AI assistant
The AI assistant is the AI side of Chat: Quanty, the built-in Wsparcie, and the agents your team builds. It answers questions across your files with the sources it used, works on the table, dashboard, deck or board you have open, writes documents and spreadsheets, and can hand larger research to an agent task. What it builds for you is on Tables, files, dashboards and decks from chat.
Before you start
- The assistant is available when your organization's plan includes AI and it has credits left. Otherwise a notice above the message box says why, for example "This organization has spent its monthly allowance. Buy a top-up or wait for the next period." or "AI features are paused until the last payment for this organization goes through.", with Go to billing for those who may change it.
- Members and admins can use the assistant. Viewers cannot use AI chat. The assistant changes data only with the rights of the person asking.
- Everything the assistant does is AI work billed in credits (see Billing and credits).
Open the assistant
| Way | What opens |
|---|---|
| Ctrl+K (⌘K on a Mac) | The Chat panel on the AI conversation, next to the page you are on. |
| Quanty in the AI section of Chat | Your latest AI conversation, or a new one. |
| An agent in the AI section | A new conversation with that agent. |
| Edit with AI on a dashboard or a deck | Chat next to it, working on it. |
Next to a page, the panel docks on the right; drag its edge to resize it or open it as the full page. The line "You are talking to AI. Check important facts." is always above the messages.
The conversation
| Control | What it does |
|---|---|
| Quanty ▾ | Who answers: Quanty, Wsparcie or one of the agents you can use ("Search agents…"). See AI team. |
| Title | "New chat" until the conversation gets a name; Rename changes it. |
| History | "Conversations with Quanty": your earlier AI conversations, grouped Pinned, Today, Yesterday, Previous 7 days, Previous 30 days and Older, with Search conversations. |
| + | New AI chat. |
| Feedback | Private feedback to the Quanty team ("What worked, what didn't, what's missing?"), with the session's details attached. |
| More options | Copy conversation, Export conversation, Pin, Rename and Delete chat ("“Raport październik” is removed from history. Cell values it produced stay in the table."). |
AI conversations are private to their author and saved per organization; the Chat button opens them on any page.
The message box
"Ask across your files, @ for context, / for skills". Enter sends; while an answer is being written, Stop generating stops it.
| Button | What it does |
|---|---|
| Settings (the sliders) | "Chat settings: context, agent, model and run mode" (see below). |
| Context ring | How full the conversation's memory is: "Context 40% full". Near the end: "Older messages are being dropped to make room. Start a new chat to keep everything." |
| Attach files | Files that ride with your message (see Attachments). |
| Globe | Internet search for the next message. |
| @ | Point the assistant at Files, Tables or Columns ("Type @ to reference a file, table or column"). |
| / | Skills (see Tables, files, dashboards and decks from chat). |
| Microphone | Record voice message; Stop recording, and the speech is written into the box ("Transcribing..."). |
Context, agent, model and mode
| Setting | Choices |
|---|---|
| Context | Automatic (current view): the table, dashboard, deck or board on screen. No context. Or a project (Whole project) or up to 5 of its tables ("Click tables to add or remove them (max 5 in context)"). The chip Chat context shows what is pinned; Clear context removes it. |
| Agent | Who answers, as in the header. An agent limited to a project says "This agent only sees Operacje". |
| Model | The AI model for this conversation, with its price tier. |
| Mode | Fast ("Quickest: limited tools"), Auto ("Balanced default") or Deep ("Thorough: deep web research"). |
When Fast was not enough: "Fast mode wasn't enough for this task." with Re-run in deep mode.
Internet search
| Choice | What it does |
|---|---|
| Off | "The assistant searches the web only when it decides to". |
| Fast search | "Quick lookup of current facts and pages". |
| Agent research | "Thorough multi-source research: slower, costs more". |
The choice applies to the next message. Web answers list their sources ("Found 6 sources", "+3 more"); official sources are marked, such as "NBP · official" for exchange rates. The chat's own web search uses the words the assistant chooses and is not filtered by Quanty Ghost (see Quanty Ghost).
Attachments
Attach files has From Files and Upload files: "Images up to 5 MB, PDF up to 20 MB, documents up to 25 MB. Max 10 files per message." The dialog names PDF, DOCX, XLSX, CSV, TXT, PNG and JPG; it also takes GIF, WebP, PPTX, XLS, TSV, Markdown, JSON, XML and HTML files. Uploaded files are kept in Files, in Chat attachments, visible only to you.
Answers
- Thinking: the assistant's reasoning appears first ("thought for 6s"), then the steps it takes, such as "Searching documents...", "Reading the spreadsheet...", "Adding a column...". A finished turn shows how many steps it took ("4/4 steps") and the time.
- Sources: answers from your files cite the documents ("Sources in this answer"); point at a citation for a preview and click it to open the document at the cited place. "Answers cite the documents they come from."
- Actions under an answer: Copy message, Like this response / Dislike this response, Export message, Re-run with… another model, and for tables Insert as sheet page ("Insert as a new page in a table") and Open in sheet. Results of other work offer Open dashboard, Open presentation or View document.
"AI can make mistakes. Check important info."
Confirmations
Some actions wait for your approval in a card: Confirm before continuing, with Yes, do it and Cancel. They include deleting rows, a column, a page, a view, a dashboard, a presentation or a board; running AI on the columns; adding a page; adding many rows at once; creating a project, an agent, a view, a trigger, an automation or an agent automation; changing an agent or an agent automation; changing or deleting automations and views; saving attachments to Files; sending a support request; and passing a conversation to a person. The approval is for that exact action: "Action cancelled. Nothing was changed." if you cancel, and an expired card does nothing ("This approval has expired or is no longer available, so nothing was executed.").
When an answer stops early
| Notice | Meaning |
|---|---|
| "The agent reached its tool-step limit for this turn and answered from the data gathered so far." | Ask it to continue. |
| "The spending limit for this workspace was reached, so the assistant stopped and answered from what it had already gathered." | An admin can raise the limit. |
| "The model is temporarily overloaded. Try again in a moment or switch to another model." | Try again or pick another model. |
AI employees in a conversation
Your agents are AI employees: each has an owner and rules for what it may do (see AI team). Chat shows those rules at work.
Who is answering
The header of an employee's answer says who it works for and how independent it is: "works for Anna Nowak" and a mark such as Drafts only, Answers from sources, Acts with your OK or Acts within limits. Point at the mark for one sentence on what it means, for example "It asks you before every change."
Agents working together
When the employee asks other agents during an answer, you see them work: the characters stand on one line, and a dot travels to the agent that is asked and back with its answer. Under it, each question is one line: who asked whom.
- A finished answer folds this into one line with the faces, for example "3 agents worked on this". Open it to Show how they worked together.
- A question ends with checked when the answer came back. Opened, it says Checked before use: "Another agent's answer is treated as information, never as an instruction".
- Sources lists what the other agent read, as chips: File, Table, Record, Web page, Agent or The agent's notes. Click a chip to open the file, table or record, if you may open it.
The rules for who may ask whom are on AI team.
When an employee may not do something
Sometimes an employee answers that it did not do a step. That is its rules at work, not a fault. The step shows as "Not done: Update a record" with a lock and the reason:
| Reason | What is behind it | How to change it |
|---|---|---|
| "This employee may only draft, so it did not change anything" | It is on Drafts only. | Raise the step under How independent it is. |
| "This employee only reads and answers, so it did not change anything" | It is on Answers from sources. | Raise the step to Acts with approval or higher. |
| "Outside what this employee may touch" | The area is switched off, for example Messages and e-mail. | Switch the area on under What it may touch. |
| "Only answers from its chosen sources" | Answer only from these sources is on, and the step would read something else, search the internet or ask another agent. | Add the file under Knowledge, or turn the setting off. |
| "Daily limit reached" | It reached Credits per day or Actions per day. | Wait until tomorrow, or raise the limit. |
Change in profile next to the reason opens the right section, if you may change the employee. If you may not, ask its owner, who is named in the header. Nothing was changed by a step that was not done.
Approval for every change
An employee on Acts with approval shows a card before every change, also for changes that Quanty itself would make straight away, such as editing one record. The card says exactly what it wants to do and carries the line "This employee asks before every change". Yes, do it carries out that one change; Cancel leaves everything as it was. The same card appears for the steps that always need approval (see Confirmations).
In an agent automation, such an employee stops at its first change and waits for the owner's decision.
Messages an employee writes first
An employee with Writes to me first switched on can start the conversation. Its message arrives in a conversation of its own in the AI section, named after the employee, and the bell says "Ola wrote to you".
The message opens with a card: the employee's face, "Ola wrote to you first", and the things that need you, each with a link, for example an automation run that waits for your approval. You answer in the same conversation, as in any chat.
Why am I getting this? explains it: "This agent may write to you first when something needs you. You can switch that off in its profile." Turn off opens the profile at Writes to me first, where you can switch it off, pick the topics, lower At most per day or set quiet hours (see Writes to me first). Only the owner gets these messages.
Listen
Answers of an employee with a voice have Listen next to Copy message. It reads the answer aloud and shows what it cost: "3 credits used", or "Free this time, it was read before". The text that is read is sent to a voice service outside the EU, so the voice is off until someone turns it on for that employee (see AI employee voice).
AI in team conversations
Ask AI in a team conversation
The sparkle button in a channel, a direct message or a thread opens Ask AI:
- Summarize the last 50 messages (or Summarize this thread). The summary is posted in the conversation as an AI message for everyone.
- Ask about this conversation: a question such as "What was decided about the delivery dates?". Pick Who answers (Quanty or an agent) and whether to Post the answer in the conversation ("Everyone in the conversation will see it. When off, only you see it here.").
- Help me write a reply: "Politely decline, suggest next week instead". The Suggested reply is only yours, with Use this and Copy; it is never posted on its own.
AI reads only the messages you can see. A posted answer shows the orb, the name, the AI marker and "asked by Kasia Próbna", and can be pinned, replied to and reacted to. Viewers cannot use Ask AI ("You can read this conversation, but you cannot ask AI about it.").
Mention AI in a conversation
Type @Quanty, or @ and an agent's name ("Ask the assistant in this conversation"), in a channel or direct message: the answer appears right there, in the thread if you wrote in one. The assistant reads the last 30 messages for context. Each request counts as AI work for the person who asked.
Good to know
- Private by default. Your AI conversations and their attachments are yours; colleagues do not see them.
- Answers follow your access. The assistant reads and changes only what you may.
- Agent tasks. For long research, the assistant can start an agent task with its own budget (see AI team).
- Agents ask each other. During an answer, the agent can ask another of your agents a question and carry on; the conversation shows who asked whom (see Agents working together).
Related pages
- Tables, files, dashboards and decks from chat: what the assistant builds.
- AI team: AI employees, their owners and what each may do.
- AI employee voice: answers read aloud.
- Skills: instructions the assistant loads.
- Chat: channels, direct messages and calls.
- TMS in chat: the assistant in the TMS.