An AI agent that triages support tickets around the clock
An MSP drowning in incoming tickets. We put an AI agent on the front line: it answers the routine ones and escalates the rest with full context.
Details anonymized and adapted for confidentiality.
The problem
An MSP of around forty people. Tickets coming in by email, by portal, by transcribed phone call. Roughly 180 tickets a day, and a big half of them are the same requests: password resets, VPN that won’t connect, missing printer, access to a network share.
The level-1 team spent its days on this. Average first-response time: 3h12 during business hours, and after hours or on weekends nothing moved until the next day. The genuinely urgent tickets drowned in the pile, and senior techs wasted real time re-reading badly qualified tickets just to understand the problem.
What we built
An AI agent wired straight into the ticketing tool, taking every incoming ticket on the front line, 24/7.
Triage and classification - On arrival, the agent reads the ticket, identifies the request type, the real priority and the client involved, then applies the SLAs. A VPN down for 40 people no longer lands in the same queue as a software request.
Autonomous handling of the routine - For known requests, the agent answers directly and executes where it can via internal APIs (AD reset, account unlock, up-to-date doc link). The client gets an answer in minutes, at 3am as easily as at noon.
Escalation with full context - Anything the agent can’t close itself goes to the right tech with a clear summary: what the client is asking, what’s already been tried, the missing info requested upfront, and the client history. No more unreadable two-line tickets.
Guardrails and traceability - No sensitive action without an explicit rule. Every agent decision is logged and reviewable, and a human keeps control over anything touching critical production.
The result
- First-response time dropped from 3h12 to under 4 minutes on incoming tickets, day and night.
- About 46% of tickets resolved with no human involvement, essentially the entire routine load.
- Level 1 freed up for real work: fewer password resets, more time on cases that need a brain.
- Escalated tickets usable on arrival: seniors no longer burn 10 minutes qualifying before they diagnose.
- 24/7 coverage without bloating on-call, nights and weekends no longer pile up until Monday.
The stack
- LLM agent (Anthropic API) orchestrated on a Python backend, with prompts and business rules versioned.
- Direct integration with the ticketing tool and Active Directory via API.
- Classification, prioritization and SLA application driven by explicit rules, not a black box.
- Escalation queue enriched automatically, every action logged and auditable.
- Human in the loop on anything sensitive or irreversible.