Escalation Pathways
What is Escalation Pathways?
Agents will encounter situations they can't handle - ambiguous instructions, conflicting information, high-stakes decisions they're not authorized to make, or tasks that exceed their capabilities. The agent needs a structured way to escalate to the human without breaking the workflow, losing context, or creating anxiety. This is different from simple error recovery because the agent hasn't failed - it's recognized its own limitations. The pattern defines four escalation types: confidence-based (uncertainty threshold), permission-based (authorization limits), conflict-based (contradictory information), and capability-based (task exceeds abilities). Each escalation preserves full context, includes a recommended action with confidence level, and allows the agent to continue from where it paused after the user responds.
Problem
Agents encounter situations they can't handle - ambiguity, conflicts, authorization limits, or capability gaps. Poor escalation design either interrupts users too frequently (escalation fatigue) or too rarely (the agent guesses wrong on high-stakes decisions).
Solution
Design structured escalation triggers with context preservation, recommended actions with confidence levels, and multiple response options. Batch non-urgent escalations, learn from repeated answers, and let users set escalation sensitivity.
Real-World Escalation Pathways Examples
Implementation
When to use Escalation Pathways, and when it backfires
Use it when
- The decision is above the agent's pay grade: it is high-stakes, irreversible, or outside the authorization you granted, and a wrong guess costs more than a pause.
- The agent's own signals say it is out of depth: confidence is below threshold, the inputs conflict, or the task needs a capability it doesn't have.
- A human can actually resolve it faster than the agent can flail. The handoff buys a real answer, not just a place to park the problem.
Don't, or minimize, when
- The agent can recover on its own: retry, re-read, or pick the obvious default. Escalating a decision you could have made is just offloading work onto the user.
- There is no human on the other end who can act, or they have no more context than the agent. A handoff into an empty room is worse than a guess.
- The cost of being wrong is trivial and reversible. Interrupting someone to confirm a low-stakes, undoable action is how you train them to ignore the next escalation that matters.
The trap
The escalation to nowhere. The agent throws up its hands, drops the user into a ticket queue, a closed support form, or worse, loops them straight back to the same bot, and discards everything it knew. The human now starts from zero: re-explaining the problem the agent already understood, re-supplying context the agent already had. A handoff that loses state isn't an escalation, it's an eviction. It is worse than the agent guessing, because at least a guess keeps moving. Just as bad is escalating too late, after the agent has already sent the email, deleted the files, or charged the card, and the 'should I proceed?' arrives as a postmortem.
Take it into your own product
- 1
Escalation is not failure. It's the agent knowing its own edges.
Error recovery is for when the agent broke. Escalation is for when it didn't: the task is just above its authority, its confidence, or its capability. Frame the handoff as a competent colleague asking a question, not an error state apologizing. An agent that never escalates isn't confident, it's reckless.
- 2
Pause before the irreversible step, not after it.
An escalation that arrives after the email is sent or the card is charged isn't a decision point, it's a confession. The whole value is catching the high-stakes action while it can still be redirected. If 'should I proceed?' shows up as a postmortem, you escalated too late and the pattern bought you nothing.
- 3
A handoff that loses context is an eviction, not an escalation.
The single thing that makes escalation worth more than a guess is continuity. Carry the full state across: what the agent did, what it completed, the decision needed, the recommended action. If the human has to re-ask the user what the agent already knew, you've just added a worse middleman than no agent at all.
- 4
Send the recommendation, not a blank question.
'What should I do?' makes the user do the agent's thinking. 'I'd route this to Legal, 62% confident. Approve, or tell me otherwise' makes the decision a one-tap confirmation. The recommendation plus a confidence number is what turns an interruption into a quick yes.
- 5
Tune the volume, and learn from the answers.
Too many escalations and users rubber-stamp everything, including the one that mattered. Too few and the agent guesses wrong on the stakes. Let users set sensitivity, batch the non-urgent ones, and when the same answer comes back three times, offer to automate it. An escalation you never stop asking is a decision you failed to learn.
Add Escalation Pathways to your product
Copy the prompt below into Claude Code or Cursor in your repo. It encodes the four moves on the left and asks Claude to find your AI decision surfaces and update them. Claude reports what it changed and asks before adding dependencies.
Check if your product already has this pattern
Upload a screenshot. We'll tell you which of the 36 patterns your AI interface uses and where the gaps are.
Audit My DesignMore in Human-AI Collaboration
Contextual Assistance
Offer timely, proactive help and suggestions based on user context, history, and needs.
Human-in-the-Loop
Balance automation with human oversight for critical decisions, ensuring AI augments human judgment.
Augmented Creation
Empower users to create content with AI as a collaborative partner.
