Why HR Should Own How Humans and AI Agents Work Together

A scene that is currently repeating itself in companies of every size, with the details changed and the shape identical.

An engineering director walks into her HR business partner's office, or more accurately into a video call, with a question that is not on any HR intake form. Her team has been working with AI agents for two quarters. One drafts code reviews and hands them to a second that checks them against the team's standards before a human approves anything. Another sits in the team channel answering status questions and escalating blockers. Productivity is up and so is friction: one senior engineer refuses to review anything an agent touched, a junior is quietly delegating work he was supposed to be learning from, and two teammates had a real conflict last week about whether an agent should have pinged a person in another department without a human deciding first. The director's question for HR is simple. "Who owns the rules for how my team works with these things?"

The HRBP checks. IT owns the security review. Legal owns the data terms. Finance owns the spend. Engineering leadership owns which tools get used. The answer to the director's actual question, how humans and agents should work together, who defines the norms, what happens when it goes wrong between people because of it, is owned by no one.

Sit with how strange that is. Companies have detailed, hard-won answers for every question about how people work together: how conflict gets handled, how work gets reviewed, how new members get onboarded, what respect looks like in a meeting, who can task whom. Those answers are the accumulated scar tissue of a century of organizational life, and HR is their custodian. A new kind of worker just joined the team, one that drafts, checks, chases, answers, and hands off work alongside everyone else, and the custodian of how-we-work-together is, at most companies, not in the room.

This essay is an argument that HR should be in the room, should in fact own the room, and that the window for claiming that ownership is short.

Why this is an HR problem and not an IT problem

The reflex at most companies has been to treat AI agents as software, and therefore as IT's domain. This made sense when AI meant a feature inside a tool. It stops making sense the moment agents become participants in collaboration: entities that hold responsibilities, get handed work, hand work to others, and interact with humans in the channels where the team's social fabric actually lives.

The evidence is in where the problems show up. Nobody files an IT ticket about the senior engineer who refuses agent-touched work; that is a norms conflict. Nobody calls the help desk when a manager starts using agent-gathered activity summaries in a way her team experiences as surveillance; that is a trust rupture, and it lands on HR's desk eventually, usually after it has already cost the team its best person. The junior engineer delegating away his own development is a capability-building failure, which is squarely a people function. The question of whether an agent may contact a person in another team without human sign-off is, whatever else it is, an org design question about interfaces and authority. Every one of these is a people problem wearing a technology costume.

Every one of these is a people problem wearing a technology costume.

The pattern even holds in the incidents that look most technical. When an agent at one company sent a customer an unapproved commitment, the postmortem found the root cause was not the model. It was that two teams shared responsibility for the agent and neither believed the review step was theirs, which is the oldest organizational failure there is, the kind HR and org design people have been untangling since before software existed. New actor, ancient failure mode, familiar owner.

There is a historical rhyme here worth taking seriously. When remote work arrived at scale, companies that treated it as an IT rollout, ship the laptops, deploy the VPN, got the tools right and the working norms catastrophically wrong, and spent years paying for it. The companies that fared best treated it as a change in how humans collaborate and put their people function at the center of designing it. The arrival of agents in teams is a larger version of the same event, and the same fork in the road is in front of every CHRO right now.

What owning it actually means

Ownership here is not a policy PDF. The companies getting this right, and there are enough of them now that the pattern is legible, have their people teams doing four kinds of concrete work, and the concreteness is the point.

First, they write working agreements for mixed teams the way good teams have always written working agreements, except the agreements now cover non-human members. What may an agent do without asking: answer factual questions, assemble context, chase a deadline it was assigned to watch. What must always route through a human: anything evaluative about a person, anything leaving the team's boundary, anything that commits the team to work. Who is accountable when an agent errs: always a named human owner, per agent, no exceptions, because accountability that isn't a person's name is a shrug. These documents are boring, and their boringness is their value. The engineering director's team conflict from the opening scene does not happen on a team that wrote the third clause down.

Second, they onboard agents like teammates rather than installing them like software, and HR runs the template. That sounds whimsical until you watch it work. An agent joining a team gets a charter: its responsibilities, its escalation rules, its owner, what it should read to understand the team's context. The team gets told, explicitly, what the agent is for and what it is forbidden to do, which preempts the paranoia that otherwise fills the vacuum. One people team I know of keeps the charters in the same system as job descriptions, which sounds like a filing quirk and is actually a governance decision: it puts the definition of every worker's role, human or not, in one reviewable place.

Third, they redesign development paths around the delegation problem, because the junior engineer quietly outsourcing his own growth is not an anecdote, he is a cohort. Every profession is about to discover that its traditional apprenticeship, learn by doing the grunt work, breaks when the grunt work is the first thing handed to agents. The people function is the only function positioned to answer what replaces it: deliberate rotations through judgment-heavy work, review responsibilities that force juniors to evaluate agent output rather than accept it, explicit skill maps for what humans must still be able to do unassisted. Companies that ignore this will discover in five years that their pipeline of seniors quietly dried up.

Fourth, and least comfortably, they police the surveillance line. Agents that sit in team channels can summarize anything, including people. The same capability that lets an agent brief a manager before a one-on-one, useful, consensual, appreciated, lets it produce activity reports that no employee agreed to, and vendors will happily sell the second thing wearing the first thing's clothes. Somebody in the company has to be the institution that says: context assembly for the human being discussed, yes; evaluation of humans without their knowledge, no; and anything feeding a performance decision goes through the person first. If HR does not claim that line, it will be drawn by default, by whoever configures the tools, which is to say it will not be drawn at all.

There is a fifth kind of work, less a task than a credential: the people team has to run its own operation this way before it can credibly design anyone else's. The HR organizations earning the room are conspicuous adopters themselves. Their agents field the first pass on policy questions in open channels and route anything personal to a named human partner, visibly, so employees always know who is in the conversation. Their HRBPs walk into manager consults with context an agent assembled overnight, the team's open questions, the stalled transfers, the survey verbatims that mention this manager, and spend the hour on advice instead of retrieval. Their performance cycles run as a chain, agent drafts and tracks, humans calibrate and decide, that the rest of the company can copy. When the engineering director asks how her team should work with agents, an HR team that works this way answers from experience, and an HR team that banned the tools internally while writing policy about them answers from a pamphlet, and every director can tell the difference in the first five minutes.

The prize, stated plainly

So far this has been the defensive case, the problems that land on HR when nobody owns the room. The affirmative case is bigger and CHROs should say it out loud in executive sessions: this is the people function's clearest route in a generation to the strategic seat it has spent decades asking for.

Every HR leader has lived the gap between the aspiration, strategic partner to the business, and the reality, keeper of processes the business tolerates. The gap persisted because HR's leverage points were weak: policies, programs, training nobody remembered. The design of human-and-agent collaboration is a different kind of leverage point. It determines, concretely and daily, how work flows, what managers spend hours on, how juniors grow, where trust accumulates or erodes. The function that owns that design owns something the CEO demonstrably cares about, because it shows up in output, retention, and risk within quarters, not years.

And the timing asymmetry is stark. Right now the norms are unformed, the vendors have not locked in defaults, and the first serious incident, the surveillance blowup, the agent-caused conflict that becomes a departure, the delegation scandal, has not yet happened at most companies. Whoever writes the rules before the first incident writes them calmly. Whoever writes them after writes them in a crisis, under legal's supervision, badly.

The engineering director from the opening scene got her answer, at her company, eventually. The HRBP took the question up the chain, the CHRO recognized what it was, and six months later the company had working agreements, agent charters with named owners, and an HR team that fields "how should my team work with agents" as a standard consult, the way it fields org design. The director's senior engineer, for the record, still reviews everything skeptically. He now does it as the team's designated owner of the code review agents, a role HR helped define, with authority to change how they work. His skepticism turned out to be the qualification.

That is the move available to every people leader reading this. The new occupants are already on the org chart in everything but ink. The only question is whether the people function writes the operating manual, or waits to be handed one.