Escalation is a routing decision

The escalation rate is a composite of things with different causes

The number on most support dashboards adds up events that have nothing in common. A ticket that went to a human because the system could not read a carrier page sits in the same bucket as a ticket that went to the brand because the brand has no policy for it. One is a capability gap. The other is a governance boundary working as intended. A single percentage cannot move in a direction that is good for both.

Driving that percentage down rewards suppression. The cheapest way to lower an escalation rate is to answer more things, and the cheapest way to do that is to lower the bar for what counts as an answer. The measurement that survives an actual operation is per-ticket: for this case, was the party that received it the party that could decide it.

The first cut is not a number at all. Informational cases resolve at the AI tier by design — Razviel's architecture routes by the nature of the case, never by a confidence value. Escalation, properly speaking, belongs to the cases that have to be decided rather than looked up.

A handoff is an artifact, not a transfer

The difference between a good escalation and a bad one is entirely in what arrives with the case. A bad escalation is a forward: the customer's last message moves to a queue, and the party that receives it starts over — asks for the order number again, asks what went wrong again, reads none of what came before because none of it traveled. The customer pays twice, once in time and once in learning that nobody was listening.

A chewed handoff is the opposite. The design captures context unconditionally, before any decision — before resolving and before escalating — so by the time a route is chosen the case is already assembled: the order, the customer's history, the policy cell consulted, what was attempted, what the customer wants, and why the case is moving rather than closing. The receiving party opens a case, not a complaint.

The design also makes the transcript the state. There is no partial state to hand across and no summary that can drift from what happened. Whoever picks the case up reads the record and knows where it stopped.

There is one setting where this has been measured. In 2014, Starmer and colleagues reported a nine-hospital study of a structured handoff program — a fixed format for what one clinician tells the next, whose final element is a synthesis by the receiver. Across 10,740 admissions medical errors fell 23 percent, and the handoffs did not take longer: 2.4 minutes per patient before, 2.5 after. The intervention was not more diligence. It was a format.

The numbers belong to pediatric wards and transfer to nothing. What transfers is the mechanism — and it applies here harder than it applied there. A format matters in proportion to how little the receiver can ask the sender. Between two clinicians that channel is narrow. Between a model and a person it is closed: the AI that worked the case does not stay available to be questioned about what it meant. It resolved or it routed, and it is gone. Whatever it did not write down cannot be recovered by asking for it.

That is the argument for specifying the format rather than trusting the summary. A human who hands off badly can be interrupted and asked; a model cannot, and the case arrives exactly as complete as the format required it to be.

Merit and capacity are two different escalations

An escalation by merit means the case requires the party it went to. Either no policy covers it, or the brand has claimed that class of case for itself. A merit escalation never comes back. Sending it back would mean sending it to a party that has already established it cannot decide it.

An escalation by capacity means nobody was available. The case was decidable at its tier; the tier was full. A capacity escalation is temporary by definition, and it reverts when capacity returns.

Keeping them apart takes two design decisions. Every escalation carries a typed reason, so the difference is recorded rather than inferred. And the queue is single: every active ticket carries one number that spans the human tier and the brand tier. A capacity escalation keeps its number, parks, and returns to its original position when a human is free.

The alternative is worse than it sounds. With two queues, a ticket parked in the second and returned to the first arrives behind everything that entered while it waited. That customer is penalized for the operation's staffing rather than anything about their case — and the penalty is invisible on the report, because the ticket was escalated, then handled.

A missing answer is not a missing context

The two real failures named at the top are opposite errors with one root: treating I cannot answer this and I do not know enough about this as the same state.

They are not. A system can capture everything a decision requires and still have no rule to decide by. That is a common case, not an edge case.

Take a customer who writes in proposing an influencer partnership to a store that runs no partnership program. The architecture captures all of it — who they are, the audience, the terms proposed, how to reach them. Nothing is missing. There is simply no cell in the brand's policy that says what a partnership offer is worth. The answer is absent; the context is complete.

That escalation is grounded, not a loop: it reaches the brand with everything the brand needs and a reason that names the gap. The loop is the other behavior — three more clarifying questions in the hope that a rule appears.

The reverse error has its own guard. A ticket with no context does not escalate at all — it does not advance past the gatekeeper. A system that escalates on insistence teaches people to insist. Requiring context first lets abuse filter itself out: refusing the triage produces nothing.

A gap in the policy belongs to the brand, by design

The obvious objection to routing an uncovered case to the brand is that it looks like passing the buck. The AI could not answer, so it goes to a person; the person could not answer, so it goes to the client. Three steps to arrive nowhere.

The objection dissolves on the question of what the human operator would do with it. A human handling a ticket acts on the brand's policy too — that is the entire basis on which they are allowed to act. Hand them a case no rule covers and they have nothing to follow. Their options are to invent something or to escalate, and inventing is the worse of the two: an invented answer becomes a precedent the brand never agreed to and will meet again.

So the routing rule is not a fallback. Where there is no policy, the design routes the case past the human tier and to the brand directly, because the brand is the only party whose answer is also a rule. The brand does not merely decide this case. It decides the class.

That has a compounding effect a defect metric would hide. Uncovered cases become readable volume at the brand tier, grouped by what they have in common, and each group names a gap in the policy matrix. The matrix is designed as a living document: the brand describes the new rule in plain language, the cell fills, and that class of case stops escalating. The rate falls as a consequence of the policy becoming complete, not as a target.

The takeaway

An escalation is a routing decision: it sends each case to the party that can decide it. The brand decides what only the brand can decide, a human operator applies judgment inside a rule that exists, and the AI resolves the rest. The real failures are escalating without context and improvising where no policy exists. For a Shopify Plus brand, a falling escalation rate is good news only when the policy became more complete, not when the bar for an answer moved.

What to check, per escalation rather than per dashboard:

1. The reason, typed. Every escalation records why it moved: no policy, the brand’s claim on that class of case, a capability gap, or no capacity. A rate that sums them can only be read once it is split.
2. What arrives with the case. The order, the customer’s history, the policy cell consulted, what was attempted, what the customer wants, and why the case is moving. A handoff that carries only the last message hands over the work of finding out.
3. Merit or capacity. A merit escalation never comes back. A capacity escalation keeps its number in a single queue and returns to its position when a person is free. With two queues, the customer pays for the operation’s staffing, and the report does not show it.
4. Where an uncovered case goes. A case no policy covers goes to the brand, whose answer becomes the rule for the whole class. A human operator handed the same case can only invent a precedent.
5. Whether the gaps close. Uncovered cases, grouped by what they have in common, name the gaps in the policy matrix. The rate falls as those cells fill — the one direction in which a falling rate means the operation improved.

A system that escalates well and a system that escalates rarely are not the same system, and only one of them is safe to grow.