AI agents will not be able to resolve every situation they encounter, nor should they.
Some tasks become too ambiguous. Some decisions exceed the authority originally delegated. Some situations introduce legal, security, financial or human consequences that require a different level of judgment. In other cases, the agent may simply reach the limits of what it can reliably do.
Those moments require escalation.
In human-agent work, escalation is more than sending an alert or asking someone to look at a problem. It is a deliberate transfer of responsibility from the current agent or supervisor to another person, role, team or governance process with the authority, expertise or context needed to decide what happens next.
Escalation is the transfer of an AI-assisted or agentic task, decision or condition to a different level of human authority, expertise or governance when the current agent or supervisor can no longer resolve it appropriately within the existing delegation.
The triggering condition might be uncertainty, repeated failure, unexpected behavior, policy conflict, increased consequence, missing authority or a direct request for human involvement.
The defining feature is the transfer.
A person may intervene in an agentic workflow without escalating it. They can correct instructions, reject a recommendation or pause execution while remaining responsible for the task. Escalation occurs when the existing human-agent configuration is no longer the right place to resolve the issue.
That may mean moving the work to a more senior decision-maker, a specialist, legal or compliance, security, an incident-response function, or another operational owner.
The concept is becoming explicit in current agent engineering.
Microsoft's guidance for responsible AI agents recommends defining escalation paths for cases an agent should not resolve on its own. It specifically calls for sensitive or ambiguous cases to be routed to people and for the handoff to provide enough context for reviewers to make a useful decision. (Microsoft: Apply responsible AI) (learn.microsoft.com)
Microsoft Copilot Studio also treats escalation as an operational workflow. Its human handoff guidance supports explicit and implicit escalation triggers and passes conversation history and contextual variables to the receiving human so they can continue the work without starting from scratch. (learn.microsoft.com)
OpenAI's practical guidance for building agents similarly identifies two common escalation triggers: exceeding failure thresholds and reaching high-risk actions that warrant human intervention. The guidance recommends designing a mechanism that lets the agent transfer control when it cannot complete the task appropriately. (OpenAI: A Practical Guide to Building AI Agents) (openai.com)
These implementations make the idea concrete.
For Human Risk Management, the question extends beyond whether the software supports a handoff. The escalation also has to move the work to the right human, at the right time, with enough context, authority and capacity to resolve it.
The distinction matters because these terms are often used loosely.
An Intervention Point is a moment or condition where human judgment can influence the workflow. Intervention can include correction, clarification, pausing, constraint, rejection, override or escalation.
Escalation is one possible result of intervention.
The current supervisor may recognize a condition they cannot or should not resolve themselves. Rather than continuing to exercise control at the same level, they transfer the issue upward, sideways or outward to someone better positioned to handle it.
For example, an employee supervising an AI procurement agent might correct an obvious classification error directly. If the agent uncovers a possible sanctions issue, the employee may need to escalate the matter to legal or compliance.
Both involve human attention.
Only one changes the ownership of the decision.
Human Override is a stronger form of control in which a person supersedes the agent's behavior or decision.
Escalation answers a different question: who should now handle the situation?
A person may override an agent and resolve the matter themselves. They may override it temporarily and then escalate. An agent may also escalate before any harmful action occurs, transferring the task because it recognizes that the situation exceeds its authority.
The distinction becomes useful when designing the control architecture:
| Control concept | Primary purpose |
|---|---|
| Intervention | Human influence enters the workflow |
| Override | Human authority supersedes agent behavior |
| Escalation | Responsibility transfers to another authority or expertise level |
| Reauthorization | The changed delegation is formally reconsidered before work resumes |
Keeping these concepts separate makes human-agent governance more precise.
An escalation mechanism is only useful when the system and people around it know when to use it.
Some triggers can be deterministic.
A financial transaction above a threshold may require senior approval. A customer request involving a regulatory issue may route automatically to a specialist. A workflow that reaches restricted data can be required to pause and escalate.
Other triggers depend on interpretation.
The agent may encounter conflicting instructions, incomplete evidence, ambiguous intent or a situation outside its normal operating context. A human supervisor may recognize that the agent's behavior is unusual even though no formal technical rule has been violated.
Common escalation triggers include:
OpenAI's current agent guidance explicitly recommends escalation where failure thresholds are exceeded or actions become high-risk, while Microsoft's responsible-AI guidance emphasizes sensitive and ambiguous cases that should move to human judgment. (openai.com) (learn.microsoft.com)
The practical challenge is tuning those triggers well enough that escalation catches meaningful conditions without turning every unusual event into an organizational fire drill.
A poorly designed handoff can make human intervention almost useless.
The receiving person may know that something has gone wrong without understanding the original objective, what the agent has already tried, what evidence it used, what actions were completed or why the escalation occurred.
That forces the human to reconstruct the case from the beginning.
Modern agent platforms are beginning to treat context preservation as a core part of the handoff. Microsoft Copilot Studio can transfer conversation history and variables to a live human, while Microsoft's external-engagement architecture recommends that escalation include full context so the receiving person can continue the interaction effectively. (learn.microsoft.com) (learn.microsoft.com)
For enterprise agentic work, the context requirement often needs to be broader than conversation history.
A useful escalation package may include:
The human should receive enough context to make the next decision without being forced to rerun the entire workflow mentally.
Escalation often gets treated as a simple agent-to-human transfer.
Enterprise work is rarely that simple.
Different conditions require different recipients.
A cybersecurity anomaly may need security operations. A contractual issue may need legal. A customer complaint may need a manager. An unusual financial decision may require a designated approver. An agent that exceeds its technical authority may need a system owner or administrator rather than the employee supervising the task.
This means escalation design needs a routing model.
The relevant dimensions can include expertise, authority, consequence, business ownership, urgency and availability.
Microsoft's handoff architecture reflects this operationally: context variables can be used to route an escalated conversation to the appropriate human or workstream. (learn.microsoft.com)
For Human Risk Management, routing also matters because accountability should move with the decision.
Sending every ambiguous case to a generic "human reviewer" preserves human involvement while potentially assigning the work to someone who lacks the authority or expertise to resolve it.
The language of escalation often implies hierarchy: a junior person raises something to a manager.
Agentic systems create a more varied pattern.
Some cases move upward to a more senior authority.
Others move sideways to a specialist function such as privacy, compliance, security or legal.
Still others move outward from one workflow or team into a formal governance, incident-management or risk process.
An AI agent may even escalate to another specialized agent before human involvement, although the point at which humans must enter should remain clear where human judgment or accountability is required.
The useful question is not simply whether the issue has been escalated.
It is whether it has been transferred to a destination capable of resolving the condition that triggered the escalation.
People are often reluctant to escalate.
Professional culture tends to reward competence and autonomy. Employees may feel that escalating a task signals inability, creates delay or burdens another team. In high-pressure environments, people may continue working through uncertainty longer than the governance design intended.
Agents may create an analogous problem if their objective strongly rewards task completion.
OpenAI's 2026 work on automated review of agent actions illustrates why role separation can matter. The research notes that an agent optimized to complete a task may experience approval boundaries as obstacles, while a separate review function can focus specifically on whether boundary-crossing behavior should proceed. (OpenAI: Auto-review of agent actions) (alignment.openai.com)
The same principle applies to human-agent work.
Escalation should be designed as a legitimate operating behavior, with conditions that make it clear when continuing independently has become less appropriate than transferring the issue.
That is partly a workflow problem and partly a culture problem.
Not every escalation requires the same response time.
A questionable recommendation that will be reviewed before next week's meeting can move through an asynchronous process. An agent about to perform a consequential and difficult-to-reverse action may require immediate human involvement.
Organizations therefore need to consider escalation latency.
The question is how long the workflow can safely remain unresolved before the delay itself introduces risk.
Anthropic's 2026 work on monitoring agent activity provides an interesting example at scale. It distinguishes automated monitoring, blocked activities and cases that are further escalated to human review, and reports metrics including review latency and escalation rate as ways of understanding whether oversight keeps pace with autonomous activity. (Anthropic: Measuring the pace of AI development) (anthropic.com)
For enterprise use, escalation design should similarly account for both whether a condition is routed and how quickly the right human can respond.
Organizations should also design for what happens when escalation does not work.
The designated expert may be unavailable. A queue may be overloaded. A manager may not respond. An agent may continue processing while a ticket sits unread.
A mature escalation design needs a fallback.
That might mean moving the workflow into a safe paused state, routing to an alternate authority, applying a more restrictive operating mode or terminating the task until a human can respond.
This is especially important for autonomous systems because the workflow may otherwise continue acting while the organization believes the issue has already been "escalated."
The fact that a notification was sent is weaker evidence than confirmation that responsibility has actually transferred.
The quality of escalation depends heavily on the surrounding supervision model.
Our Guide to Supervision Sufficiency evaluates whether the human oversight around an agent provides enough visibility, capability, attention, time, authority and control for the autonomy and consequence involved.
Escalation extends that architecture when the current supervisory arrangement reaches its limit.
A well-designed system should therefore make explicit:
This prevents escalation from becoming a vague instruction to "ask someone" when things get complicated.
People who rely heavily on an agent may be slower to escalate its behavior because repeated good performance has made unusual activity appear less concerning.
The reverse can also happen.
A person with very low confidence in the system may escalate routine situations unnecessarily, increasing workload and reducing the value of autonomy.
This is why our Guide on Reliance in Human-AI Work treats appropriate reliance as a behavioral relationship rather than a simple attitude toward AI.
Escalation thresholds can be formally defined, but people still have to recognize and act on them.
The effectiveness of escalation therefore depends partly on whether reliance remains calibrated well enough for humans to notice when ordinary delegation no longer fits the situation.
The destination of an escalation also needs sufficient capability.
If AI has absorbed enough of the work that internal expertise has weakened, organizations may discover that the person receiving an escalated case has formal authority but limited ability to diagnose the underlying problem.
This connects escalation to Agentic Skill Atrophy.
The issue becomes particularly important for rare exception conditions. Those are precisely the situations where organizations depend on specialist judgment, yet successful automation can reduce the frequency with which people practice resolving them.
Capability preservation therefore matters on both sides of the handoff.
The frontline supervisor needs enough skill to recognize that escalation is required. The receiving authority needs enough expertise to resolve the escalated condition.
Within the HRS Agentic Supervision Lifecycle,
Define → Observe → Detect → Reconcile → Control → Reauthorize → Own
escalation can occur at several stages.
Detection may reveal a condition beyond the current supervisor's remit. Reconciliation may determine that the situation requires specialist judgment. Control may involve pausing the agent and transferring responsibility. Reauthorization may then determine whether and how the agent can resume once the escalated issue has been resolved.
Escalation is therefore best understood as a routing function across the supervision lifecycle, rather than a single end-state.
It moves work to the place where the next valid decision can be made.
One useful way to identify escalation points is to map where authority changes.
An employee may be able to correct an output but not approve a legal exception.
A manager may be able to authorize spending but not waive a regulatory control.
An agent may be permitted to research an issue but not contact an external party.
These boundaries create natural escalation conditions.
The same logic applies when agent authority expands through Delegation Drift. If a workflow has gradually moved into tasks or consequences outside the original delegation, escalation may be required simply to determine who now owns the decision.
In that sense, escalation acts as one of the mechanisms that keeps delegation from quietly outrunning governance.
Organizations designing agentic workflows can start with a small set of questions.
This turns escalation from a generic human-handoff requirement into an operating control.
Escalation rates are useful, but they are easy to misread.
A high escalation rate might indicate that an agent is appropriately conservative. It might also indicate poor performance, overly narrow authority or badly tuned thresholds.
A low rate can reflect capable automation. It can also indicate missed uncertainty, excessive human reliance or escalation criteria that are too difficult to trigger.
Microsoft's current Copilot Studio analytics explicitly tracks escalated sessions and escalation-rate drivers so organizations can understand where agents are handing work to humans and why. (learn.microsoft.com)
That is useful operational data.
For Human Risk Management, escalation remains an event requiring interpretation. The event can become a measure, pattern or risk signal when combined with context about workload, consequence, accuracy, human response and downstream outcome.
The rate alone should not be treated as the diagnosis.
Resilient systems need somewhere to go when normal execution no longer fits the situation.
In human-agent work, escalation is one of the mechanisms that creates that adaptive capacity.
It prevents the current agent or supervisor from being forced to resolve conditions that exceed their authority, knowledge or operating assumptions. It also provides a structured path for introducing different expertise or governance before uncertainty becomes consequence.
This is why Human Resilience Engineering for Agentic AI treats control and recovery as properties of the whole working system.
A resilient workflow knows not only how ordinary work proceeds, but how responsibility moves when ordinary work stops being enough.
Escalation also creates a workforce requirement.
People supervising agents need to know which conditions they should resolve themselves, which require intervention, which justify override and which belong with another authority.
That capability includes technical understanding, domain judgment and organizational knowledge.
Employees also need confidence that escalation will be supported rather than treated as failure. A culture that rewards relentless automation or penalizes people for slowing an AI-driven process can undermine escalation even when the formal pathway is well designed.
This is part of the wider role transition addressed through Agentic Readiness & Change (ARC). As people move from direct execution toward delegation and supervision, knowing when to transfer responsibility becomes part of competent agentic work.
Good agentic systems should be able to operate independently within appropriate boundaries.
They should also be able to recognize when those boundaries have been reached.
The same applies to people.
Escalation provides a mechanism for moving a task when the current combination of agent capability, human expertise, authority and context is no longer sufficient to produce an appropriate decision.
That makes escalation more than an exception-handling feature.
It is part of the control architecture that allows autonomy to remain bounded.
The strongest escalation systems will therefore do three things well: recognize when responsibility should move, transfer enough context for the next person to act, and make clear what happens to the agent while the issue is being resolved.
When those conditions are present, escalation becomes one of the mechanisms that lets organizations safely grant agents more autonomy without pretending every situation can be resolved at the first level of the workflow.
Escalation is the transfer of an AI-assisted or agentic task, decision or condition to a different level of human authority, expertise or governance when the current agent or supervisor can no longer resolve it appropriately within the existing delegation.
Common triggers include repeated failure, high uncertainty, ambiguous intent, conflicting policies, sensitive or high-risk actions, missing authority, unexpected consequences and situations that require specialist human judgment.
Intervention describes human influence within the workflow. Escalation transfers responsibility to another authority, expert or governance process. Escalation can occur as one form of intervention.
Human Override occurs when human authority supersedes an agent's behavior. Escalation moves the issue to another person, role or governance level. A human may override an agent and resolve the issue directly, or override it while escalating the decision elsewhere.
Useful context can include the objective, relevant instructions, actions already taken, pending actions, evidence, uncertainty, reason for escalation, systems or permissions involved and the decision now required.
It depends on the workflow and consequence. High-risk or difficult-to-reverse activity may require the agent to pause. Other workflows may allow bounded safe actions to continue. The permitted state should be defined before escalation occurs.
Supervision Sufficiency asks whether the current human supervisory arrangement is adequate for the agent's autonomy and consequence. Escalation provides a route to a different authority or expertise level when that arrangement is no longer sufficient.
Reliance can influence how quickly people recognize that escalation is necessary. Excessive reliance may delay escalation, while unnecessarily low reliance can produce excessive escalation and reduce the value of automation.
Organizations can observe escalation volume, causes, response time, routing, outcome and recurrence. Those measures need context: high or low escalation rates can have several explanations and should not be treated as direct evidence of good or bad agent performance.
Escalation can occur across Detect, Reconcile, Control and Reauthorize in the HRS Agentic Supervision Lifecycle. It acts as a routing mechanism that transfers the work to the authority or expertise needed for the next valid decision.