An assistant that makes things up is worse than none
A confident wrong answer costs more than a missing one: someone acts on it, and now your support desk created the incident instead of resolving it. So we built the assistant around one rule - it only says what it can back up.
Grounded in your knowledge, not the internet
When a user asks a question, the assistant searches your workspace first:
- Your knowledge base. Articles are automatically chunked and vector-embedded, so retrieval is semantic - a question about "expense reimbursement" finds the article titled "Declaring costs".
- Your active service announcements. If maintenance is planned or an outage is ongoing, the assistant says so before anyone files a duplicate incident.
Answers come with citations to the articles and announcements they were drawn from. Users can click through and verify. The model is Gemini on Vertex AI, hosted in the EU region on Google Cloud, and responses stream in as they are generated.
"I don't know" is a feature
When retrieval comes back empty, the assistant does not improvise a plausible answer from its general training. It says it does not know - and offers the useful next step: drafting an incident so a human agent picks it up. A knowledge gap becomes a signal for your team, not a hallucination for your user.
Permissions are enforced in the data layer, not the prompt
Here is the part most AI bolt-ons get wrong. Telling a model "do not reveal restricted information" is a suggestion, not a control. Prompts can be talked around.
Our assistant runs as the logged-in user:
- When it looks up incidents, the query carries that user's role and identity. A member asking "what is the status of my requests?" sees only their own. An agent sees what an agent may see. The assistant cannot retrieve what the user could not open in the UI.
- When it searches the knowledge base and document library, role-based visibility applies to the search itself. Restricted content is never in the assistant's context to begin with.
There is nothing to jailbreak, because there is nothing hidden behind a polite instruction. The same access rules that protect your API protect your AI.
Confirm before create
The assistant can turn a chat into a structured incident - but it never files one on its own. The flow is deliberate:
- The user describes the problem in plain language.
- The assistant drafts the incident: title, description, a suggested category and priority.
- The user reviews the draft and confirms - or edits it, or cancels.
Only after that confirmation does the incident exist. No surprise tickets, no AI acting behind your back, and a clean record of what the user actually approved.
That is the bar we hold the assistant to: cited when it answers, honest when it cannot, and never holding more access than the human it serves.