An AI copilot for maintenance that shows its working.
Most AI in this category is a chat box bolted onto a database. This one is a Model Context Protocol agent with tools, a scope it cannot widen, and an approval gate in front of every write.
What it does
Grounded answers with citations
Questions are answered from the tenant’s own read models and published language, retrieved through a vector index over every ingested entity. When an answer draws on retrieved records, those records are shown beside it, so an engineer can check it against the source instead of taking it on trust.
Two retrieval tools, not one pipeline
The agent chooses between semantic search over the vector index and a query against the equipment knowledge graph, in a bounded reasoning loop. Tenant and organisational scope are injected server-side by closure, so the model cannot reach outside them.
Read-only until you say otherwise
Write mode is opt-in, off at the start of every session, and warned. With it off the copilot cannot mutate anything at all.
Allow-listed actions, approved one by one
With write mode on, the copilot may propose an action from a server-side allow-list. Nothing runs until you approve that specific action, and each is approved on its own. The approved call goes through the same authorized use case the dashboard screen calls.
Root-cause explanations from work history
Ask why an asset keeps failing and the retrieval runs over that asset’s own closed job cards, certificates and status history, with those records shown so you can check the answer against them.
Downloadable reports and draft checklists
The copilot produces reports, extracts and draft plans as documents you can download. It is instructed to build them only from records retrieved out of your own tenant and to say so plainly when a fact is missing. Every one is a draft for a competent person to review, not an authority to act on.
Chat attachments are read, not absorbed
A file attached to a conversation is extracted to text for that conversation only. Promoting it into durable, searchable knowledge is a separate action you approve.
Your own history, under your control
Conversations and generated documents are owner-private. You browse, reopen, archive and delete your own.
How it fits the rest of the platform
The copilot owns no business rules. It reads the published language of every other context and, when you approve a write, calls that context’s own use case. Adding a module to a tenant’s subscription widens what the copilot can answer about, with no work on the copilot itself.
Questions about the copilot
Can the copilot change data on its own?
No. Write mode is off by default, and even with it on the copilot can only propose actions from a server-side allow-list. Nothing executes until a person approves that specific action, one at a time.
Where do its answers come from?
From the tenant’s own records, retrieved through a per-tenant vector index and equipment knowledge graph. Where an answer draws on retrieved records, those records are shown with it.
Does it see data outside my organisational scope?
No. Tenant and organisational scope are bound server-side into the tools the agent can call, rather than asked for in the prompt.
See it against your own equipment data.
A walkthrough on your asset classes, your certificate disciplines, your job cards, not a canned demo.