AI Agent Human Handoff: Writing Rules That Work

It is Tuesday afternoon, you are on a call with another client, and a message arrives saying the work was not what they expected and they want to cancel. The person who usually handles that kind of note is in the room with you, so the only thing that can answer in the next few minutes is the agent that has been taking the routine questions. It will either write back on its own or stop and send the thread to the account manager, and the difference is a rule written before this message showed up. That rule is the human handoff. Without one, the agent handles every case until it mishandles one that matters, or it stalls on every ambiguous moment and becomes slower than the person it was supposed to help. An operations agent runs when you write both instructions, what it may finish alone and what it must bring to a person. Bain's Automation and AI Pathfinder Survey of 951 companies in June 2026 found that only 7% run fully autonomous agents in production.

Why does a weak handoff rule cost more than no rule at all?

An agent with no handoff rule tries to finish everything, and your team finds the mistake after the client has already read it. A weak rule costs more, because it sends low-stakes notes to your team all day. Then a message that really does need the account manager arrives, the rule fires, and nobody answers quickly, because the team is still clearing notes that did not need them.

What the studies say about letting an agent decide alone.

Renieris, Kiron, Mills, and Kleppe, writing in MIT Sloan Management Review in September 2026, reported that 72% of the expert panelists agreed that treating AI agents as fully autonomous decision-makers will fail to maintain accountability. The agent can execute the message. It cannot own the apology, the credit, or the conversation that follows a bad one.

Bain's June 2026 Pathfinder survey of 951 companies found that 50% of the companies that delivered their cost-reduction targets ran agents at guardrails-level autonomy or above, compared with 38% of the companies that fell short. Guardrails-level means the agent handles what it was built for and brings everything else to a person. One mishandled escalation costs more to repair than a month of routine output.

Which situations should always route to a human?

Three kinds of situation belong with a person no matter how capable the agent has become. Put them in on the first day, and change them only in a review you scheduled.

Decisions the agent cannot undo.

Any action the agent cannot take back belongs with a person: a refund, a canceled subscription, a contact removed from a campaign, or a statement published under the company name. Write an explicit rule for each action that cannot be undone within 60 seconds, and name the person who receives it.

Conversations that have turned heated.

A thread that shifts into a complaint, a billing dispute, or plain frustration belongs with a person, and so does a message from a customer whose account value is above a threshold you choose, or who flagged a concern in the previous 30 days. If the message contains "cancel", "not what I expected", or "disappointed", the agent should stop and route the full conversation to the account manager.

Cases outside what you have actually tested.

IBM's 2025 study of 1,700 senior data and analytics leaders, run from July through September 2025 across 27 countries and 19 industries, found that 79% of organizations were still early in defining how to scale and govern AI agents. In that same study, 83% believed the benefits outweigh the risks, and 77% were comfortable relying on agent outcomes. If you have not watched how the agent handles a situation, that situation goes to a person. The post on AI agents versus RPA walks through where a modern agent handles variation and where an older rule-based bot still draws a different line.

How do you write a handoff rule your agent can follow?

A handoff rule has three parts: the trigger, the action, and the context the person receives. Miss the trigger and the rule never fires. Miss the action and the agent pauses without doing anything useful. Miss the context and the person starts from a blank screen.

Writing a trigger the agent can actually check.

The trigger has to be something the agent can read. "If the client seems upset" fails, because "seems upset" is not a field. "If the message includes 'cancel', 'refund', or 'disappointed'" is a condition it can evaluate the same way every time. Start with three to five phrases that mark the cases your team handles by hand, and test them against the last 90 days of odd cases before you turn the rule on.

For tone, use a score band. Most platforms expose a sentiment estimate on a numeric scale, so fire the rule when that score drops below a value you chose. Set the first threshold toward more handoffs, and tighten it after 30 days of looking at what fired. The post on writing agent instructions covers the five-field format that keeps these rules consistent, including the exception that tells the agent what it may still say while it waits.

Writing the action and what travels with it.

Name exactly what the agent does: stop responding, forward the thread, create a task, or alert a named person. "Escalate", "flag", and "alert" produce different behavior on different days. If the action does not name a person, a place, and a format, it is not an action yet.

A worked example of a handoff rule.

Trigger: the message contains "cancel", "refund", or "disappointed", or the sentiment score is below -0.6. Action: stop. Forward the full conversation to the account manager with the subject "Review: [client name], [date]". Include the client's original message, the last three agent responses, the account tier, and the date of the last billing event. Do not send an automated reply. Stop.

What context should the agent include in the handoff?

A handoff with no context is a notification. The person still opens the CRM, finds the conversation, and rebuilds what happened, which takes 10 to 15 minutes per case. Five of those in a day is an hour you could have attached to the alert.

The four things that should travel with every handoff.

• The full conversation or task history, starting from the client's first message.
• The trigger that fired, for example "rule fired on the word cancel".
• Account details: tier, contract value, last billing date, and any open issues from the previous 30 days.
• Any action the agent took before the handoff, including messages it sent and data it looked up.

If the person has to open a second system before they can answer, the package is incomplete.

The format has to match where your team already works.

If your team lives in email, forward a structured message with labeled sections. If they live in a project tool, create a task with those four items in the description.

Chan, Ezell, Kaufmann, and colleagues, at ACM FAccT in 2024 (arXiv:2401.13138, January 2024), named three ways to keep an agent visible: an identifier, real-time monitoring, and an activity log. The package you send at the handoff is that activity log, so build it into the rule.

How do you calibrate the escalation threshold?

The threshold is the line between "the agent handles this" and "a person reviews this." Too high, and real issues slip through. Too low, and the team is buried in alerts that did not need them. A missed high-stakes case costs more than a reviewed low-stakes one, so start with more human review and tighten the line after you have seen real cases.

A 30 day loop for tightening the line.

Run 30 days on the conservative threshold, then sort every handoff into cases the team acted on and cases they closed without changing anything. If more than 30% are the second kind, raise the threshold on the trigger that generated most of them. If the team handled a case the agent should have handed over, lower the threshold for that trigger and run another 30 days.

What the performance numbers suggest you aim for.

The same Bain 2026 survey of 951 companies found that 38% still require a person's approval before each agent action, and that 32% use a guardrails-plus-exception model, where the agent acts inside its limits and flags cases at the edge of its confidence. That 32% group delivered better outcomes than companies using either pure autonomy or a person approving every action, which is the pattern to copy: the line moves outward only as you verify more kinds of cases.

The September 2026 MIT Sloan article says a number that lives only in a policy memo is not a threshold, so write it into the instructions.

What makes a handoff rule fail in production?

Most handoff rules that fail do it in one of three ways, and it is cheaper to name them before the first live week than after a client has already read the wrong reply.

Three ways the rule breaks, and the fix for each.

The trigger is too broad, so the agent hands work over too often, the team stops answering, and a real case is missed in the noise. Narrow it to specific words or score bands from cases you have already seen.

The action is ambiguous. "Escalate to the team" produces an alert nobody owns, which is a polite way to make a cancellation everybody's job and therefore nobody's on a Tuesday. Name the person, the tool, and the format: "Forward to the account manager by email with the subject Client escalation: [name]."

The context is thin, so the person cannot act without opening other systems. Make the four-part package a required output, and ask a teammate who has never seen the case to decide in under two minutes.

Watching the rule drift after the first month.

A rule tuned on last quarter's messages may fire too often, or not often enough, on this quarter's. Review every handoff rule every 60 days, and track how often a handoff did not need a person and how often a case that needed a person never fired. If either rate rises above 20%, revise the rule before the next review.

Our explainer on the White House superintelligence order and accord maps the accord's four layers onto a team running agents: written limits on what each agent can touch, a named owner who reviews its output, a second reviewer who spot-checks it, and a weekly report to the team's leader.

If the rule you still need is the one that stops a cancellation, a refund, or a promised time from going out while you are with another client, write that one first. The weekly owner report is where the pattern of those cases shows up, and you can describe the job on the get your free Systems Plan.

Common questions.

What is an AI agent human handoff rule?

It is the instruction, written into what the agent reads before it acts, that tells it when to stop and bring the task to a person. A rule you can run has three parts: a trigger that says what fires it, an action that says what the agent does next, and a context package that says what the person receives.

When should an AI agent hand off to a human?

Hand the work over when the situation is outside the agent's scope, past the authority you gave it, or past the emotional line you set. The cases that show up most often are a customer who expresses frustration a second time, a commitment that would pass a dollar or time threshold the agent is allowed to make, a confidence score below the level you accept, a legal, medical, or compliance question, and a customer who asks to speak with a person.

What should go in an agent handoff context package?

Include a short account of the conversation so far, the trigger that fired so the person knows why they received it, any commitment the agent already made, the customer's stated goal or complaint, and the next action you recommend. The person should be able to act without reading the full transcript from the first line. A missing package is the most common reason a handoff fails.

How do you know if your handoff thresholds are calibrated correctly?

Watch two rates side by side. The handoff rate is the share of agent sessions that end in a handoff. The unrecovered handoff rate is the share of those handoffs that do not end in a resolved outcome for the customer. If your handoff rate is below 5%, the thresholds are too high and the agent is finishing situations it should have brought to a person. If your unrecovered rate is above 20%, the packages are incomplete or the thread is going to the wrong person. For the first month, read the actual handoff transcripts every week and adjust the words that fired wrong in either direction.

What is the difference between a guardrail and a handoff rule?

A guardrail stops the agent from taking an action you have prohibited. For example, the agent cannot promise a refund above $500. A handoff rule sends the work to a person when a condition is met. Guardrails are the stop, and handoff rules are the routing, so most production agents need both. The Bain 2026 survey found that 32% of companies use a guardrails-plus-exception model, where the agent acts inside its guardrails and the exceptions trigger a person's review.

Can I use a generic handoff rule template?

A generic template can give you the shape: the trigger, the action, and the context package. The thresholds still have to match your business. Start with the shape, run ten real sessions, and read every handoff. Move the triggers based on what fired wrong in both directions. For most teams that calibration takes two to four weeks, not two hours.

Where these figures come from.

Renieris, Kiron, Mills, and Kleppe published "Responsible AI Means Knowing the Limits of Agent Autonomy" in MIT Sloan Management Review in September 2026. Their panel is the source of the finding that 72% agreed treating agents as fully autonomous decision-makers will fail to maintain accountability. The article is at sloanreview.mit.edu.

Bain's Automation and AI Pathfinder Survey 2026 covered 951 companies in June 2026. It is the source of the 7% running fully autonomous agents in production, the 38% that require a person's approval before each agent action, the 32% on a guardrails-plus-exception model, and the comparison in which 50% of companies that met cost-reduction targets ran agents at guardrails-level autonomy or above, against 38% of companies that fell short. The write-up is at bain.com.

IBM's Institute for Business Value CDO Study 2025 surveyed 1,700 senior data and analytics leaders in 27 countries and 19 industries from July through September 2025. It is the source of the 79% still early in defining how to scale and govern AI agents, the 83% who believe the benefits outweigh the risks, and the 77% comfortable relying on AI agent outcomes. The newsroom summary is at newsroom.ibm.com.

Chan, Ezell, Kaufmann, and colleagues published "Visibility into AI Agents" at ACM FAccT in 2024, also released as arXiv:2401.13138 in January 2024. The paper is the source of the three visibility categories used here: identifiers, real-time monitoring, and activity logging. It is at arxiv.org/abs/2401.13138.