Can AI safely handle a share of 20,000 support tickets a month?

A design exercise showing how I would run discovery and a pilot. No ticket data has been analysed, no system has been built, and no automation rate is assumed.

Context

A SaaS business handles about 20,000 support tickets a month and response times are rising. The COO wants to explore whether AI can take on a meaningful share without damaging customer experience or security.

Problem

Rising response times, with leadership pressure to automate. But some tickets touch customer data, billing and account changes, where a wrong answer or an unapproved action is expensive.

Approach

Segment tickets by volume, complexity and risk. Separate read-only answers from actions that change account or billing state. Route by intent, risk and confidence. Keep approvals and an audit trail on anything that changes an account. Validate it all with a shadow-mode pilot.

What I did

Scenario analysis, discovery plan, architecture, guardrails and pilot design. The approval, scoping and audit patterns draw on what I built in Snitch (CS-03).

Key decisions

Analyse ticket data before choosing a solution	How much can be automated is unknown until the tickets are segmented.
Separate reading from changing	Read-only answers and state-changing actions carry very different risks.
Proposals and human approval for any account change	The worst case becomes a rejected proposal, not a wrong refund.
Shadow mode before automation	Measures quality on real tickets without customers bearing the risk.

Risks managed

Wrong customer’s data exposed
Unapproved or duplicated account changes
Confident wrong answers from stale documentation
High-risk tickets misrouted to automation
Prompt injection via ticket text
Customer trust damaged by poor automated replies

How success would be measured

Automation rate per category, not reopened
Resolution quality on QA samples
Reopen and escalation rates vs baseline
CSAT vs baseline
Routing accuracy on high-risk tickets
Safety incidents, with a target of zero

Intended outcome

A defensible answer to “how much can we automate, where, and at what risk?”, backed by pilot data instead of an assumed percentage.