If you don’t see this in your navigation, your administrator can switch the module on under Administration → Modules, or it may not be included in your plan.
Before anyone can use it
Copilot is off until somebody in your organisation agrees to switch it on, and it runs on your organisation’s own API key with the model provider your administrator chooses. Both have to be true: consent with no key has nothing to send with, and a key with no consent sends nothing. Both are set on Administration → AI, together with which capabilities are allowed and how many questions the organisation may ask in a day. While it is off, this screen says “The copilot is off. Nothing is sent anywhere until somebody here agrees to it — see the Consent tab”, and every Copilot panel elsewhere says it is switched off instead of offering a box to type into.Where you use it
What you can ask for
Your administrator chooses which of these are allowed. Anything not allowed is refused by the server, not merely hidden.
You never choose a capability before asking. There is one box: you type the question, and which
capability answers is worked out from the sentence. While that is happening the conversation’s
capability reads being worked out; afterwards it names the one that answered.
Who can do what
Asking and accepting are separate on purpose: sending a record out to be read and writing to a
record are different risks. See Roles.
Asking
On this screen, while Copilot is on and you may ask, Ask about the whole site offers a text box and Ask. On a record, the box is Ask about this record. Above the conversations the header counts questions left today — the organisation’s daily ceiling, shared by everybody. When it reaches nought the box says “No questions left today. The count starts again tomorrow.” The answer comes back through a queue, so there is a pause: the conversation says “Asked. The answer comes back through the queue, so this can take a moment.” If nothing has arrived after a minute and a half it says so, and suggests your administrator checks the service that sends questions out; the answer still appears when it arrives. Inside a conversation, Ask a follow-up. continues it. A record can carry several conversations: asking from the record’s own box starts a new one beside the earlier ones, which are listed collapsed to their last line so that an earlier thread can be reopened (Open) rather than re-asked.What each turn shows
Charts in an answer
A question about numbers — “how many emergency jobs did each site have last quarter” — is often answered with a chart, shown under Charts above any suggestions. The chart is not a picture of the asker’s figures: it is a question composed from your sentence and run again, with your access, each time the conversation is opened. A technician limited to one site and their supervisor can open the same conversation and see different charts, and both are right. Because you never saw the filter the model chose, the chart says what it covers before you read it:
A question asking which records — “which pumps at the north site are overdue” — is answered with a
list instead of a chart: the matching records in a table, run again with your access each time, with
“The first … of …” beneath it when there are more than it shows. A list cannot be kept on a board.
Where a value in your question matched nothing but the rest of the narrowing still found rows, the
answer is drawn and the value that matched nothing is named under it — “Nothing in field is called
“X”.” — so a figure lower than you expected comes with its reason.
Under the chart, a line names what it counted (the source, the grouping and any split). If you may
create dashboards, Choose a board… and Keep this add it to one of your boards as an ordinary
widget, after which it reads Kept on a board. See Dashboards.
Suggestions
Copilot proposes; it does not write. A suggestion — a failure mode, a cause, a priority — sits under Suggested until somebody decides on it, and nothing reaches a record on its own. Each suggestion shows the field, the proposed value, the reasoning, Rests on: the evidence it used, and the confidence the model gave itself, presented as the model’s own figure rather than as a measurement.
Decided suggestions move to Decided on the conversation, with the outcome and any note.
The Copilot screen
The header counts conversations, suggestions waiting and questions left today (a dash while Copilot is off).Is it any good
An accept rate per capability, over suggestions somebody actually accepted or rejected: Capability, Offered, Accepted, Rejected, Written on (accepted and applied to the record) and Accept rate. A capability whose accept rate is under about one in three is highlighted, because its suggestions are mostly costing attention to be wrong; your administrator can switch that one off on Administration → AI without giving up the others. There is no single headline figure, on purpose. The capabilities fail in different ways, and an average across them is true of no part of the product. Until anything has been decided the tab says “Nothing decided yet”.Consent
Filling a form from a photo, a document or your voice
Two buttons on create forms — an asset, for instance — start a record from something other than a keyboard. They need Copilot switched on with Fill forms allowed; without it, they are not there. Neither appears when you edit a record: they exist to start a new one, not to rewrite one with a history.Fill from a document
Fill from a document (“Read a new … off a photograph or a document”) takes a photograph — a nameplate, say — or an uploaded PDF or image, and fills the matching fields in the unsaved form. While it works the button reads Reading…. Each field it filled carries a small marker until you change it, so you can tell what was read from what you typed. A picker field is only ever filled with one of your own configured values, never an invented one. A picker that depends on another field — a state that depends on a country — is never filled, because filling one half of a pair risks a mismatch nobody notices. A filled picker may show Loading… briefly while its options catch up. Nothing is saved until you press Create. The document itself is read and discarded; once the record exists you can attach the file to it in the ordinary way.Say it
Say it (“Describe a new … out loud”) records you describing the item — “it’s a KSB pump, model…” — and runs the transcript through the same filling. The button reads Listening — stop while it records. What was heard shows the transcript above the fields, so you can check what was understood rather than trusting it. Speech needs an OpenAI key, set separately on Administration → AI. If your organisation uses ChatGPT it is the same key; on Claude it is a second key, because Anthropic has no speech-to-text service. Without one, Say it does not appear. The recording is never kept — only the transcript, on the record of the question.Related
Administration → AI
Consent, the provider and key, allowed capabilities and the daily ceiling.
Dashboards
Where a chart from Copilot can be kept.