Human Handoff Protocol for AI Agents: The 12-Field Context Contract

What is a human handoff protocol for AI agents?

A human handoff protocol for AI agents is a versioned operating contract. It defines when an AI yields control, which person receives the case, what context follows, and how ownership can return. Done properly, escalation becomes a controlled transfer of evidence, state, responsibility, and deadlines rather than a simple route change.

Decagon defines AI agent handoff as transferring a conversation with its context when the AI cannot or should not continue alone. The same rule applies inside a company. A finance approver, HR specialist, security reviewer, or support representative needs to know what happened, what remains unresolved, and which decision now belongs to them.

This transfer is one control within a broader human-in-the-loop operating model. The protocol governs the handoff itself. An AI agent governance framework sets permissions, oversight, risk controls, and accountability across the full process.

What context should an AI transfer to a human?

Decagon and Smith.ai guidance supports transferring a concise issue summary, detected intent, full timestamped transcript, customer and account data, prior interactions, attempted actions and outcomes, collected resources, sentiment and urgency, escalation reason, policy or security constraints, channel and session identifiers, and the unresolved task with its next owner.

1. Issue summary: A plain-language account of the problem and its current state.
2. Detected intent: The outcome the user appears to want.
3. Timestamped transcript: The complete conversation with timestamps.
4. Customer and account data: Identifiers needed to find the correct record.
5. Prior interactions: Relevant CRM history, open cases, promises, and previous outcomes.
6. Actions taken: Every attempted step, its inputs, result, and resulting state change.
7. Resources collected: Uploaded files and other records collected during the interaction.
8. Sentiment and urgency: Signals that affect priority or the transfer experience.
9. Escalation reason: The exact rule, request, uncertainty, or failure that triggered transfer.
10. Constraints and flags: Applicable policy, regulatory, legal, fraud, privacy, or security conditions.
11. Channel and session identifiers: Stable references that connect the case across systems.
12. Next step and ownership: The unresolved task, recommended action, owner, deadline, and return criteria.

The summary and transcript do different jobs. A summary lets the recipient scan the case in seconds. The transcript preserves exact wording, sequence, timestamps, and evidence. Send only the summary and critical detail disappears. Send only the transcript and the recipient must reconstruct the case under time pressure.

When should an AI agent escalate?

Smith.ai, Decagon, and Trackmind guidance supports escalation when the user asks for a person, confidence falls below the approved risk tolerance, the conversation loops, resolution takes too long, complexity exceeds the agent’s scope, sentiment turns negative, policy requires review, or fraud, legal rights, security, urgency, VIP treatment, or high-consequence actions require human judgment.

• Direct human request: Transfer regardless of the AI’s confidence.
• Low confidence: Escalate when uncertainty exceeds the task’s approved tolerance.
• Repeated loop: Detect recurring questions, answers, or failed intent classifications.
• Unresolved duration: Transfer before delay creates a larger operational or customer problem.
• Scope or complexity breach: Stop when the required judgment exceeds the agent’s authority.
• Negative sentiment or urgency: Raise priority and shorten the route to a person.
• Mandatory review: Route financial disputes, legal matters, privacy requests, and security cases to people.
• High-consequence action: Require human judgment when an error is costly, hard to reverse, or difficult to explain.

Set thresholds by consequence, not convenience

A confidence score should never decide escalation by itself. Trackmind recommends assessing the maximum cost of error, reversibility, explainability, task-specific accuracy, consequence severity, and time sensitivity. A reversible scheduling change can tolerate more autonomy than a legal dispute or payment release. Use an AI agent risk assessment to set thresholds by task instead of copying one percentage across every workflow.

How should an AI-to-human transfer work?

Retell AI recommends a warm transfer in which the recipient can review a summary, attempted steps, and relevant information before taking control. Tell the user why control is changing, confirm that their existing information will follow, route by skill and priority, and define what happens when the intended recipient is unavailable.

Keep the user-facing message direct: the request needs a specialist, the information already supplied will transfer, and the user will be told what happens next. Leading workflow platforms document that placing the human in the same support console used by the AI preserves information the customer already supplied. Ownership must move to a person or queue, never an undefined waiting state.

If nobody is available, retain the payload, assign a fallback owner or queue, record the deadline, and tell the user what happens next. Never return control to the AI without saying so. The return needs an explicit condition, such as human approval, corrected data, or completion of a restricted action.

How should handoff context be represented and connected?

Smith.ai guidance supports representing handoff context as a machine-readable object and connecting it to CRM or ticketing systems through APIs and webhooks. Pair that object with a human-readable briefing and deliver it to the support console, CRM, or ticketing environment people already use. Keep identifiers stable, record state changes, and accept updates in both directions so ownership and outcomes stay current.

{
  "schema_version": "1.0",
  "handoff_id": "ho_1042",
  "issue_summary": "Purchase form and quote amounts differ.",
  "detected_intent": "Submit a valid purchase request",
  "conversation": {
    "transcript": [
      {"at": "2026-10-02T14:03:00Z", "speaker": "user", "text": "Please submit this quote."},
      {"at": "2026-10-02T14:03:08Z", "speaker": "agent", "text": "The two amounts do not match."}
    ]
  },
  "subject": {
    "customer_id": "cust_example",
    "account_id": "acct_example",
    "prior_interactions": ["pr_1008"]
  },
  "actions_attempted": [
    {"action": "extract_quote_amount", "outcome": "12500 USD"},
    {"action": "compare_form_amount", "outcome": "mismatch"}
  ],
  "resources": ["quote_1042.pdf"],
  "signals": {
    "sentiment": "neutral",
    "urgency": "standard",
    "confidence": 0.38,
    "complexity": "medium"
  },
  "escalation_reason": "Conflicting financial values",
  "constraints": {
    "policy_flags": ["human_verification_required"],
    "security_flags": []
  },
  "session": {"channel": "workflow", "session_id": "sess_204"},
  "unresolved_task": "Verify the correct amount",
  "recommended_next_step": "Finance reviews the form and quote",
  "ownership": {
    "queue": "finance_review",
    "deadline": "2026-10-02T16:00:00Z",
    "return_condition": "Verified amount entered by Finance"
  }
}

Publish the schema with required fields, allowed values, and version history. Reject incomplete payloads before accepting a transfer, and preserve the original evidence rather than replacing it with an AI summary.

The receiving system should acknowledge case creation, return the assigned owner and status, and send later changes back to the orchestrator. This two-way pattern keeps the AI, human queue, and user-facing status consistent. It also supports AI agent orchestration, where workflows and people coordinate around one authoritative task state.

How should teams implement and test the protocol?

Implement the protocol as an operational product, not a prompt adjustment. Define trigger policy, publish the schema, connect systems, assign owners, rehearse fallback states, train recipients, and audit real transcripts. Test correct escalation and correct non-escalation. Unnecessary takeovers expose weak thresholds just as clearly as missed transfers.

1. Inventory tasks and identify the decisions, data, and consequences that require human control.
2. Define mandatory triggers, configurable thresholds, prohibited actions, and direct-request handling.
3. Publish the versioned context schema and map every field to its source system.
4. Connect the AI, CRM, ticketing, or workflow system through APIs and webhooks.
5. Assign skill-based destinations, priority rules, deadlines, fallback owners, and return conditions.
6. Train recipients to scan the briefing, verify evidence, and update the shared case state.
7. Run controlled tests, audit transcripts, and revise trigger or payload versions based on observed failures.

Measure continuity, speed, and routing quality

Measure repeated-information rate, transfer success, routing accuracy, time to human response, resolution time, average handling time, re-escalation, abandonment, and customer satisfaction. Review them together. A fast transfer with poor routing has failed. Shorter handling time is not progress when users must repeat the case.

Put this work on a staged AI agent implementation roadmap. Start with bounded cases, inspect failures, and widen autonomy only after the human transfer path works reliably.

How is a human handoff different from an agent-to-agent handoff?

A human handoff prepares a person to assume responsibility, so it needs a scannable brief, evidence, user communication, ownership, and deadlines. MindStudio explains that an agent-to-agent handoff prioritizes precise, predictable, machine-readable inputs. Yannic Kilcher’s Agent Handoff Protocol announcement describes moving an objective, conversation, and resources between AI applications and positions the protocol as complementary to MCP.

Do not use these labels interchangeably. A machine can parse a technically valid payload that is painful for a person to scan. A polished human summary can omit fields another agent requires. Build the presentation for its recipient while preserving one authoritative task state underneath.

How Cogniver helps build human handoffs into approval workflows

Cogniver turns internal handoff rules into directed approval workflows with branching, merging, and multi-step chains. Required document uploads keep evidence attached to the request. Approvers can enter verified values at their step, and later routing decisions can use those values.

An AI Router sends each request down exactly one branch using exact amount rules or an AI-applied plain-words policy. Every routing point includes a mandatory default branch. When the evidence is unclear, the request follows the designated human path instead of stalling or forcing the AI to guess.

Each workflow also gets its own isolated AI agent, trained by organization administrators on that workflow’s rules and configuration. It can answer questions, route requests, chase approvers, or sit inside the flow as an approval step. When a request needs human judgment, the default branch can send it to an approver, who can verify values before later routing continues.