Guide

The escalation rules are the product

Updated

Buyers compare voices, accuracy claims and integrations. The thing that actually determines whether customers have a good experience is a list most buyers never ask to see.

Why the edge dominates

The routine calls will go fine on any competent system: that is the easy part and every vendor demonstrates it.

What varies is the tail, and the tail is where your reputation sits. A confident wrong answer is worse than no answer, because the caller believes it and acts on it, and you find out later from a complaint rather than from a log.

What a good trigger list contains

An explicit request for a person, honoured immediately and without a negotiation loop. An unrecognised intent after one clarification rather than three. Detected distress or anger. Anything involving money, a complaint, a cancellation or a legal matter. A named account or VIP list. And any question about safety.

Six triggers, all of which can be written down. Ask for the list in the sales process, and ask what happens on each one out of hours.

Where the handoff goes

To a person who is actually there. An escalation to a queue nobody staffs at 9pm is a worse outcome than the AI saying it cannot help, because it adds a wait to the failure.

Establish the coverage hours of the human layer separately from the coverage hours of the AI layer. They are frequently different and vendors rarely volunteer that.

What you should be able to review

Transcripts, the containment rate, and specifically the calls that were not escalated but should have been. That last set is the one that tells you whether the configuration is right.

Ask how you get at it. A system you cannot audit is a system whose failures you learn about from customers.

Configure for handing off too often at first

The safe direction to be wrong in is escalating too much, and the safe sequence is to start cautious and tighten as the transcripts show what the system handles reliably.

Vendors sometimes push the other way, because a high containment rate is a good number in a review. It is your number to set, and the cost of the two errors is not symmetrical.

Compare on the things that actually differ

What each model handles, what happens at the edge, who is accountable, and what it may do outbound.

See the comparison