Reporting Lines for Software: The Quiet Rise of the Agent Operations Function

Somewhere in a Columbus insurance company, a procurement agent just renegotiated a vendor contract. Nobody told it to. It had standing instructions, a budget ceiling, and the authority to act. By the time a human read the summary, the renewal was signed.

This is not a hypothetical. Agentic AI systems — software that pursues goals across multiple steps, tools, and hours without hand-holding — are moving out of the demo theater and into the org chart. And they are dragging a question behind them that most enterprises have managed to avoid until now: when software behaves like an employee, who manages it?

The Accountability Vacuum

The traditional answer has been "IT owns the tool." That worked when AI was a recommendation engine or a chatbot. It collapses when the system holds credentials, transacts with vendors, and makes judgment calls that affect customers.

IT departments, stretched thin and measured on uptime, are not staffed to supervise autonomous behavior. HR, meanwhile, has no framework for a worker who never sleeps, never resigns, and cannot be interviewed. The result is a vacuum — and in American corporate life, vacuums get filled by whoever moves first, not whoever is best qualified.

We have seen this movie before. Shadow SaaS created a generation of ungoverned subscriptions. Shadow agents will be worse, because they do not merely incur costs. They take actions.

What an Agent Operations Team Actually Looks Like

The enterprises getting this right are assembling what some are calling agent operations — a small, cross-functional function that treats autonomous systems the way airlines treat aircraft: monitored, logged, and never unsupervised on the tarmac.

In practice, the team carries four responsibilities:

• Behavioral oversight. Defining what each agent may and may not do, and reviewing deviation logs the way a compliance team reviews trade records.
• Identity and access. Agents get named credentials, scoped permissions, and offboarding procedures. If your contractor badge process is stricter than your agent provisioning, you have a problem.
• Escalation design. Deciding in advance which decisions require a human signature — and making sure agents actually stop when they hit that line.
• Performance review. Yes, really. Agents drift. Their environments drift faster. Periodic evaluation is not optional.

None of this requires a new philosophy degree. It requires the same discipline enterprises already apply to privileged access and vendor risk — extended to a new class of actor.

IT, HR, or Neither?

Here is the opinion part: neither department should own this alone, and pretending otherwise will cost someone their quarter.

IT has the technical fluency but not the authority over business outcomes. HR has the governance instincts but, frankly, no mandate for non-human workers and little appetite to acquire one. The honest answer is that agent operations belongs to the business line that benefits from the agent, with shared accountability — the same model we use for financial systems. Finance owns the ledger; every department owns its numbers.

The 2026 org chart will not have a box labeled "AI Manager." It will have something less glamorous: a standing committee, a named operations lead, and a budget line. The companies that formalize this now — before an agent does something memorable in the wrong way — will spend their energy on optimization. The rest will spend it on incident response.

The second employee has shown up for work. Somebody should be assigned to manage them.