What changes when the chatbot is retired?
A support chatbot matches what a customer types to an intent, serves a canned answer or a menu, and is measured on how many conversations it kept away from an agent. A Customer Support Digital FTE on ftes.ai reads the ticket, answers from the knowledge base with a citation or escalates as a gap, routes anything on the always-escalate list to a person with a four-part summary, and is measured on first-line resolution per category with escalation quality beside it. The migration keeps what the chatbot learned about what customers ask and drops the design goal that made customers hate it.
The comparison of a Digital FTE with an AI agent and a chatbot in the library gives the three-column view; this guide is about the move.
What should you keep from the chatbot?
Its analytics: the intents customers actually hit, in what volume, and where they abandoned, which is the best available map of which categories to migrate first and which articles are missing. Its intent-to-article mapping, where one exists, as a starting point for confirming the knowledge base coverage per category. Its escalation triggers, as a starting list for the always-escalate list, which will grow. And, for a short overlap, the chatbot itself on the categories not yet migrated, if it can be scoped that way.
Keep nothing of its conversational design. Menus, loops, and dead ends are what the migration is for.
What moves to the FTE, and in what order?
First, record the baseline per category from the helpdesk and the chatbot's analytics: first-line resolution where it can be computed, first response time, abandonment, and complaint volume about the bot itself. Second, write the always-escalate list with receivers and the summary standard, and confirm the articles for one high-volume category the chatbot's analytics show was hit most. Third, connect the helpdesk with a dedicated scoped account, route that category to the FTE, and retire the chatbot for it, keeping a direct path to a person visible to customers. Fourth, run gated for a month, approving cited replies and the articles proposed from gaps. Fifth, at the first review, move the next category and consider the Promotion Standard for the first.
Retire the chatbot category by category, not all at once, unless its harm is such that immediate retirement is the safer choice.
- Baseline from the helpdesk and the chatbot's analytics
- Escalate list and summary standard first
- One category migrated and the bot retired for it; a path to a person always visible
What breaks in the first month?
The deflection dashboard, and any target anyone had on it. The FTE does not report deflection, because deflection rewards keeping customers away from people; it reports first-line resolution, which rewards answering, and escalation quality, which rewards handing over well. A team that had a deflection target needs a new target from the first week. The knowledge base breaks too, as in every support migration: the chatbot's canned answers hid gaps the FTE's citations expose, and the gap escalations in week one become proposed articles. And customers' expectations break in a good direction: some will type agent out of habit, and the FTE's reply, with a citation or a summary to a person, is the first time the habit was unnecessary.
The myth piece on AI support making customers angry describes what the chatbot was doing.
What does the record show after a month?
First-line resolution for the migrated category against the baseline. First response time including overnight, which the chatbot also covered but without answers. Escalation volume by reason and escalation quality, which the chatbot could not produce because it had no summaries. Knowledge gaps closed as proposed articles approved. Complaints about the bot itself, from the helpdesk's tags, against the baseline. Approval rate with send-back reasons, showing whether the category's voice has settled. And cost per resolved ticket from the role's usage against the chatbot's licence plus the agent time it deflected into later.
The metric explainer on first-line resolution describes why deflection was the wrong number.
What should you refuse during the migration?
Keeping a deflection target for the FTE. Migrating every category at once, because the knowledge base cannot absorb the gaps. Removing the visible path to a person, ever. Promoting a category to auto-send in the first month. Reusing the chatbot's canned answers as knowledge base articles without checking them against the fact gate's sources, because canned answers were written to end conversations rather than to be true. And measuring on a quiet month.
The escalation grid framework describes the routing the migration writes on day one.