An Unexpected Overtime in the Age of AI
On an autumn night in 2025, only one light remained on in the building housing the IT service desk of the City of Raleigh, North Carolina.
Under that light, the service desk supervisor stared at a stream of conversation logs on his screen, brow furrowed. A month earlier, he had been thrilled about the city’s adoption of an AI agent—the ServiceNow-deployed IT Helpdesk Agent, which promised to automatically handle high-frequency issues such as password resets and software installation guidance. He had naively believed this would give his team some breathing room.
But now, his workload had increased rather than decreased.
During the day, he managed his five human employees; at night, he had to “train” the AI agent—correcting its mistakes, checking for missed information, and monitoring every interaction it had with citizens. What he had thought would be a digital colleague now felt more like an intern requiring round-the-clock supervision.
He couldn’t help but recall a remark from Jacqui Canney, ServiceNow’s Chief People and AI Enablement Officer:
“Managers had a hard job before. Now, they have a harder job.”
This is merely the tip of the iceberg. A quiet battle over the “hidden management cost” of AI agents is now unfolding within organizations across the globe.
ServiceNow’s Shift: From Copilot to Agent
1. Early 2025
The enterprise software giant ServiceNow made a strategic decision: to fully embrace AI agents.
Unlike the “copilot” model popular in 2023–2024—where humans write prompts and AI assists with fragmented tasks—ServiceNow’s AI agents were given a brand‑new identity: autonomous executors.
These intelligent agents, called “AI Specialists,” can complete entire workflows end‑to‑end in three domains: IT service management, customer relationship management, and security and risk. They no longer require human confirmation at every step; instead, they receive a goal, break down tasks, call tools, and deliver results—much like a real employee.
At the same time, ServiceNow launched its “AI Control Tower”—a command center designed for management. What can this control tower do?
- Observe agent behavior: replay every decision path like a dashboard camera.
- Track ROI: precisely calculate how much money each agent call costs and how much value it creates.
- Govern and circuit‑break: allow managers to intervene with one click when agent behavior goes off track or costs become abnormal.
ServiceNow’s hope was that companies would buy AI agents like software and manage them effortlessly through the control tower.
But the real story is far more complex than the product brochures suggest.
2. A Contradictory Signal: Management Isn’t Getting Easier
In an interview with The Deep View, Jacqui Canney revealed a telling nuance:
“I actually hope our managers aren’t thinking ‘I’ve got five agents on my team.’ Instead, they should see agents as embedded parts of new workflows.”
The subtext: ServiceNow has discovered that if managers continue to treat agents as “colleagues,” they will fall into a huge management trap.
What trap? Blurred responsibility.
When traditional software fails, it’s a bug—call IT to fix it. When a human employee fails, it’s a performance issue—the manager has a conversation. But when an AI agent fails?
- Is it insufficient model capability?
- Is it bias in the training data?
- Is it the manager failing to set up the right prompts?
- Or is it the user’s unclear question?
No one is naturally responsible for it. Yet the person who ends up cleaning up the mess is still that supervisor sitting in front of the service desk.
Raleigh’s Front‑Line Report: A Manager’s Nightmare Week
Scene: The First Month of a Municipal IT Service Desk
Let’s return to Raleigh.
Chief Information Officer Mark Wittenburg described the case to the media in detail:
After deploying the AI agent, the service desk supervisor’s first week was a shock.
Monday: The agent went live. The supervisor spent half a day manually importing the city’s internal knowledge base—3,000 frequently‑asked questions, system permission guides, emergency procedures—into the agent’s training set. But import was not a one‑time task; because the agent kept encountering new questions not covered in the knowledge base, the supervisor had to supplement it constantly.
Wednesday: The agent began answering independently. The supervisor found that for clear instructions like “reset my password,” the agent performed well. But for ambiguous descriptions such as “I can’t log in—maybe my account is locked, or maybe it’s a browser issue,” the agent started giving incomplete or wrong answers. The supervisor had to label every suspicious session: “Correct,” “Partially Correct,” or “Wrong.”
Friday: The agent had processed 147 requests with an accuracy rate of about 82%. That meant the supervisor had to manually verify the 18% of erroneous cases, apologize to users, and make corrections. Worse, he discovered that in one session the agent had inadvertently leaked an internal server IP address—fortunately not causing a security incident, but it sent a chill down his spine.
Week summary: The supervisor originally spent 40% of his time on team management and process optimization. Now that 40% was entirely consumed by agent supervision, plus an extra two hours in the evening. His team did feel some relief (the agent handled repetitive tasks), but he himself was trapped in an unprecedented, high‑intensity “human‑machine sandwich” state.
Wittenburg admitted:
“That’s been a transition for the supervisor.”
The Core of the Conflict: The “Invisible Transfer” of Management Duties
Companies often calculate the ROI of AI agents using a financial model: software license fees + API call fees + implementation costs, compared to labor hours saved. But in Raleigh’s case, one hidden cost item was completely overlooked: the manager’s attention cost.
The manager’s role shifted from “managing people” to “managing people + managing agents + managing human‑agent interaction quality.” These three activities require entirely different skill sets:
- Managing people: needs empathy, motivation, performance feedback.
- Managing agents: needs technical understanding, log analysis, model debugging.
- Managing human‑agent interaction quality: needs process design, exception handling, rapid decision‑making.
Few managers possess all three capabilities. Even fewer companies provide additional training or compensation for this.
Financial and Security Time Bombs
If Raleigh’s story brought “management burden” to the surface, a warning from Jayney Howson, ServiceNow’s Chief Learning Officer, shone a spotlight on another layer of hidden cost: token economics.
1. Token: The “New Oil” of the AI Era
In the world of generative AI, a “token” is the billing unit, roughly equivalent to 0.75 English words. On the surface, the cost of a single API call is pitifully low—a few dollars per million tokens. But when AI agents run at high frequency across an organization, the bill expands rapidly.
Howson points to a troubling trend: the combined use of employees and agents is quietly driving up costs.
For example: A marketing specialist uses an AI agent to draft a customer email (agent calls GPT‑4, consumes 500 tokens), then manually revises it and asks the agent to polish it again (another 300 tokens). Next, the agent calls an internal database to pull last week’s sales data and generates an analysis paragraph (1,500 tokens). A single simple email task can involve hundreds of token calls—while the manager remains unaware, until month‑end when a bill for tens of thousands of dollars arrives.
Even more alarming is the risk of data leakage. When agents are granted access to sensitive internal systems (e.g., HR systems, customer databases), every call may transmit data fragments to external models. Without strict “data permission boundaries” and “output auditing,” a small mistake can become a major compliance disaster.
Howson’s original words:
“If managers aren’t prepared, they will be left cleaning up the mess.”
2. A Fictional Yet Highly Realistic Scenario
Consider a typical mid‑sized company:
- Deployed 5 AI agents (IT support, HR Q&A, sales assistant, finance reconciliation, compliance review).
- 200 employees frequently interact with the agents every day.
- Agents call each other to complete complex tasks (e.g., “generate a contract for a new client and check compliance”), forming agent chains.
After one month, the manager faces three “surprises”:
- Billing surprise: The expected monthly AI cost of $5,000 becomes $35,000. The cause: unsupervised circular calls between agents (A calls B, B calls A, infinite recursion).
- Security surprise: While processing a contract, the compliance review agent sends the client’s non‑public financial data as context to a third‑party model, whose logs are externally accessible.
- Labor surprise: The IT manager spends an entire week manually tracing the abnormal call chain, writing new guardrails, and explaining and remediating the situation with affected clients.
In this mess‑cleaning process, no one rewards the manager. Instead, executives only ask: “Why did AI governance get out of control?”
AI Agents Are Not Employees—They Are a Runtime System
The ServiceNow case is a landmark because it forces the entire industry to confront a fundamental question:
What exactly are we managing?
The traditional answer: people and tools. The new answer: a runtime system that exhibits autonomous behavior, continuously evolves, and blurs accountability.
1. From “Software Lifecycle” to “AI Runtime Governance”
In the past, enterprise software deployment followed the classic “requirements‑development‑testing‑launch‑maintenance” model. After launch, software behavior was deterministic—Excel doesn’t make calculation errors because it’s in a bad mood.
But an AI agent is entirely different:
- It has no finite state machine; its behavior is based on probabilistic models.
- Its output shifts with model version updates, prompt tuning, and context changes.
- Its errors are emergent—even if each step is correct, the combination can be absurd.
This means enterprises can no longer manage AI agents like they manage software. They must establish a completely new governance paradigm: runtime governance.
Runtime governance demands:
- Real‑time monitoring: not a weekly review, but tracking the agent’s decision path every second.
- Dynamic guardrails: not predefined rules, but real‑time adjustments of permissions and boundaries based on agent behavior.
- Accountability tracing: every error must be attributable to a specific model, prompt, data, or management action.
ServiceNow’s AI Control Tower is essentially an attempt to implement this runtime governance. But as the Raleigh case shows, tools alone are far from enough—managers need new skills, new organizational support, and new incentives.
2. The Future of Managers: From “Running the Business” to “Running AI Operations”
The most powerful sentence in the case comes from Jacqui Canney:
“I actually hope our managers aren’t thinking ‘I’ve got five agents on my team.’ Instead, they should see agents as embedded parts of new workflows.”
This is not just a change in wording; it is a fundamental shift in worldview.
“Having agents on the team” means the manager still sees themselves as a manager of people, with agents as extra “digital subordinates.” This mindset leads the manager to micro‑manage every agent, ultimately sinking into the quagmire of micromanagement.
“Workflows are re‑embedded by AI” means the manager’s core task becomes designing, maintaining, and optimizing human‑machine hybrid workflows. Under this view:
- AI agents are not subordinates; they are autonomous nodes in the process.
- The manager’s value is no longer “controlling every step” but “ensuring the entire process converges on cost, quality, and risk.”
This requires managers to possess three new core capabilities:
- Process engineering: ability to map business flows and identify which steps are suitable for agents and which must remain human.
- AI economics: ability to calculate token ROI for each step and optimize calling strategies.
- Exception design: ability to pre‑set automatic fallbacks, human backup, and post‑incident recovery mechanisms when agents fail.
Unfortunately, the vast majority of mid‑level managers today do not have these capabilities. Even more unfortunately, no business school systematically teaches “AI process governance.”
The True Watershed for Enterprise AI Adoption
The ServiceNow case leaves us not with an easy answer, but with a heavy exam.
Question No. 1: Are You Willing to Acknowledge the “Hidden Costs”?
Many companies still calculate AI agent ROI using Excel models and are delighted to find “payback in less than six months.” But they overlook:
- How much is the manager’s extra time worth, converted into salary?
- How much customer churn cost is caused by agent errors?
- How much additional insurance and audit expense is incurred due to data leakage risks?
- How much “decision‑delay cost” arises from unpredictable agent behavior?
Acknowledging these hidden costs is the first step to maturity.
Question No. 2: Are You Willing to Restructure Management Capability Models?
ServiceNow has already begun internal action: they have formally incorporated “AI enablement” into all managers’ job descriptions and established a mandatory “AI governance certification” course. The curriculum includes:
- How to read an agent’s trace log?
- How to set prompt guardrails?
- How to calculate token ROI for each process?
- How to design human‑machine breakpoints?
This is no longer a “nice‑to‑have” skill; it is a fundamental capability for future managers.
Question No. 3: Are You Willing to Invest in “AI Observability Infrastructure”?
Without measurement, there is no management. ServiceNow’s AI Control Tower provides a template, but it may not fit every enterprise. The key is that enterprises need to build their own:
- Agent behavior logging system: record every input, output, and intermediate reasoning step.
- Cost attribution system: trace token consumption to specific departments, processes, managers, and agent instances.
- Anomaly circuit‑breaking system: automatically pause and notify the manager when a single agent’s call cost exceeds a threshold or when sensitive data is attempted to be transmitted.
The construction cost of these infrastructures is not trivial, but they are the only guarantee against “uncontrolled chaos.”
The Real Winners in the Age of AI Agents
Today, in 2025, every tech company is talking about AI agents. But the ServiceNow case reveals a sobering truth:
The first enterprises to deploy AI agents are not necessarily the winners. The real winners are those that can govern AI agents with the lowest management cost and the highest reliability.
What will become of that service desk supervisor in Raleigh? If he receives adequate training and tools, he may slowly transform from an “agent babysitter” into a “workflow architect”—no longer checking every agent answer line by line, but designing an automated quality sampling and feedback loop. His team will no longer be bogged down by simple repetitive labor, but will focus on complex requests that truly require human empathy and judgment.
If he does not receive support? He will burn out, resign, and become another silent casualty on the road to enterprise AI transformation.
And the ultimate message from ServiceNow is:
Don’t ask “What can AI agents do?” Ask “Is our organization ready to manage AI agents?”
This quiet war over “hidden management costs” has just begun. The enterprises that win this war will define the organizational form of the next decade. And those that fixate solely on technical ROI while ignoring governance systems will eventually discover—
The most expensive cost is always the one that never appears on an invoice, hidden in the tired eyes of managers and the low‑value redundancy that employees are forced to perform.