Chat — AI assistant

The assistant is your home screen

The AI assistant chat is the first screen you see after signing in. Ask it anything about your active project — issues, assets, documents, the planner, RFIs — in plain language, and it looks the data up for you and answers with sources.

You can return to chat at any time from the drawer (hamburger icon or swipe right from the left edge), and your past conversations are always available in the drawer's chat history.

The chat header — project chip, new chat, and badge

The chat header has four controls:

• Left: the menu chip opens the drawer.
• Centre: the project chip shows your active project's name. Tap it to open the project picker and switch projects, or to join one with an invite code. It reads "My Workspace" while you are in your personal workspace.
• Right: the bell opens your notifications; it carries a badge with your unseen notification count.
• Far right: the compose icon starts a new chat.

Until you join your first project, an onboarding banner guides you to join one.

How to choose which assistant answers

Cx Assistant has several assistants, each focused on one kind of work. You pick one when you start a chat.

1. Start a new chat with the compose icon in the header.
2. In the toolbar under the message box, tap the chip that shows the current assistant's name (next to the + button).
3. The "Choose assistant" sheet opens. Each row shows the assistant's name, a short description, and a "Default" or "Beta" chip where it applies.
4. Tap the ⓘ icon on a row ("About this assistant") to see its card: "What it can do", "What it can't do", "Try asking" examples, and "Available on". Tap back to return to the list.
5. Tap a row to select it, then send your first message.

The assistant is fixed for the whole conversation. The chip only shows on a new, empty chat — to switch assistants, start a new chat. If you see no chip, only one assistant is available to you there. A warning icon next to a name means that assistant can make changes that cannot be undone, such as removing members or locations.

The assistants you can choose from

Inside a project, the "Choose assistant" sheet lists:

• Commissioning Assistant (Default) — works with this project's issues, assets, documents and scripts, draws and edits 3D layouts, reads CAD drawings and administers the project (sets and statuses, members, locations, workflows and settings) — each of those as far as your role allows.
• General Assistant — engineering knowledge, web search and your own files; no project workflows.

In My Workspace the sheet lists the My Workspace Assistant (Default) — your personal files, engineering knowledge and web search, no project data — and the General Assistant.

Before you join a project it lists Getting Started (Default), which creates your first project and guides you through the platform, and the General Assistant.

Building dashboards in chat is only available in the web app. Every assistant works with your own permissions: the Commissioning Assistant only draws layouts or administers the project for people whose role grants it (the layouts and project administration permissions, granted to project admins by default), and every change it makes is checked like a change you would make yourself.

Asking a question — streaming replies and stopping

Type your question in the composer and send it. The reply streams in live, word by word, with a status chip and typing indicator showing what the assistant is doing.

While the assistant is answering, the Send button becomes a Stop button — tap it to stop generation immediately if you already have what you need or want to rephrase.

Tool chips, thoughts, and sources — seeing how the answer was built

While answering, the assistant may use tools to look things up in your project. Each tool call appears as a chip with a live label describing what it's doing, and you can see its result.

You can also expand the thoughts panel to read the assistant's reasoning, and a sources list under the answer shows where the information came from, with inline citation chips in the text itself.

Opening citations — documents, issues, and assets in place

Answers cite their sources. Tap a citation chip to open the source without leaving the chat:

• Document citations open the document viewer.
• Issue citations open the issue's detail sheet right there.
• Asset citations open the asset's detail sheet right there.

This lets you verify any claim the assistant makes in one tap.

Approvals: control what the assistant may change

You decide how much the assistant can do on its own. The approval mode chip is in the toolbar under the message box, with a shield icon and the current mode's name. Tap it to choose:

• Auto — the assistant runs all actions without asking. This is the default.
• Ask for critical — only critical actions (permanent deletions, member removal) ask first.
• Ask for edits — every action that changes project data asks first.

The mode is remembered per chat. Reading data never requires approval — only changes do.

A few actions ask in every mode, Auto included: creating a project, adding a company to the company directory, deleting a planner activity, and anything a connected service sends on your behalf, such as an email or a meeting invitation. For those, the card shows the full details — recipients, subject and message — so you can read exactly what will be sent. Connected services are set up in the web app, in Settings → Connections.

Deleting a planner activity: the assistant can remove an activity from the planner when you ask, in the Commissioning Assistant. It always asks first — the card names the activity and warns that everything under it goes too — and it needs the planner.delete permission, like deleting it yourself.

How to answer an approval card

When the assistant needs your go-ahead, it pauses and an "Approval required" card appears in the conversation. You answer it right on the phone. It lists each action the assistant wants to run, with a checkbox per row.

1. Read each row. The title says what will happen; the lines below show the details. Long text, such as an email body, folds — tap "Show more" to read all of it.
2. Untick any action you do not want.
3. Tap Approve to run the ticked actions (the button reads "Approve 2" when only some are ticked), or Decline all to run none.
4. Optional: before approving, tick "Don't ask again for … in this chat" so that kind of action stops asking for the rest of the conversation. Critical actions never offer this.

Declined actions show as "Declined" and the assistant carries on without them. If you send a new message instead of answering the card, the waiting actions are dropped and do not run.

If you are not looking at the chat when the assistant stops for approval, you get a notification: "The assistant is waiting for your approval". Tap it to open the chat.

The Tasks list: follow a long job step by step

For a job with several steps, the assistant writes itself a task list and shows it to you. A card titled "Tasks" appears just above the message box.

• The header shows progress as done/total, for example 2/5.
• A pulsing dot marks the task the assistant is working on right now (In progress).
• An empty circle is a task still to do (Pending).
• A green tick is a finished task (Completed).

Tap the header arrow to collapse or expand the list. The card disappears by itself once every task is completed. The list is the assistant's own plan — you do not edit it; to change course, send a new message.

Research agents: helpers that read in the background

When a question needs a lot of reading — many documents, long reports, several topics at once — the assistant can hand parts of that reading to research agents and write its answer from their reports. A research agent only reads; it never changes anything. (This needs an up-to-date version of the app.)

The agents work in the background: the assistant keeps working on the rest of your request while they read, and it can start several for one message. If it has nothing left to do but wait, it may send a short first message and then continue the answer once the reports are in.

Each agent is one compact row in the conversation:

• The row reads "Agent (Researcher)" and a short description of the job, with its state on the line below.
• While the agent works in the background, its icon pulses and the row says "background task".
• When the assistant is waiting for the agent, the row says "running" with the time so far.
• When the agent ends, the row shows the steps it took, the time and the tokens it used, for example "5 steps · 41s · 2.4k tokens". A warning icon with Stopped, Timed out, Ran out of steps or Failed means it could not finish.

Tap a row to open its card underneath, with the steps the agent took. On the card, tap "Open thread" to open the "Research thread" sheet with the Briefing the assistant gave, the Steps taken and the final Report.

Limits: the assistant can start up to 9 research agents for one message, and three of them work at a time — the others wait their turn. Once the limit is reached, the assistant does the rest of the reading itself. Stopping the answer also stops the agents that are still running.

You do not need to do anything — the rows are there so you can see where the answer came from.

Documents the assistant writes: Word, PDF and exact-layout sheets

Ask the assistant to write a document and it saves a real file in your files, with a preview and versions like any other document.

There are two kinds:

• Reports, minutes, memos and other flowing text. The assistant writes the text and saves it as a Word document (the default, editable), a PDF or a Markdown file. Say which format you want.
• Exact-layout sheets, for when the layout itself is the point: a client's form or checklist reproduced from its original, a one-page test sheet, a certificate or cover with a logo in a corner, tables with merged cells and shaded bands. The assistant lays the page out itself (it writes it as an HTML page) and saves it as a PDF. Nothing is added to such a page — no cover, no logo, no footer — unless you ask for it.

Getting a good exact-layout sheet:

• Attach the original form, or name the document to copy. The page is A4 unless you say otherwise.
• Logos and pictures must be images in your files or attached to your message — the assistant cannot take a picture from a web address. It can cut a logo or a figure out of a page of one of your documents, save it as an image, and place it on the sheet.
• The assistant looks at the finished pages and the page count itself. Ask for corrections in plain words ("move the logo to the top right", "keep it on one page"): each correction saves a new version of the same file, not another file.

Where it goes: the folder you name, or your own Chat files when you name none.

Creating a document is a change, so it follows your approval mode, and it needs the documents.create permission in a project. The Commissioning, Project Management and My Workspace assistants can write documents.

Leaving a chat while the assistant works — background work and notifications

You do not have to wait on the chat screen. If you open another chat, go to another screen, or put the phone away while the assistant is working, it keeps going and the finished answer is waiting when you come back. (This needs an up-to-date version of the app.)

You can have the assistant working in up to 3 chats at the same time — start a long task in one chat, then start a new chat and ask something else. Each chat works on one request at a time.

In the drawer's chat history, a dot next to a chat shows its state:

• Pulsing dot — the assistant is working in that chat.
• Green dot — finished, with an answer you have not seen yet.
• Amber dot — waiting for your approval.
• Red dot — stopped with a problem.

Opening the chat clears its dot.

When an answer finishes in a chat you are not looking at, you get a notification: "The assistant finished answering". With the app closed or in the background it arrives as a push notification; tap it to open that chat.

The first time you walk away from a working chat, the app asks "Get notified when it's ready?". Tap "Turn on" and then allow notifications in the system prompt. "Not now" is fine too — the offer is made only once, and you can turn notifications on later from the notification settings in the app.

Long conversations — the context ring and summaries

Every chat has a limit on how much the assistant can keep in its working memory at once. The small ring next to the message box shows how much of that memory the conversation is using; it fills as the conversation grows.

When a conversation gets close to the limit, the assistant summarizes the older part of it and continues from that summary, so you can keep working in the same chat (needs an up-to-date version of the app). This can take a minute or two — the status line says "Summarizing earlier conversation…" while it happens. A thin line saying "Earlier conversation summarized" then appears in the transcript, and the ring drops. Nothing you wrote is removed: every message stays above the line for you to read; only what the assistant keeps in memory changes. Tap "View summary" on the line to read what the assistant continues from.

Summaries keep record ids, decisions, open tasks and the exact next step, but fine detail from early in the conversation may be condensed. For a new topic, a new chat is still the cleanest start.

Sending a message with a weak connection — "Your message was not sent"

If the connection drops for a moment while you write, you can still tap Send. The app holds your message and sends it by itself as soon as the connection is back — you do not need to retype it.

There is one exception. If, while you were offline, you moved to a different chat, the app does not post the held message into the wrong conversation. It removes the message from the screen and shows an alert: "Your message was not sent. The connection dropped while you were writing it and that chat is no longer the one open. Open the chat and send it again."

When you see that alert, open the chat you were writing in from the drawer's chat history and send the message again.

How to create a project from the chat

You can set up a new project by asking for it — also before you have joined any project.

1. Open the chat and ask, for example: "Create a project called Riverside Data Center for Acme".
2. Give the two things the assistant needs: the project name and the client company. The project code is optional — a short one is made from the name if you do not give one. You can also ask which project templates exist and pick one.
3. An "Approval required" card appears with the project name, code and client. Creating a project always asks, whatever your approval mode. Tap Approve.
4. A "Project created" card appears. Tap Open to switch to the new project.

You become the owner of the new project. If the client company is not in the company directory yet, the assistant tells you and asks before adding it. This works with the Getting Started, Commissioning, Full, Project Management, My Workspace and General assistants.

Attaching photos and files to your message

Tap the + button in the composer to attach context to your question:

• Take photo — use the camera.
• Photo library — pick images from your phone.
• Document — pick a file from your phone.
• Project documents (or My Files in My Workspace) — browse the documents already in the app and attach files, or a whole folder, without uploading them again.
• Paste image — shown when you have copied an image.

You can attach images (JPG, PNG, GIF, WEBP, HEIC), PDFs, and Office or text documents (Word, Excel, PowerPoint, TXT, CSV). Limits: up to 10 files per message, 60 MB per file. Videos cannot be attached from the phone — use the web app for those.

Attachments appear as a thumbnail strip above the input with upload progress. If an upload fails you get an error with a retry option, and each chip has a remove control. If a file is refused, a message tells you why. Send your message and the assistant can read the attachments.

Voice dictation — talk instead of typing

When the input is empty, a microphone button appears in the composer. Tap it to dictate:

• The button pulses red while the app is listening.
• Your words appear as a live transcript merged into the input, so you can keep editing by hand.
• Dictation follows the app's language, set in Profile.

Tap again to stop listening, review the text, and send.

Thinking mode — deeper answers for harder questions

The composer has a thinking-mode selector with four settings:

• Non-thinking — fast, direct answers for everyday questions.
• Thinking · Low — brief reasoning; quick answers with a sanity check. This is the default and the recommended setting.
• Thinking · Medium — balanced reasoning for complex tasks.
• Thinking · High — the deepest reasoning; thorough analysis for hard, multi-step problems, at the cost of a slower reply.

The setting applies per message, so you can raise it for one hard question and drop back down afterwards.

A context-usage donut next to the composer shows how much of the conversation's capacity is used — start a new chat if it fills up.

Mentioning members with @ and assets with #

While typing, two shortcuts bring project context into your message:

• Type @ to get an autocomplete of project members and mention one.
• Type # to get an autocomplete of assets and reference one directly.

Pick from the suggestion list that appears above the input. References give the assistant an exact anchor, so answers about that person or asset are precise.

Copy, feedback, and regenerate

Under each assistant message you'll find:

• Copy — copies the message text (a brief "Copied" confirmation appears).
• Thumbs up / thumbs down — rate the answer to help improve the assistant.
• Regenerate — available on eligible messages; asks the assistant to produce a fresh answer to the same question.

Chat history — finding past conversations

Open the drawer (hamburger icon or swipe right from the left edge) to see your chat history below the module list. The active chat is highlighted; tap any conversation to reopen it, and scroll down to load older ones. The New chat button at the top of the history starts a fresh conversation.

Inside a long conversation, scroll up to load older messages; a floating "Jump to latest" chip returns you to the newest message.

Quick prompts on a new chat

When a conversation is empty, suggestion chips with quick prompts appear above the composer. Tap one to send it as your question — a fast way to see what the assistant can do or to kick off common requests without typing.

The assistant respects your permissions

The assistant works with the same access you have. It answers from the active project shown in the header's project chip, and it can only read and change what your own project role allows — it never bypasses permissions to reach data you couldn't open yourself.

Chat itself needs no special permission; anyone signed in can use it.

Chat not responding — troubleshooting

If the assistant misbehaves:

• Reply stalls or errors — tap Stop if it's still running, then resend, or tap Regenerate under the failed answer. Check your internet connection.
• Answers reference the wrong project — check the project chip in the header and switch to the right project.
• Attachment stuck — use the retry control on the failed chip, or remove and re-attach it. A file is refused when it is over 60 MB, is a type the chat cannot read, or is the 11th file in one message.
• The assistant seems to be waiting and nothing happens — look for an "Approval required" card in the conversation (an amber dot in the chat history marks it). It is waiting for your answer.
• Actions keep asking for approval — the approval mode chip under the message box may be on "Ask for edits". Switch it to Auto or "Ask for critical". A few actions (creating a project, sending an email through a connected service) always ask.
• An alert says "Your message was not sent" — the connection dropped and you had moved to another chat. Open the chat you were writing in and send it again.
• You cannot find the assistant chip — it only shows on a new, empty chat, and only when more than one assistant is available to you. Start a new chat to switch.
• No notification when an answer is ready — turn notifications on in the app's notification settings, and check that your phone's system settings allow notifications for Cx Assistant.
• Dictation not working — make sure the app has microphone permission in your phone's system settings, and note the mic button only appears while the input is empty.
• Long conversation getting slow or the context donut is full — start a new chat with the compose icon.
• Still stuck — close and reopen the app, and contact support from Profile → Support if the problem persists.