A few months ago, a mid-sized fintech company gave its support team an AI agent that triaged tickets, drafted responses, and escalated the messy ones. Nobody officially “onboarded” it. It didn’t show up in the org chart, didn’t have a manager, and didn’t go through a performance review. It just started working – and the team started relying on it the way you’d rely on a new hire who happened to never sleep.
That’s the strange part of where we are right now. AI agents aren’t a future problem for HR to plan around. They’re already sitting in the workflow, doing real work, and nobody quite agrees on who’s responsible for them.
What we mean by “AI agents as colleagues”
An AI agent isn’t a chatbot you ask a question and forget about. It’s software that can take a goal, break it into steps, use tools or other systems, and carry a task through to completion with minimal hand-holding. Book a vendor meeting. Pull last quarter’s numbers and flag the anomalies. Draft a first-pass contract review. The agent doesn’t wait for instructions at every step – it acts, checks its own work to some degree, and comes back with output.
That’s a meaningfully different relationship than “using a tool.” A spreadsheet doesn’t make decisions. An agent does, even if narrow ones. Teams are starting to talk about them the way they’d talk about a junior teammate: reliable at some things, prone to specific mistakes, needing oversight in certain areas and none in others.
The numbers are moving faster than most policies are
The scale of this shift is easy to underestimate if you’re only watching it inside your own company. Gartner has predicted that 40% of business applications will include task-specific AI agents by the end of 2026, up from less than 5% in 2025. Separate research from Salesforce found the average company now runs around 12 AI agents, with that number expected to reach 20 by 2027 – and roughly half of those agents currently operate without any connection to each other, let alone a coordinated oversight structure.
Adoption is running ahead of governance. That’s the part worth sitting with. Companies are handing agents real responsibility before they’ve decided who owns the outcomes.
It’s also worth being honest that this isn’t a guaranteed success story. Gartner has gone on record predicting that more than 40% of agentic AI projects now underway will be shelved by the end of 2027 – not because the technology fails outright, but because of rising costs, unclear business value, and weak risk controls. In other words: the agents themselves aren’t usually the problem. The management around them is.
So who ends up managing an AI agent?
In practice, it’s rarely one clean answer. A few patterns are showing up across the companies actually doing this:
The team lead who inherited it: Someone on the ground – a support manager, an ops lead – ends up de facto responsible because the agent sits in their workflow. They didn’t ask for the role. They just noticed the agent’s mistakes first.
IT or a platform team, for the technical layer: Access, integrations, uptime, and security sit with whoever already manages the company’s software stack. But they’re rarely equipped to judge whether an agent’s decisions were good ones – that’s a domain question, not a systems question.
HR and people leaders, for everything that touches humans: This is the genuinely new layer. If an agent is doing work that used to belong to a person, someone has to think through what that means for role clarity, skill development, trust, and how employees are expected to work alongside something that isn’t a colleague but also isn’t quite a tool. That’s squarely an HR and culture question, and it’s one most job descriptions haven’t caught up to yet.
The honest answer, for most organizations right now, is: partially all three, clearly none of them. That gap is exactly where things go wrong – an agent drafts something inaccurate, or takes an action nobody reviewed, and the postmortem reveals three teams each assumed someone else was watching.
What HR leaders should actually be doing about this
A few things that separate the organizations handling this well from the ones improvising:
Name an owner before the agent goes live, not after something breaks: Even a lightweight version – “this agent’s outputs are reviewed weekly by X” – closes most of the gap.
Treat agent oversight as a skill, not a side task: People managing agents need to know what “good” output looks like well enough to catch the bad kind. That’s a training investment, not something you assume people will figure out.
Decide what “autonomous” actually means for each agent: Full autonomy on a low-stakes task (drafting a meeting summary) is very different from full autonomy on something customer-facing or financial. Most failures trace back to this line being fuzzy rather than the agent being incompetent.
Keep the human relationships in the room: An agent can draft, sort, and summarize. It can’t build trust with a client, read a tense meeting, or notice when a teammate is burning out. Leaders who are thinking seriously about this – including Subramanyam Sreenivasaiah’s take on building AI and human-centric workplaces together – keep coming back to the same point: the technology changes the how of work, not the why people stay somewhere.
This is a culture question before it’s a tech one
It’s tempting to hand this whole topic to IT and move on. But the organizations getting real value out of AI agents are the ones treating this as a workplace design problem first. Who’s accountable? How people are expected to check an agent’s work without resenting it. Whether employees feel like agents are taking grunt work off their plate or quietly replacing their reason for being there. Those are culture questions, and they’re the same ones HR has always been the best-positioned team to answer.
Harvard Business Review’s recent analysis of failed agentic AI rollouts makes a similar point from the other direction: most failures aren’t about the model being bad, they’re about organizations skipping the unglamorous work of deciding ownership, scope, and review before deployment.
The teams that will look back on this moment as a smart bet are the ones asking “who’s managing this?” now, not the ones asking it after something goes wrong. If your organization is already thinking about how AI is reshaping people practices and culture, it’s worth digging into the broader conversation happening in Amazing Workplaces’ ongoing HR and culture coverage – this shift is going to touch almost every part of how workplaces are built from here.
