Agent Human Handoff Runbook: Teach AI Digital Employees When To Escalate

Agent human handoff is the operating pattern where an AI digital employee pauses at a relationship decision, business exception, sensitive authorization, or evidence gap and hands a complete decision packet to the right human owner. The pain point is not that a person stays involved; it is that manual, time-consuming, repetitive approvals interrupt the wrong people and still miss the risky decision. The more useful enterprise pattern sits between two extremes: not everything stays manual, and the Agent does not become fully autonomous without review. Let the Agent handle tedious, verifiable preparation work. Let humans decide the parts that depend on judgment, policy exceptions, or trust.

This is not the same as an exception queue. The article on Agent exception queues explains what should happen when a workflow fails, gets stuck, or hits an unexpected edge case. Agent human handoff is planned delegation. A customer refund exception, supplier pricing decision, legal clause escalation, or external email send is not necessarily a failure. It is simply not the kind of decision an Agent should make alone.

Axon makes this pattern practical by turning handoff into workflow design instead of a last-minute mention in a chat. The Agent can call Skills to search material, read files, generate summaries, and prepare evidence. Trust Mode stops high-risk actions at confirmation points. The workspace keeps artifacts and run records. For the broader boundary design, read the article on human approval boundaries.

A useful handoff is not "please review." It states the current status, verified evidence, available options, recommended judgment, and what the Agent should do after the human responds.

Handoff Is Division Of Labor, Not Failure

The most wasteful office handoff is not that an Agent asks a person. It is that it asks too late, asks the wrong person, or asks without enough context. Operations leaders are already interrupted by repetitive approvals every day. The genuinely important risk decisions get buried inside vague messages. Agent human handoff is meant to make the Agent say, in operational language, "this is what I finished, this is the judgment I should not make, and this is the decision I need from you."

Most handoffs fit into three groups:

• Business judgment: whether to make an exception for a VIP customer or accept a delayed delivery explanation.
• Risk authorization: whether to send an external email, overwrite an important file, or create or delete a calendar event.
• Evidence gap: conflicting sources, missing fields, or inconsistent file versions.

These groups may require different owners. Routing every handoff to an administrator creates another bottleneck. A handoff runbook should identify the accountable role before the Agent starts running.

A Role Map Keeps Handoff From Drifting

Situation	What the Agent completed	Human owner	Response returned to the Agent
Customer exception	Summarized order, communication history, refund policy	Customer success lead	Approve, reject, offer credit
External sending	Drafted email, listed recipients and attachments	Business owner	Send, rewrite, hold
Compliance concern	Extracted clauses, marked risk phrases, collected evidence	Legal owner or counsel	Continue, revise, seek outside view
Data conflict	Compared file versions and listed differences	Data owner	Choose source, request missing material

Four Questions Every Handoff Packet Must Answer

A good handoff packet should save the reviewer from rerunning the entire workflow:

1. What was the goal of this run, and why did it trigger handoff?
2. Which sources did the Agent read, and which artifacts can the reviewer inspect?
3. What options are available now, and what risk comes with each one?
4. After the owner decides, what should the Agent do next?

If the packet cannot answer those questions, the handoff turns into "can you explain again?" and the team falls back into manual coordination.

Owner Routing Belongs Before The Run

The Agent should not reach the end of a workflow and then search for an owner. Refund exceptions go to customer success, compliance questions go to legal, and data conflicts go to a data owner. The earlier these rules appear in the Agent brief, the fewer irrelevant interruptions the team gets.

A Refund Review Example

Imagine a customer success team wants an Agent to prepare high-value refund requests every afternoon. Low-value, policy-clear requests can become reply drafts. High-value requests or requests involving prior complaints require human review. The handoff packet can be explicit:

{
  "handoffType": "customer_exception",
  "runId": "refund-review-2026-05-24-1500",
  "agentGoal": "Prepare refund decision material for high-value requests",
  "triggerReason": "Refund amount exceeds standard policy and customer has two prior complaints",
  "evidence": {
    "orderFile": "workspace/refunds/2026-05-24/orders.xlsx",
    "conversationSummary": "workspace/refunds/2026-05-24/customer-8241.md",
    "policyExcerpt": "workspace/refunds/2026-05-24/refund-policy.md"
  },
  "options": [
    {"label": "approve refund", "risk": "sets exception precedent"},
    {"label": "offer credit", "risk": "may not satisfy customer"},
    {"label": "escalate to manager", "risk": "slower response"}
  ],
  "ownerRole": "customer-success-lead",
  "agentNextActionAfterDecision": "draft reply and keep Trust Mode on Confirm before sending"
}

In this example, the Agent does not decide the relationship outcome. It reads, organizes, explains, and drafts. The human keeps the decision that affects the customer relationship.

When Handoff Should Become An Exception Queue

Handoff is planned delegation. Exception queues handle unplanned failure. Mixing the two makes operations noisy.

• If the trigger is "amount exceeds threshold," "customer manager required," or "confirm before sending," it is handoff.
• If the trigger is "file read failed," "missing source," "output schema invalid," or "permission denied," it belongs in an exception queue.
• If the same handoff is rejected repeatedly because one field is always missing, the team should fix the Agent design rather than ask the owner to compensate forever.

In Axon, these signals belong in the AI Agent control plane, where queues, status, and artifacts are visible to operators. The NIST AI Risk Management Framework is useful because it treats AI risk as a managed process. Google's SRE guidance on incident management is also worth borrowing for ownership, escalation, and review, while remembering that planned handoff is not an incident.

FAQ

Q1: If the Agent asks people, does automation still help?
Yes. Agent human handoff should appear only at high-risk, relationship-sensitive, or evidence-limited nodes. Reading documents, summarizing material, preparing files, and drafting responses can still be automated.

Q2: How is a handoff packet different from an approval popup?
An approval popup usually asks "approve or cancel." A handoff packet gives evidence, options, risk, owner, and next action so the reviewer can make a real decision.

Q3: Do small teams need this much structure?
They can start with a lightweight Markdown packet. The required fields are runId, trigger reason, evidence path, available options, and owner role.

Start With One Repeated Escalation

Start with one repeated human interruption, not the whole company workflow. After you download Axon, choose a scenario such as refund exceptions, external email sending, or contract risk confirmation. Let the Agent generate a handoff packet and ask the owner to make one real decision from it. Then read more about Trust Mode and exception queues, and turn Agent human handoff into a collaboration rule instead of a rescue habit.