Escalation Boundary Design (Onboarding)

Escalation Boundary Design (Onboarding)

1. 01

   Inventory every inquiry type the client's front line handled in the last 90 days

   Pull chat transcripts, email threads, and call notes, then bucket them into intents such as pricing, order status, booking, and complaint. The buckets become the raw material for deciding what a bot may answer alone.

2. 02

   Score each intent on two axes: answer confidence and consequence of error

   A store-hours question is low consequence; a refund above a set dollar threshold or a contract term question is high consequence. Intents scoring high on consequence route to a human regardless of how well the model answers.

3. 03

   Write the escalation triggers as literal conditions, not judgment calls

   Examples: customer types the word cancel, sentiment drops twice in one session, the bot cannot cite a source for its answer, or the thread exceeds six turns without resolution. Conditions a reviewer can test beat instructions like use good judgment.

4. 04

   Define the human handoff payload so the agent inherits full context

   The receiving rep should see the transcript, the detected intent, the customer's account record, and what the bot already promised. Platforms such as Chatling and FastBots support context-carrying handoff, so configure it rather than rebuilding it per client.

5. 05

   Set a fallback path for after-hours and queue-overflow windows

   Decide whether the bot captures a callback request, opens a ticket, or states a response window. Leaving this undefined is where clients form the impression that the bot simply ignores people.

6. 06

   Agree on the metrics that define a failed conversation before launch

   Track containment rate, escalation rate, and post-escalation satisfaction separately. A high containment rate with low satisfaction usually means the bot is trapping people rather than resolving them.

7. 07

   Get written client sign-off on the boundary document

   The signed version becomes the reference when a client later asks why the bot did not handle a specific case. It also sets the scope for the optimization retainer.