Assisted inbound reply triage (L2)
Form fills, chat transcripts, and inbox replies get classified and summarized before a human touches them, so real buying questions do not sit behind vendor pitches and job applications.
The steps
- 01
Write the classification set
Tool: Manual
Five categories maximum: buying question, support, partner, vendor pitch, noise. More categories means lower agreement and slower triage. Owner: demand gen lead. DoD: categories defined with two examples each.
- 02
Measure agreement before trusting it
Tool: ChatGPT
Classify two hundred historical inbounds and compare to human labels. Publish the agreement rate per category and keep humans on any category below eighty percent. Pitfall: routing on an unmeasured classifier and losing real buyers in the vendor bucket. DoD: agreement rate published per category.
- 03
Summarize with the original attached
Tool: HubSpot
Every routed item shows the summary and the full original text. Reps must be able to read the raw message in one click. Owner: marketing ops. DoD: summary and original both visible on the record.
- 04
Sample the noise bucket weekly
Tool: Slack
Read twenty noise classified items each week. The cost of this play is entirely in false negatives, so inspect them deliberately. Owner: demand gen lead. DoD: weekly noise sample with misroutes logged.
Tools in this playbook
Building the team behind this workflow?
Next playbooks
Unfamiliar terms are defined in the AI and Revenue Dictionary. Related frameworks live in the framework library.