AI Chat
AI Chat answers questions about the project you are in, by going and reading the project. Not a search box and not a chatbot with a manual: you ask what is waiting on your approval, and it looks at the workflow runs, finds the six bookings, and tells you which they are and who raised them.
It is in the project sidebar under AI Chat.
One project at a time
A conversation belongs to a project, and every question in it is answered from that project's data. There is no cross-project chat, and no conversation carries over when you switch jobs. Change project and you get that project's chat.
What it can reach
The blank state says so out loud, in a row of chips under It can reach:
- Contracts, claims, variations, budget and the approvals waiting on you
- Drawings, the register, the text on the sheets, whiteboards and photos
- Mail, correspondence, site messages and submittals
- Procurement, packages out to tender and what closes next
- Program and site, lookahead, the site diary, meetings and bookings
- Safety and quality, inspections, observations, incidents, permits, SWMS and who is on site
That is a complete statement rather than a sample. What sits behind each chip, and what is deliberately outside all six, is in What the AI can read.
Opening a new chat
A new conversation opens on What can I help with? and a band headed Waiting on you, which is not a set of examples: those cards are built from your own pending approvals and assigned work, so the first thing you see is the real list. When there is nothing waiting, five general starters take their place.
Those cards are read straight from the database, so opening the chat page costs nothing in AI credits. Only a question does.
Asking
Type into the composer and press Enter. Shift and Enter together give you a new line.
Three shortcuts sit in the composer, and each saves the assistant a step it would otherwise have to work out for itself.
Slash commands
Typing / opens a palette of two dozen ready-made questions, grouped under the same six headings: /approvals for what is waiting on you, /claims for the progress claim position, /lookahead for a two-week look-ahead, /rfi for unanswered mail, /drawing to find one, /read to answer from what is printed on a sheet, /diary, /permits, /swms, /defect to raise an observation, and so on.
Picking one fills the composer with an editable question rather than sending it, so you can add "for level 3" before you press Enter.
Referencing a record with @
Typing @ opens a list of records on the project, so a question can point at exactly the one you mean instead of describing it. Documents, whiteboards, contracts, meetings, observations, mail, submittals, incidents, inspections, permits, inductions, safety data sheets, equipment, bookings, procurement packages and programs can all be referenced this way.
Attaching a file
The attach button takes files as well as records. A file is uploaded to the conversation, up to 25 MB each, and shows as a chip above the composer until you send.
Ask mode and Act mode
Beside the send button is a two-state control, and it decides whether the assistant may propose changes at all.
Ask is read only. Every tool that could write anything is withheld, so nothing can be proposed because nothing can be named. Use it when you want an answer and nothing else.
Act allows the assistant to propose an action: raise an observation, send a reply, create a permit, complete a workflow step. It still cannot do any of it on its own.
Approve or reject an action
When the assistant wants to act, the answer stops and an approval card appears in the thread, in amber, saying exactly what it proposes to do. Nothing happens until you press Approve, and Reject discards it.
Two things hold underneath that card. The assistant can only propose something you are allowed to do, because every tool re-checks your own permissions before it runs. And an approval raised in Act mode stays approvable afterwards: switching back to Ask does not strand a card you have not answered yet.
Sixteen of the sixty things the assistant can do are actions in this sense. The rest are reads, which run without asking.
The working timeline
Above each answer is a line saying what the assistant did to produce it: which modules it loaded, which searches it ran, how many rows each came back with. It builds live while the answer streams and collapses to one reopenable line once it lands.
It is worth opening when an answer surprises you. Most surprising answers turn out to be a search that found nothing rather than a judgement you disagree with, and the timeline is where that shows.
Citations
Under an answer are chips naming the records it was built from, each linking straight into its own tool. They carry the reference a person uses, OBS-42 or INC-7 or the mail number, never an internal id.
That is the check worth doing before acting on an answer: open a chip and read the record.
Conversations
The rail on the left lists your conversations, most recent first, and a title is generated from your first question so you are not asked to name it.
Your chats are private to you
Not a permission you can grant, not something an administrator can read: a conversation is keyed to the person who created it, and nobody else on the project, host or guest, can list it or open it. See Privacy and training.
Rename, export or delete a conversation
Each conversation can be renamed, deleted, or exported as a Markdown file carrying the whole transcript, which is the way to put an answer into a document, an email or a project record.
What it costs
A question spends AI credits from your organisation's weekly allowance, and a question that has to read several tools spends more than one that does not. The composer shows what is left once the allowance is half used. See AI credits.
Working inside a host's project spends the host's credits, not your own, wherever you happen to work.
Where else the assistant appears
AI Chat is the main surface, not the only one. The same model sits behind the 3D model chat in the BIM viewer, the workflow builder's describe-a-workflow panel, Describe your report in the report builder, the rich text editor's rewrite assistant, and the extraction that reads a safety data sheet or a method statement when you upload it. All of them spend from the same weekly allowance and all of them are listed separately in its breakdown.
Two things to know before you rely on it
It reports what is in Teralo. If the diary was not written, the assistant will tell you the diary was not written. That is the correct answer and it is not always the one people expect.
Times are read in the project's time zone, which comes from the project's address in settings. A project that has never had one set falls back to UTC, which on an Australian site is the previous afternoon. Worth setting before you ask the assistant to create anything with a time on it.
And the standing one: it can be wrong. Read the citations before you act on an answer, particularly a number.
PermissionsHost only
Can access AI Chat is the whole of it. There is no separate view and create split, because a chat belongs to one person: you either have the tool on a project or you do not.
It is a project permission, so it can be granted on one job and not another, and it is granted by the host organisation on the template you are on. It is also what gates the AI Chat dashboard widget and any connected outside assistant. See How permissions work and Connectors.
Inside the conversation, permission is checked again per tool. Being able to open AI Chat grants nothing beyond being able to ask.