Support automation without losing the human touch

Triaging tickets, suggested replies, and escalation rules—built so agents stay in control.

A growing software company needed to reduce first-response time without firing the knowledge base or angering power users. Leadership wanted automation that agents could override in one click. The program had to respect that senior agents often deviate from scripts for good reasons—any suggestion UI could not fight them for control.

Challenge

Previous experiments with a generic bot increased handle time because agents didn’t trust suggestions. Product worried about wrong answers on billing and security topics. Marketing feared brand tone drift if drafts sounded like every other SaaS chatbot.

Approach

Mapped top 40 ticket types; labeled which were safe for draft suggestions vs. required human-only handling.
Built retrieval over approved docs only; blocked categories for refunds and account deletion.
Shadow mode for two weeks: suggestions visible only to leads until quality thresholds held.
Added lightweight feedback buttons (“helpful / not helpful”) wired to weekly review—so quality improved without a heavy ML team.

Outcomes

Median first response improved while escalation rate stayed flat.
Agents adopted the tool after two training sessions—documentation owned by support leadership.
Quarterly review cadence for prompt and doc updates written into ops rhythm.
Power users kept shortcut macros; the AI layer augmented rather than replaced expert workflows.

Stack & constraints

Helpdesk API + vector store over internal Notion/Confluence exports; no training on customer messages.