Deflection Baseline Audit (Onboarding)

Deflection Baseline Audit (Onboarding)

1. 01

   Pull 90 days of raw ticket exports before any routing logic is configured

   Request the full export from the client's helpdesk, not a filtered view. You need channel, timestamp, requester, first-response time, resolution time, and the raw subject and body text for every record.

2. 02

   Tag a 200-ticket random sample by hand into intent buckets

   Two people tag independently, then reconcile disagreements. The disagreement rate is your realistic ceiling for automated classification accuracy, so record it before a vendor demo sets expectations.

3. 03

   Separate volume that is genuinely deflectable from volume that only looks deflectable

   Password resets and order-status questions usually qualify. Billing disputes, cancellation threats, and anything mentioning legal or regulatory language do not, regardless of how repetitive they appear in the export.

4. 04

   Measure the current cost per ticket across the sample

   Multiply average handle time by a loaded hourly rate for the tier that actually touches the ticket. A 4-minute average at a $38 loaded rate gives the client a number they can compare against any platform quote.

5. 05

   Write the escalation boundary into the client's own words

   List the exact conditions that force a human: sentiment below a stated threshold, a named enterprise account, a second contact on the same issue within 24 hours, or any message containing a refund demand above a set dollar figure.

6. 06

   Confirm the client's helpdesk exposes the fields your routing rules depend on

   If priority, account tier, or SLA clock lives in a custom field that the API does not return, the routing logic will silently misroute. Verify field-level access before scoping the build.

7. 07

   Publish a one-page baseline the client signs off on

   Include ticket volume, current first-response median, deflection ceiling estimate, and the escalation rules. This document becomes the reference point when someone later claims the automation 'isn't working.'