AI
AI Agents
The digital colleagues that do work for you — an AI Agent is a configured role with a specific set of actions in Gfacility, from advising to acting autonomously.
Updated May 18, 2026
Configuration · AI · 8.1
An AI Agent is a configured “digital colleague” with a bounded role — a ticket classifier, a KB assistant, a no-show detector. Per agent you record which actions it may take, at what level (advise / assist / act), and which system context it uses.
Why this matters to the business
"AI everywhere or nowhere"
Agents per use case → AI exactly where it adds value, not as a one-size-fits-all.
"Handler loses control"
Level per agent → advise (human decides) to act (human reviews after).
"AI hallucinates"
System context (see 8.3) restricts what the agent "knows" — only your KB and data, no wild inventions.
"Costs spiral"
Consumption control (see 8.4) caps per agent — no surprises on the invoice.
What do you configure per agent?
| Field | What it's for |
|---|---|
| Name & purpose | "Ticket Classifier", "KB Assistant" — purpose clear to admins and end users (transparency). |
| Scope | Which object types does it act on (tickets, bookings, KB articles)? |
| Use case (see 8.2) | Which specific tasks does it perform? One or more. |
| System context (see 8.3) | Which prompt and which data sources it may use to answer. |
| Level | Advise · Assist · Act. Determines whether human approval is needed. |
| Trigger | When does it kick in? On ticket creation, on status change, on demand? |
| Consumption limit | How many requests / tokens per day, week, month? Cap on cost and usage. |
| Status | Active · Test · Disabled. Quick switch-off when needed. |
Agent examples
Ticket Classifier
Reads the report, suggests category + priority + workgroup. Usually starts at level "advise".
KB Assistant
Finds KB articles matching the issue and proposes an answer to the handler or requester.
No-show detector
Tracks check-in patterns, flags risk bookings, can (at level 3) auto-release them.
Catering Optimizer
Suggests catering quantities based on historical no-shows and attendance.
Anomaly Detector
Spots patterns in tickets (sudden 30+ "WiFi issues") and alerts service management.
Communications Drafter
Drafts email replies for the handler — who edits and sends.
Which decisions will you make?
Which agents go live?
Start with 1-3 at "advise" for low-risk use cases. Scale after measurement (see 6.3).
Transparency
Show end users that AI is answering? Label, badge, footnote — required by the AI Act for some categories.
Approval per agent
Promotion to level 2/3 by steering-group decision. No ad-hoc upgrade by one admin.
Roll-back procedure
Status "Disabled" stops the agent immediately. Who's allowed to do that?