What does the myth say?
That customers who meet an AI in a support channel feel fobbed off, that they want a human and the AI is there to keep them from one, and that any company deploying AI support is trading satisfaction for cost. The belief is widely held, often from personal experience, and heads of support who hold it are protecting something real: the relationship their team has with customers on their worst day.
The examination is about which AI support the belief describes, and what a Digital FTE does differently.
Where does the belief come from?
From phone trees and chat bots designed to deflect: menus with no human option, bots that answer a different question than the one asked, loops that return to the start, and confident replies that are wrong. From deployments measured on deflection rate, which rewards keeping customers away from people. And from the experience of typing agent or human repeatedly into a chat window. The belief is a correct description of a design goal many companies chose.
It becomes a myth when applied to a design whose goal is the opposite.
What is true in it?
That a customer with a complaint, a refund, an account problem, or a difficult day does not want a machine and should not get one. That a wrong answer delivered fluently is worse than no answer. That a bot which cannot say I do not know will eventually say something harmful. That a company which measures deflection will build the thing customers hate. And that a Digital FTE deployed with a wide remit, a thin knowledge base, and promoted categories before the record supports them would produce exactly the experience the myth describes. The design has to be right, and the manager has to keep it right.
The honest claim is that the design makes the good version the default and the bad version hard.
What does the mechanism actually do?
A Customer Support Digital FTE on ftes.ai answers only what the knowledge base and policy pages answer, with a citation on every reply, and escalates anything else with a note that the base has a gap; it does not extrapolate. The always-escalate list, refunds, credits, account access, complaints, legal or safety language, distress, is fixed in the remit at every autonomy level, and each escalation reaches the person who owns that category with a four-part summary: what was asked, what was found, why it did not resolve, a suggested next step. Categories earn auto-send one at a time from a clean record; money never does. There is no menu, no loop, and no measure of deflection anywhere in the workspace.
The measures are first-line resolution per category, first response time, and escalation quality, which reward answering well and handing over well.
- Answer: from the base, cited, or escalate as a gap
- Escalate: fixed list, four-part summary, to the right person
- Measure: resolution, response time, escalation quality; never deflection
Why does a good handover change the experience?
Because the thing customers hate second most, after being trapped, is repeating themselves. An escalation that arrives as a summary means the person picks up with the context already known: what the customer asked, what was checked, why it needs a person. The customer is not re-asked for their order number. The team's time goes to the hard case rather than to reconstruction. And because the first line handles the known answers at any hour with a citation, the hard case reaches a person sooner than it did when the queue was full of password resets.
The escalation design piece in the library sets out the summary standard and the routing.
How can you check for yourself?
Run one high-volume, low-risk category gated for a month. Read every escalation and ask whether the summary would have let you pick up without re-asking anything. Read the send-back reasons and see whether they are about facts, which means the base needs work, or tone, which means the voice is transferring. Read the replies that cited an article and check the citations. Survey satisfaction in that category if you survey at all, and compare with the baseline; where you do not survey, the platform marks it unmeasured rather than claiming it. Promote nothing until the record is clean.
If the escalations arrive with context and the citations hold, the experience is the one the myth says is impossible.