IT Service Desk Agent Client Simulation

This is a 45-minute client simulation in which the interviewer plays a customer who wants an agent that takes actions, and answers only what is asked. Every action the customer names hides an approval, an audit trail or a data problem that surfaces only if you ask about it. With a partner, hand them this page and have them reveal a constraint only when your question earns it; alone, read the opening brief, write your questions down, then open one constraint at a time, and only for a topic you actually asked about.

The customer is the head of IT operations at a mid-sized manufacturer with several plants and a central office. He runs the service desk, which is partly in-house and partly outsourced.

"Our service desk is drowning. I want an AI agent that reads every ticket, fixes what it can, like password resets, access requests and software installs, and closes the ticket. People should only see what the agent can't do."

He cares about the backlog and response times, about plant downtime, and about getting through the next audit without findings.

"Right, thanks for making the time. The backlog is the problem. Tickets sit for days, the plants complain, and my team spend their time on password resets instead of real problems. Our ITSM platform has an AI add-on, but it only suggests articles and nobody reads them. I want something that actually does the work. I've got budget this year. What would you build?"

Discovery

• "Which ticket types make up most of the volume, and how long does each take today?"
• "Why aren't people using the reset portal?"
• "Which actions need an approval today, and who checks those approvals later?"
• "What's the most expensive ticket to get wrong?"
• "What can an integration do in your platform, and who has to sign that off?"

Framing back

• "So most of the volume is resets and unlocks, but plant staff can't verify themselves. Access changes need a manager's approval, and auditors check that. A misrouted line-down ticket can stop production. And write access needs the change board. Is that fair?"

Shaping a first slice

• "I'd start by helping your desk, not replacing it. The agent reads every new ticket, sets category and priority, sends anything that looks line-down straight to plant support, and drafts a reply from an article your desk trusts. Your agents accept or edit."
• "That needs no write access, so we don't wait for the change board."
• "On resets, the real fix may be getting plant staff enrolled. No agent should reset a password for someone it can't verify."
• "Closing tickets comes later, for reversible ticket types, once the numbers show it's safe."

Risks and next steps

• "The worst failure is a line-down ticket marked low. So we test on [number of historical tickets] past tickets first, and anything uncertain goes up, not down."
• "Access stays in your approval workflow. Later the agent can pre-fill the request, but a manager still approves."
• "Next steps: an export of last quarter's tickets, the articles your desk trusts, a change request for read access, and a session with your desk lead and the outsourcing manager."

Score each dimension from 1 to 5 with the scorer below; a 4 or 5 needs a quoted line of evidence, and a pass is no dimension below 2 and a mean of 3 or more.

• Designing automatic password resets without asking how the requester proves who they are.
• Treating an access request as a single tool call, when the approval and the audit trail are the real product.
• Routing tickets by CMDB owner without asking how current the CMDB is.
• Measuring the agent on tickets closed, which rewards wrong closures, instead of reopen rate and requester confirmation.
• Forgetting the desk team and their contract, so the people who must use the agent work against it.
• Letting the agent attempt fixes on line-down tickets instead of escalating them at once.