What the AI can read
The short answer is: exactly what you can, on the project you are in, and nothing else. This article is the long answer, because "what can it see" is the first question anybody sensible asks about an assistant with access to a live project.
It runs as you
The assistant has no account and no access of its own. Every lookup it makes runs a permission check against your template on this project, at the moment it runs, and the check is the same one the tool's own screen runs.
So the assistant is not a way around a permission. If Contracts is not in your sidebar, asking about contracts returns a refusal, which the assistant then reads back to you in plain words rather than as an error. Two people asking the same question on the same project can legitimately get different answers, and that is the system working.
The same holds for anything it does. Sixteen of its abilities create or send something, and each one checks the matching create permission before it runs, after you have approved it. See AI Chat.
One project, and only that project
A conversation is bound to a project when it is created. Every lookup is scoped to that project, and a record from another project cannot be reached even by naming it, even if you are on that project too.
Records that belong to your organisation rather than to a project are outside its reach for the same reason: they have no project, so they fail the scope test. See Organisation level and project level.
The areas it can reach
Twenty-one areas, loaded on demand as a question needs them:
| Area | What it can do there |
|---|---|
| Search, read in full, send and reply | |
| Documents | Search the register and the sheets, read pages |
| Contracts | Search, full detail with claims and variations, budget summary |
| Procurement | Package status and detail |
| Inspections | Search, module status, start one |
| Observations | Search, raise one |
| Incidents | Search incidents and injuries, report one |
| Submittals | Search, full detail, create a draft |
| Permits | Search, create |
| Method statements | Search, create |
| Safety data sheets | Search, register one |
| Site diary | Search, create an entry |
| Site messages | Search, create |
| Meetings | Search, full detail with agenda, create |
| Bookings | Search, detail, list resources, check availability, create |
| Equipment | Search plant inducted on the project, register |
| Inductions | Search workers, worker detail, site attendance, licence status |
| Program | Search, detail, look-ahead, task search |
| Photos | Search by description, tags, album or date |
| Whiteboards | Search by name |
| Workflows | What is waiting on you, run detail, act on a step assigned to you |
Alongside those, three things it can always do without loading anything: a project summary, the list of what is waiting on you, and a directory search.
On workflows it can only act where you can. Approving through the assistant goes through the same engine as the button on the record, which means the same check: you must be a participant on that step and hold whatever the module requires. It cannot approve on your behalf somewhere you are not the approver. See Approving and rejecting.
What it cannot reach: Passive Fire, Timesheets and organisation records
Worth stating plainly, because a confident "I could not find any" reads the same whether the tool is empty or absent.
Two project tools have no coverage at all: Passive Fire and Timesheets. Ask about either and the assistant has nothing to look in.
Nothing at organisation level. Estimating, Tenders, Cashflow, Calendars, the organisation registers for equipment, permits and method statements, organisation documents and whiteboards, and organisation compliance documents are all out of scope.
Reports and dashboards themselves. The assistant cannot open a saved report or read a dashboard. It reads the same underlying records the reports read. See The report builder.
Anything your permissions do not cover, and, if you are a guest, anything belonging to another guest organisation. That second one is not a permission at all. Another subcontractor's submittals are not in the data the assistant loads, so no template and no phrasing produces them. See Hosts and guests.
Other people's conversations. Chats are private to whoever created them.
Documents are the special case
Everything else is a database lookup. Documents are the one place the assistant reads content rather than fields, and it is worth knowing how, because it decides which questions work.
The register, then the sheets
A document search matches the title, the number and the notes, and the text printed on the pages. A result that matched on page text comes back with the page numbers and a snippet from each, so the answer can say "it is on sheet 12" rather than "it is in that drawing somewhere".
Reading a page
Given a document and a page number, the assistant can read what is actually on that page and answer from it: a schedule row, a dimension, a fire rating, a note. Three pages per call, and it will call again for more, up to a practical limit of about eighteen pages in one question.
This is the one lookup that is not free. See AI credits.
A scanned drawing has no text layer to read
Page text comes from the text layer of a PDF, extracted once when the document is uploaded. Three consequences:
- A scanned or photographed drawing has no text layer, so it is searchable by its register fields and nothing else. There is no character recognition step.
- Only PDFs are indexed. A Word document, a spreadsheet or a DWG is found by title and number.
- A PDF over 100 MB is skipped, and one over 25 MB cannot be read page by page.
It matches words, not meaning
Content search is a literal match rather than a semantic one, so it finds FRL 120 and CT-9 reliably and does not find "the fire rating of that wall" by inference. Ask with the term as it is written on the drawing.
It gives you references, not ids
Everything the assistant names, it names the way the register does: OBS-42, INC-7, a mail number, a booking number. Internal identifiers never appear in an answer, which is deliberate: a string you cannot look up is worse than no answer.
The same rules apply to an outside assistant
Connecting Claude or ChatGPT to Teralo gives it this same set and these same checks, resolved per user, with a read-only option on top. Nothing about connecting widens what can be reached. See Connectors.