Subscribe Home Conversations On AI App Development CRM Enterprise IT Ethics & Governance Futures HR Industries ServiceNow on ServiceNow Platform Foundations Products & Solutions All topics For Leaders In IT & Dev Customer Experience Finance, Operations & Strategy Employee Experience Security & Risk News & Events People & Culture My List Explore All
November 19, 2026 7 min The 5 rules of AI agent delegation Delegating to AI agents is a management skill. The more freedom AI agents get, the more thought their boundaries deserve. AI Research
Evan Ramzipoor headshot
Evan Ramzipoor Editorial Strategist, ServiceNow
Thomas McKinlay headshot
Thomas McKinlay Founder and CEO, Science Says
A 3d cube with pastel highlights
Top takeaways AI agents can handle emotionally complex judgment calls and high-volume work if they get precise instructions.  AI agents need clear boundaries, human-override points, and auditable records to keep them accountable.  Delegating to AI changes human behavior. Research shows it can make people more willing to act unethically. 
Alt text

This article draws on the latest scientific research on human-AI collaboration and delegation, from institutions including Johns Hopkins University, the University of Southern California, Massachusetts Institute of Technology, and the University of Oxford. 

AI has moved beyond answering questions. Autonomous AI agents can now take a chain of actions on their own, such as updating a customer’s account and issuing a partial refund. As their capabilities grow, managers are assigning them more work. 

Delegating to an AI agent is different from delegating to a person. A manager who hands an employee a goal and steps back can count on shared experience, professional judgment, and the unspoken understanding that comes from working in the same organization. 

An AI agent has none of that context. Those instincts simply don’t carry over. 

The more autonomy you give AI, the more thought you actually need to put into its boundaries, monitoring, and fallback plans. Tinglong Dai Professor, Johns Hopkins University

“One of the greatest misconceptions organizations have is that more autonomous AI means less human oversight,” says Tinglong Dai, a professor at Johns Hopkins University. “It’s the opposite, and we call this the ‘autonomy paradox’: The more autonomy you give AI, the more thought you actually need to put into its boundaries, monitoring, and fallback plans.” 

AI agent delegation is quickly becoming a core management skill. The latest scientific research points to five rules that can help. 

1. Delegate tasks and judgment calls 

Imagine you’ve built Kessyria, an AI agent that handles customer queries. You point it at your ticketing system and let it take on the most common questions.  

Researchers at Stanford and MIT found that when a Fortune 500 software company gave its customer support agents an AI assistant, productivity rose by 15% on average and 30% for junior employees. Those gains showed up within the first month and held in the months that followed. 

The ServiceNow Enterprise AI Maturity Index 2026 found that organizations with AI deeply embedded reported productivity gains 5.6 times higher than less-mature businesses. 

Now imagine Kessyria handling a complex issue the way it does a password reset. A frustrated customer asks for a refund after a product glitch hits at a critical moment. Kessyria pastes the refund policy page into the chat without acknowledging the customer’s distress. 

AI agents manage volume well. Emotionally complex tasks are more difficult because they call for empathy and judgment. With the right instructions, AI agents can handle those too, and that’s where gains can compound. 

Research led by Sichuan University in China found that AI agents given explicit instructions for emotionally complex tasks outperformed humans by 49%. When the instructions also specified an emotional tone for particular situations, such as firm and negative with slow payers, the margin rose to 94%. 

AI agents can take on judgment-heavy work. How well they do it depends on how clearly humans define the job.

AI agents can take on judgment-heavy work. How well they do it depends on how clearly humans define the job. 

Recommendation: Map which recurring tasks can be automated, and write specific instructions for each. When a high-volume task also requires judgment, be even more specific about how the AI agent should respond, including both what it should say and the tone it should take. 
 

2. Set visible boundaries 

Kessyria is performing well, so you expand its authority. It can now waive fees, change account terms, access customer records, and coordinate with other AI agents. 

There’s just one problem: Nobody has documented what it can and can’t do. Customers start getting replies that mention sensitive details, such as a prior billing dispute or their home address, that have nothing to do with their query. 

Many organizations are in the same position. Their AI agents often lack proper governance. According to the 2026 ServiceNow Enterprise AI Maturity Index, 49% of employees say AI agents expose their business to data privacy and security risks, and 53% say AI agents are causing people to lose control over decisions. 

Research led by Johns Hopkins University lays out a framework called “assured autonomy” that describes how a well-governed AI agent deployment should work. The authors recommend that organizations define exactly how far an agent’s autonomy extends, across four elements: 

  • Scope: the actions the AI agent can take and the systems it can see and control 
  • Authority: what the AI agent can commit to on its own and when a decision passes to a human 
  • Safety constraints: hard limits it can’t cross, such as company policies or federal regulations 
  • Oversight mechanisms: monitoring, governance policies, and human sign-offs 

All four need to be built into the AI agents and the systems that govern them, as well as documented so people know how those systems work. 

Illuminated boundary lines

“Delegating to an AI agent is not merely assigning a task,” explains Yue Zhao, an assistant professor at the University of Southern California. “It is granting a bundle of authority. Organizations therefore need to specify both the outcome they want and the authority the agent may exercise to achieve it.” 

Recommendation: Build the four limits into an AI agent’s permissions and workflow: access scoped to the query, hard approval caps, limits it can’t cross, and defined triggers for human handoff. 

Then publish a one-page summary of those settings for everyone who works alongside the AI agent. And keep it in sync with the agent’s configuration. 

3. Make every AI agent auditable 

Six months in, Kessyria answers a high-stakes customer question. An executive at the client company reads the answer, disagrees, and asks a simple question: Why did it suggest doing that? Nobody can say for sure. 

Kessyria reasoned through five steps, pulled data from three systems, passed its findings to another AI agent, and took action that no one can reconstruct. 

Traditional software follows fixed code. When something goes wrong, engineers can replay the same input and see why an AI agent acted. An agent chooses its own path as it works: which systems to query, what to conclude, which agent to hand the task to.  

Many systems aren’t set up to record exactly what an AI agent did, who authorized each step, or what it was allowed to do. 

Research led by MIT and the University of Oxford argues that every AI agent action should trace back to a named human owner, including when one AI agent hands work to another. 

Maintaining that record is a good practice for any organization. In regulated industries such as finance and healthcare, a missing chain of evidence can become a regulatory problem. 

“When a regulator asks in 18 months whether that overprivileged agent was really cut off at a point in time, the answer should come back from the target system, not from the workflow that asked politely,” says Rahul Prakash, vice president and general manager of identity security at ServiceNow. 

Organizations...need to specify both the outcome they want and the authority the agent may exercise to achieve it. Yue Zhao Asst Professor, University of Southern California

Recommendation: “Preserve trustworthy evidence afterward, including approvals, retries, fallbacks, and delegation handoffs,” Zhao says. “That evidence makes the agent auditable, and auditability makes accountability possible.” 

 

4. Put humans in the loop 

You run an audit to see how employees are working with Kessyria. Some follow every recommendation it offers without question. Others override it constantly. Both think they’re doing the right thing. 

A recent framework from researchers at the University of Aberdeen and the University of Vaasa, which synthesizes research on human-AI teams, explains what’s happening. Teams struggle to know when to trust AI agents.  

People who defer to algorithms too readily, treating their outputs as authoritative, show what’s called “automation bias.” People who distrust AI entirely experience something called “algorithm aversion.” They may refuse to forgive its mistakes, or feel they’ve lost control over their own decisions. 

To help humans recalibrate, the same researchers recommend structured reviews of where and why human and AI judgment diverged so that people learn over time when to trust an AI agent. 

Where an AI agent sits in a workflow matters too, Dai says. “In our research at [Johns] Hopkins, we’ve found that it makes a lot of difference whether AI goes first, acts as a second opinion, or checks a human decision.” 

Track each point where humans enter a workflow. People may defer to an AI agent at one point and resist it at another. 

Recommendation: Review an AI agent’s override rate (how often people reject its actions) at each stage of its work, along with whether those overrides were correct. Researchers recommend starting with these benchmarks and adjusting them over time: 

  • An override rate below 10% suggests automation bias. People are deferring to the AI agent too much. Add clearer limits on its autonomy, such as required manual checks for important decisions, and prompts that encourage people to check its outputs. 
  • A rate above 50% suggests excessive distrust. If the AI agent was right most of the time, improve how it explains its recommendations, train employees to interpret its outputs, and draw clear lines between decisions the agent makes alone and those that need human verification. 
People walking through an illuminated hallway

5. Be aware of risks 

Kessyria is now calibrated, auditable, and operating within clear boundaries. Then your manager realizes they can set an upselling revenue target and let the AI agent work out how to hit it, with no instructions or restrictions. The agent might get there by unethical means, such as passing off upsells as mandatory updates. 

Across 13 experiments led by a team at the University of Duisburg-Essen, researchers found that people become more willing to act dishonestly when they delegate decisions to AI. 

When participants could simply set a goal, such as “maximize profit,” and leave the rest to the AI agent, 85% let it act unethically—by inflating numbers, for instance. When participants had to spell out exactly what the AI should do, 25% still told it to cheat. 

Delegation puts psychological distance between a person and the outcome. Zhao has seen the effect in his own research: “When people give loose or incomplete instructions and frame them conveniently, models will go beyond the scope of the task.” 

Nobody has to say, “Break this rule.” They can tell an AI agent to maximize a metric and let it decide how. 

Recommendation: Design AI agent interfaces that make it difficult to claim ignorance of how the agent will interpret instructions. Encourage users to write firm prohibitions into their tasks (for example, never misstate terms and conditions to close a sale). And put clear accountability and managerial review in place. 

Picture Kessyria a year from now. It still resolves tickets in minutes. Its limits are written where everyone can read them, every action traces to a named owner, and someone watches the override rate. A human signs off on the revenue target and what the AI agent can and cannot do to reach it. 

It’s effective, the people who delegate to the AI agent do so with care, and the risks are much smaller. 

Find out how ServiceNow can help put AI agents to work for you. 

Next up
Dive into more conversations AI App Development CRM Enterprise IT Ethics & Governance Human Resources Industries ServiceNow on ServiceNow Platform Foundations Products & Solutions All Topics
Stay in the know Join Us
stay in know image
Alt