes. The standard is reasonable measures that fit the audience, written down.

Adam: What's the second rule?

Mia: Telling people when an AI is talking. According to the Commission's AI Act page, the transparency rules started to be enforced on the second of August. Under Article 50, chatbots and other AI systems that interact with people must disclose that they're AI, unless that's obvious from context. So an HR chatbot that answers employees has to say it's a bot.

Adam: But I thought some of this got delayed.

Mia: Some of it, yes. The Digital Omnibus package, which is a set of changes to the Act, moved the high-risk use cases listed in Annex III, employment among them, to the second of December 2027. It moved the high-risk systems built into products to the second of August 2028. But the Commission says the transparency rules weren't delayed. For anything where AI makes or shapes a decision about a person, the high-risk rules from Part 1 are the relevant ones, and those have the later date.

Adam: So the chatbot's disclosure is already live, and the decision-making rules are still ahead.

Mia: Right. Now, how does a team begin? The governance picture is the weak point. Ted Flett, a Canadian employment lawyer, wrote a report on SHRM's latest research for HRD Canada on the second of September. SHRM is a large HR professional association. Fewer than half of organisations have an AI policy at all. Where one exists, about thirty percent of employees have breached it, and of those, thirty-one percent said they'd do it again to save time.

Adam: Those are survey numbers, passed through a lawyer's article.

Mia: Exactly, conference-session and survey figures relayed by a lawyer, so treat them as indicative. The point stands, though: tools arrive faster than rules.

Adam: So what's the sequence?

Mia: Four steps, and the thread through them is to start small and make someone responsible. First, one named owner. The Commission says no AI officer is mandated, so that's a choice, not a legal requirement. But each failure we just covered had a gap where an owner should be. Nobody checking the airline bot against the policy, nobody pulling the New York bot when its errors were public.

Adam: What does an owner actually do, though? A title doesn't fix anything.

Mia: Fair challenge. An owner is one person who knows which AI tools the team uses, what data goes into them, who reviews their output, and who can switch them off. Those four things, nothing grander.

Adam: Okay. Step two?

Mia: A small first use with a check built in. For a chatbot, that means a narrow set of questions answered from approved policy text, with a way to reach a person. For drafting, a human reads every policy and checks it against the law and the company's actual practice, and nobody publishes the first draft.

Adam: And if the chatbot touches employee data?

Mia: Then data protection rules apply, and in many countries works council involvement, as we covered in Part 2. In Germany the Hamburg labour court held that merely letting staff use ChatGPT didn't trigger co-determination, but a tool that monitors or evaluates them can.

Adam: Step three?

Mia: Literacy for the HR team first. Article 4 is satisfied by reasonable measures that are recorded. So a short practical session on what these tools do, where they invent things, and what must never be pasted into them, with a note of who attended, is a defensible start.

Adam: And the fourth is the bot announcing itself?

Mia: Yes. The chatbot's first message should say it's an AI system, and its answers should point to the policy they come from. Beyond being the law, it's the cheapest trust measure there is, and the IBM story shows how fast trust goes when people feel something was imposed on them.

Adam: Here's what I'm wondering. If bots take the easy questions, what's left for the people in HR?

Mia: The honest summary is that routine work shrinks, judgment work grows, and nobody has measured it well yet. A May job-change tracker rates the pressure on HR business partners as moderate. Routine employee enquiries and benchmarking are being automated, analytics and succession planning are being assisted, and the role isn't disappearing. Gartner has a research note on the HR business partner role, dated the twenty-sixth of August, that argues something similar, though only its abstract is public, so we couldn't see the detail. These are interpretive sources, not measurements.

Adam: So what can be said with more confidence?

Mia: What follows from the cases. When a chatbot takes the simple questions, the ones that reach a person are the hard ones: disputes, health, family situations, conflicts. And the IBM history is a warning that the human route has to stay open. If you work in HR, that means the part of your day that's left is the part with the most at stake, so the human route to you has to be easy to find.

Adam: And there's a new duty on top of that, isn't there?

Mia: There is. HR becomes the department asked to answer for AI across the whole company: who delivers the literacy training, who tells employees what is used on them, who the owner is. An earlier Gartner survey found that thirty-five percent of the 133 HR leaders who responded expected to lead their organisation's AI ethics approach. It also found that forty-five percent of a larger group of 167 were working with legal and compliance on generative AI use cases. So the role gains a duty that's neither technical nor purely people work.

Adam: Which brings us back to the front door IBM built, and the managers who were sent through it too early.

Mia: Right. So here's the one thing to try this week, whatever your role: write down the name of the one person who owns your team's AI tools, even if that person is you. The reason is that every failure we covered had a gap exactly there, and a name on paper is the cheapest way to close it.

Adam: That's a short list. One name.

Mia: One name, yes. That closes the series: AI in hiring, AI and the people already employed, and now HR's own tools. Want a show like this about your own work? walkingpoints.fm

Adam: Thanks for walking with us.

Mia: Take care of yourself out there, and enjoy the rest of your walk.

Heard something wrong? Tell us