Skip to content
Flexday AI Docs
Sign in with GitHub

Use cases › Service Desk Copilot

Teams experience

What Harbourline employees see when they chat with the service desk bot in Microsoft Teams, and the rules the Bot and its Agent follow.

Written for
  • Everyone

Last reviewed

For most Harbourline employees, the Service Desk Copilot is simply a contact in Microsoft Teams. They ask for help in their own words. The Bot answers from Harbourline's IT knowledge, and when that is not enough it raises a complete ticket in ServiceNow and gives them the number.

A Teams personal chat: the employee reports a VPN problem, the Bot suggests two documented fixes with sources, the employee asks for a ticket and gives their email, the Bot asks for the site, then confirms INC0012345; alongside, four notes about knowledge, asking before acting, governed Flows and limited status look-ups
Figure: a conversation with the service desk bot.

What employees can do

They askThe Bot does
"How do I set up the VPN on a new laptop?"Searches IT knowledge and answers, with the articles it used under Sources.
"My laptop can't connect to the VPN."Suggests the documented fixes first.
"Please raise a ticket."Collects the work email, site, device and what was tried, confirms, then raises it and gives the number.
"What's happening with INC0012345?"Reports its state, assignment group and last update.
"Stop" (as the whole message)Stops any messages the Bot would start. Answers to their questions still come.

How a conversation works

  1. The message arrives. Microsoft delivers it to the Bot; its signature is checked and a duplicate delivery is dropped.
  2. The conversation continues. The same person in the same chat carries on the same session, so the Agent remembers what was said a moment ago.
  3. Teams shows the typing indicator while the Agent works.
  4. Guardrails run first. Personal data is masked, and a request to reset someone else's password is refused by a blocked topic before the model sees it.
  5. Knowledge before opinion. With grounding on, the Agent searches before it states a fact about Harbourline's systems, and says so when it cannot find one.
  6. Tools are limited to grants. Search IT knowledge, look up one incident by number, and run the Raise incident Flow. Nothing else.
  7. The answer arrives complete. A long answer is split into several messages at paragraph breaks.

What the Bot will not do

LimitWhy
Show forms or filesTeams conversations with a Bot are text only, so the Agent has forms switched off and gathers details by asking.
Show runbooksRunbooks are tagged it-staff, and Teams does not tell the platform a person's groups. IT staff use the dashboard's chat for those.
List someone's tickets by nameTeams identifies people by their Microsoft account, not their email, so look-ups are by ticket number only and return fields any employee may see.
Link to cited articlesCitations are names, because each document sits behind the platform's own access rules.
Message people who have never used itA conversation exists only after someone interacts with the Bot, or it is installed for them.

Messages the Bot starts

The P1 alert Flow messages on-call engineers in their personal chats with the Bot. Before any message the Bot starts:

  • The Bot must be live.
  • The person must not have opted out. Typing STOP (or UNSUBSCRIBE) as a whole message in a one-to-one chat opts out; START (or RESUME) opts back in. Removing the app also opts out.
  • The Solution's allowance of 60 messages an hour, shared by every sender, must not be used up.

Each message is recorded in the Bot's Deliveries view, and a retry never resends one that already went out.

Getting the Bot to people

  1. Harbourline's builders create the Bot with a managed registration, so Flexday AI holds the Microsoft registration.
  2. Verify, then Go live once the Agent is published and switched on.
  3. A Microsoft 365 administrator publishes it to Harbourline's Teams catalogue in one step, signing in for that publish only; nothing is stored.
  4. Harbourline's Teams admins pre-install the app for on-call engineers, so they have a conversation the P1 alert can reach, and the builders copy those conversations into the P1_ONCALL_CONVERSATIONS Variable.

Measuring it

The Agent's Analytics view shows conversations, people, the tools used, the articles cited, guardrail events and feedback. Harbourline reviews the questions that ended in a ticket each month and adds articles for the most common ones, which the Doc Base indexes automatically.