Design an Agentic Support Ticket Triage System
Let an AI agent investigate each new support ticket with read-only tools and set its category, priority, team queue and links, sending uncertain cases to people, never overwriting a person's change, and keeping every ticket within a 10-minute triage target through an outage-sized surge and a fixed model quota.
Problem statement
Design an AI agent system that triages new support tickets.
For this exercise, assume a software company's support desk uses an existing ticketing system. Customers open tickets by email or a web form. The ticketing system sends a webhook, at least once, for every new ticket and for every change to a ticket's fields, saying who made the change.
Its API can read a ticket and update the ticket's triage fields; each update names the ticket version it was based on and is refused if the ticket has changed since.
Today support staff triage every ticket by hand: they choose one of about 40 categories, a priority from P1 (urgent) to P4, and one of about 30 team queues, link the ticket to an open incident or to an open ticket it duplicates, and write a two-line internal summary.
The agent should do this work. For this exercise, also assume it may call read-only tools: look up a customer's account and plan, search recent tickets, list open incidents from the status service, and search the help articles.
It may change only the triage fields and add internal notes; it never replies to the customer, closes a ticket or changes anything else. Ticket text comes from customers and can contain anything, including instructions aimed at the agent.
When the agent is unsure, its suggestion goes to a human triage queue instead of being applied. A person can change any triage field at any time, and a person's change always wins over the agent's.
Every new ticket must be triaged, by the agent or by a person, within 10 minutes of arriving. Staff should be able to see why the agent chose what it did, and the team needs a safe way to measure the agent and to try a new prompt or model.
What to build
- Every new ticket gets a category, priority, team queue and internal summary within 10 minutes, set by the agent or, when the agent is unsure, by a person working from the agent's suggestion.
- Before deciding, the agent can look up the customer, similar recent tickets, open incidents and help articles, and it links a ticket to an open incident or to an open ticket it duplicates.
- The agent changes only triage fields and internal notes, never overwrites a person's change, and is not steered by instructions inside ticket text.
- Staff can see why the agent chose each field; every correction by a person is recorded so the agent's accuracy can be measured and a new prompt or model can be tried safely.
At scale
- For this exercise, assume about 100,000 new tickets a day, about 1 a second on average and up to about 4 a second in busy hours.
- For this exercise, assume one agent run makes about 4 model calls using about 8,000 tokens in total and usually finishes within a minute, and that this system's model quota is fixed at 4 million tokens a minute.
- For this exercise, assume a major outage can bring an extra 20,000 tickets within 30 minutes, most of them about the same problem, on top of the normal flow.
- For this exercise, assume the human triage team can handle about 1 ticket a second in total.