When to Route an Escalation to a Person vs. a Second AI Pass

Every trigger lands in one of three buckets. The bucket is decided by one question: what kind of failure is this?

The middle column is the one most teams underuse. A retry is cheap compared with a human minute, and it is often enough.

A second pass earns its place when the first failure was informational. That covers four common cases.

Missing or stale knowledge. When the agent answers only from an approved knowledge base, a question it has no fact for goes to a person rather than a guess (Elva). Before that handoff, one retrieval retry against the same approved sources is worth it, because the gap is often a retrieval miss rather than a true knowledge gap.

A tool that failed. A failed action fires when an integration call errors or times out, and the handoff should carry the action attempted, the error, and what the user saw (Userflow). A single retry of that tool, then a re-answer, resolves many of these without a person.

Ambiguous input. When the customer's sentence supports two readings, a second pass fed with the likely reading, or a single clarifying question, both beat a human handoff for a simple intake call.

A model that is weak on this task type. Some task classes are simply harder for some models. Cascades exist precisely to send hard queries to a stronger model (Concentrate). When the trigger records a task type rather than a content problem, route it to a model with a better record on that class instead of a person.

One rule makes this safe: a second pass only changes the input. It does not change the permission. If the first pass was not allowed to execute the action, the second pass is not either.

Some failures are about authority. More context cannot fix them because the missing ingredient is a human who can carry the consequence.

Support guidance is consistent on the non-negotiable topics: refund requests, safety concerns, formal complaints, legal or accessibility requests, and maintenance emergencies should always reach a person, with no confidence score required (Conduit). The reasoning is downstream consequence: a wrong or delayed answer here becomes a review, a chargeback, or a liability claim.

The same logic covers irreversible or protected actions and compliance flags such as a legal threat, health data, a payment dispute, or exposed personal information (Userflow). Weigh every action by reversibility before you decide who may take it. A draft email is reversible. A sent refund is not.

Two more cases belong here. A policy conflict between two approved sources is a judgment call about which source wins, and that is an owner's decision, not a model's. And a model that cannot articulate why it is unsure gives you nothing to act on. Research on verbalized confidence found the stated score separated right from wrong answers only weakly, with an average AUROC of about 62.7 percent in the study's tests (Userflow, citing Xiong et al.). A confidence number with no explanation is a coin toss dressed up as a metric.

Finally, an explicit request always wins. "Talk to a person," "this isn't helping," and even "forget it" should all count, and every request should be logged (Userflow). When a customer asks for a human, a second AI pass reads as evasion.

The short answer deserves a mechanical form. One retry, with a changed input, then a person if the bar is still not cleared.

def route(case, first_result):
    trigger = classify(first_result)      # why did it fail?

    if trigger in AUTHORITY_TRIGGERS:     # irreversible, money, legal,
        return handoff(case, first_result)  # policy conflict, explicit ask

    if trigger in INFO_TRIGGERS:          # missing context, tool failure,
        retry = second_pass(case, changed_input(trigger))
        if clears_bar(retry):             # same bar as the first pass
            return accept(retry)
        return handoff(case, retry)       # no third pass

    return accept(first_result)           # low score alone: clarify, do not escalate

Three details matter more than the code.

The bar does not move. The retry must clear the same quality threshold as the first pass. A lower bar on the second attempt just lets worse answers through.

The input must actually change. Re-sending the same prompt to the same model with the same context is not a second pass, it is a repeat. Change the retrieved documents, the model, the tool state, or the clarification, or do not retry at all.

The reason travels. Whatever decides the retry should record why it fired, and whatever leaves for a person should carry the transcript, the attempted steps, the customer context, and the trigger (Userflow). Without the reason, you cannot tune the rule later.

Take a support queue of 1,000 conversations a day. Suppose 180 of them raise a trigger. Assume, for illustration, that 40 percent of those triggers are informational and 60 percent are authority cases.

At these assumptions, the retry cuts human handoffs from 180 to 140, a reduction of about 22 percent. That is the prize, and it comes with two caveats. First, the retry adds a model call for every informational trigger, so the saving is net only when a human minute costs more than that call. Second, the retry must not be applied to the 108 authority cases, because a second model pass on a refund request has no upside and some downside.

The numbers are illustrative. The decision rule is not: measure your own mix of triggers, then check whether the retry resolves enough cases to justify the call.

An escalation rule that cannot be inspected cannot be improved. Three practices keep it honest.

Record every route and why it ran, including each attempted step, because evaluation has to cover the whole policy rather than one model call (AI Engineer). If you only log the final answer, you cannot tell whether the retry helped or merely added latency.

Limit how many times an agent may ask a caller to repeat, and offer another path such as a transfer when the limit is reached (VocalOps). A loop with a confused caller is worse than a clean handoff.

Run new scenarios in a confirmation mode first. One receptionist product lets staff approve proposed actions before anything writes to the system, then switches to automatic once the judgment has proven out over real calls (Elva). Apply the same pattern to the retry: watch it decide for a week before you let it route unsupervised.

If your queue runs on a mixture of models and task types, a single endpoint for text, image, audio, and embedding calls, prepaid and priced in dollars, keeps the retry logic simple, and Quanta Inference offers that through one API key. Bring the routing decisions into review with your own support data before you tune thresholds.