Aiô answers 24 hours, 7 days. When it does not solve it, the conversation becomes a ticket for human support (9 a.m. to 6 p.m., Monday to Friday) — see the Support Center.
Aiô is the assistant that lives in the bottom-right corner of every screen. It answers from the documentation and from the state of your environment. It's not a generic chat: when you ask "why did my pipeline fail?", it opens the run that failed, reads the task that broke and the end of the log, and answers about that failure.
Availability#
- On any screen of the system, via the Aiô button or the ⌘/ shortcut (on Windows, Ctrl+/).
- The assisted answer depends on the AI provider being configured and on the month's AI budget. Without either one, Aiô still delivers the documentation excerpts that match the question — and tells you which of the two was missing.
Permissions#
- Anyone in the workspace can use it.
- Aiô only sees your workspace. Credentials never reach the model: passwords, tokens, and addresses are stripped from the error message before it leaves the server.
What it does well#
| You ask | What it does |
|---|---|
| "Run the ERP pipeline now" | Finds the pipeline by name and proposes the run; one click from you and it fires |
| "Schedule the Online store for 6 a.m." | Builds the schedule and shows it for you to confirm |
| "Why did my pipeline fail?" | Opens the run, reads the task that broke and the end of the log, and answers about that failure |
| "Are Tuesday's sales missing?" | Counts the table's rows per day and says which days are empty — before claiming data is missing |
| "Recover days 10 to 12" | Shows what will be redone in that period and what stays untouched, and waits for you to confirm |
| "Create a customers Silver" | Reads the real columns of the Bronze table and proposes the Silver table |
| "How do I add clicks to the states mart?" | Reads the columns the table has today and the recipe that builds it; tells you whether it's a calculated column in the dashboard or new data in the table, and what's missing (sometimes just running the pipeline again) |
| "Why doesn't my map show data?" | Reads the fields of each visual in the dashboard (Axis/Location, Values, tooltip) and points out what's missing |
| "How do I build a sales-by-month dashboard?" | The path anchored in your tables and the dashboards you already have, by name |
| "What is incremental load?" | A short explanation and what it's for in practice |
It does it, but not on its own
When what you ask can be done by the system, Aiô does not teach you the path: it builds the action and shows a card with Confirm and Not now. Nothing happens before your click. The card's title and detail come from the system, not from the assistant's text — what's written is exactly what will happen.
It knows where you are on the screen
Not just "which screen": when the screen can tell, it receives the exact point.
- The wizard step. In New pipeline, it knows you are on step 3 of 6 — and answers that step, instead of teaching the whole setup from the start.
- The required field that is missing. "Why can't I move on?" usually has a one-line answer, and it gives it by the field's name.
- The reason it is blocked, when the screen itself works it out ("Choose at least one table").
- The error in front of you. If there is a visible error message, that is what it answers about — without asking you to retype it.
- The open resource. On
/dashboards/…or/dags/…it knows which dashboard or which run you have on screen, and doesn't ask "which one?". - The real button names of that screen, so it cites the label you see instead of describing the path from the outside.
What it does
not read from the screen Nothing you type. Field values, passwords, tokens, table cell content and SQL never leave your browser: what travels is label and state ("the Watermark field is empty"), never content. Any identifier that shows up in an error message (national ID, e-mail, phone, address) is removed before the message leaves, and the symptom — "timeout while connecting" — is preserved. A screen that cannot tell where you are does not guess: it asks.
The actions it already knows how to propose:
| Action | How it appears on the card |
|---|---|
| Run a pipeline | "Run the pipeline “…” now" |
| Rerun only what failed | "Rerun only what failed in “…”" |
| Recover a period (days) | "Recover … in “…”" — with a notice of whether history outside the period is preserved or not |
| Set the schedule | "Schedule “…”: … at …" |
| Turn the schedule on or pause it | "Turn on the schedule for “…”" / "Pause the schedule for “…”" |
| Create a Silver table | "Create the Silver table “…” from …" |
| Turn on micro-batch (refresh every few minutes) | "Turn on micro-batch in “…”: every … minutes" |
| Run a published flow | "Run the flow “…” now" |
| Start or stop a Spark cluster | "Start the cluster “…”" / "Stop the cluster “…”" |
| Create a pipeline from a connection you already have | "Create the pipeline “…” from the connection “…”" |
After Confirm, the card counts the time while it runs — you can keep chatting, because the action runs on the server — and shows a link to follow it, such as Follow the run. If there's an error, the button becomes Try again. On Not now, it replies "OK, I didn't do anything."
What it doesn't confirm for you
Recovering a period in a table that only appends rows can duplicate what already exists. In that case the card says "Can't recover … from here" and explains the path through the Runs screen, where you decide while looking at the preview.
This isn't bureaucracy. Running a pipeline costs processing and writes to the datalake; turning on a schedule changes what happens overnight, when no one is watching. An assistant that misunderstood the sentence can't be the last one to decide — so the final word is always yours. Every confirmed action is recorded in the Audit log, flagged as proposed by Aiô and confirmed on screen.
Screenshot and audio
You don't need to put what you're seeing into words:
- Screenshot — click the image icon (Attach a screenshot) or paste with
Ctrl+V/⌘V. It reads the screen and any error message on it. The image is shrunk in your browser before uploading. Up to 3 attachments per question. - Audio — click the microphone (Record audio) and speak, up to 90 seconds; Stop recording ends it. It transcribes and shows what it heard ("I heard: …"), so you can check it understood correctly. If you only send the audio, the transcript becomes the question.
Screenshots and audio are kept in the support case: whoever looks at the ticket later sees exactly what you saw.
Three safeguards it respects, worth knowing:
- It doesn't claim a feature doesn't exist. If it didn't find it, it says it didn't find it.
- It doesn't make up menu paths. When the answer mentions a screen that doesn't exist in the real menu, the panel shows a notice asking you to check the name using search (⌘K) before moving on.
- It doesn't make up numbers, deadlines, prices, or limits.
How to use it#
Open it on any screen
The Aiô button in the bottom-right corner, or ⌘/ (Ctrl+/ on Windows).
Ask for what you need
Write, speak, or paste a screenshot. It can be a question ("why did it fail?") or a request ("run the ERP pipeline"). While it investigates, you'll see "Investigating: runs, tables, and the documentation…".
Confirm, when it proposes something
If the request can be done by the system, a card appears with what it's going to do. Read it and click Confirm — or Not now.
Adjust the window, if needed
Enlarge window and Shrink window change the size; the header drags the window around the screen. Minimize keeps the conversation — useful when it covers exactly the field you asked about.
End and rate
Click End conversation, at the bottom of the window. Give a rating, check what best describes it, and click Close.
At the end you should see: the message "Conversation closed. Thank you!" and the link View the conversation, which opens the conversation in Support.
Every conversation becomes a support case#
Each conversation is recorded in Support (avatar menu), with the question, the answer, the attachments, the originating screen, the time of each message, and your rating. What Aiô couldn't solve is the most useful material there is about where the product or the documentation is falling short.
- Esc and Minimize keep the conversation. The X (Close window) doesn't end it either: the support case stays open, and Aiô remembers what was already said.
- If an answer arrives while the window is closed, the Aiô button gets a new answer badge.
- To really end it, use End conversation or the header button End and start a new conversation.
- The View my conversations and tickets link, at the bottom of the window, opens the Support screen.
The rating, and when it becomes a ticket
The End the conversation screen asks How did it go? (rating from 1 to 5) and What best describes it, with the tags "Solved my problem", "Clear explanation", "Missing detail", "Not what I asked about", "Step-by-step didn't work", and "Took too long". There's a free-text field ("Want to add details? (optional)").
Check Didn't solve my problem. and the button becomes Close and open a ticket: the entire conversation goes to the team as a ticket. Changed your mind? Back to the conversation.
Limits and costs#
- The assisted answer consumes the workspace's AI budget (see Usage and quotas). Once the budget runs out, Aiô keeps delivering the documentation.
- It reads the state of your environment (runs, failures, end of log, row counts per day, table names, columns from any layer, the recipe of the Silver/Gold tables, the fields of visuals, schedules) — not the content of your data row by row. From the screen, it reads the point you are at (step, empty field, visible error) — never what you typed.
- Schedule: say the time as on the screen, in Brasília time ("at 7 a.m.").
- The actions it proposes are only the ones in the table above, one at a time. It does not create connections, doesn't touch credentials, and doesn't delete anything.
- New ingestion: if you already have the connection saved, it reads the tables it offers and proposes the pipeline — name, tables, schedule, and whether the first load runs now. If the connection doesn't exist yet, it shows the step by step for Connections › New connection, knowing which step of the wizard you are on. A connection failing its test does not become a pipeline: it says so and asks you to reconnect.
- It never asks for a password, token or key in the conversation — and does not offer "send it to me and I'll set it up". Credentials are typed on the screen, in the encrypted form: what you write in the chat is stored in the support case and, if it becomes a ticket, other people read it.
- Creating a Silver table through Aiô copies the chosen columns from Bronze, with no sum, grouping or join. For a Silver that aggregates (for example, "IPCA by quarter"), it hands you the SQL ready to paste into the Silver's Advanced mode.
- Rerun only what failed applies to a task that failed. To redo a task that succeeded, use Runs › Recover… › New task attempt: Aiô shows the path instead of proposing the whole run.
The Aiô window#
- Size: drag any edge or corner of the window to make it taller, shorter, wider, or narrower — the edge lights up when you hover over it. The size is saved for next time.
- Right-click inside the conversation: Copy (the selected text) and Paste into message.
If something goes wrong#
| Symptom | Likely cause | What to do |
|---|---|---|
| "The assisted answer is unavailable in this environment" | AI provider not configured | Talk to the workspace owner; the documentation still appears |
| "This month's AI budget has run out" | AI cap reached | Adjust the limit or wait for the month to turn over |
| "I couldn't reach the server right now. Try again in a moment." | The connection dropped midway | Send it again; the conversation is still saved |
| "I couldn't access the microphone. Check the permission." | The browser blocked the microphone | Allow the microphone for the site and record again |
| "I couldn't confirm it in the menu" notice | The answer mentioned a screen name the menu doesn't have | Check the name using search (⌘K) before following the step |
Next task#
Last reviewed: 2026-10-08.