What to ask Orion for

And there is one request almost nobody thinks to make: ask it about itself. "How does the shared agenda work?" · "What changes if I declare a group as external work?" — it consults this manual in your language and comes back with the right passage, instead of you having to hunt down the chapter. It works in voice conversation too. And when the manual has no page on the subject, it says so — it does not make the answer up.

When it gets something wrong and you ask why, the rule is not to make up the cause. It says what the checks in that conversation showed (for example, "the 6:06pm read did not return that event") and, when it does not know the reason, it says so: it cannot see its own code or internal records, and the team can look into it. And there is a safety net: when the system notices a technical explanation with no check at all, it reminds it that it cannot see that and asks it to rewrite.

Work and deliverables

"How is my team's dashboard doing?" · "Which deliverables are late?"
"Create a deliverable 'Launch campaign' for the Marketing team."
"Add the task 'review the copy' to the Campaign deliverable." · "Mark that task at 100%."
Recurring deliverable (a standing goal, with no final deadline — monthly sales, NPS, uptime): "log June's result for the Sales deliverable: 87 thousand." Orion keeps the value for each period; this kind of deliverable never "finishes" at 100%.
"Create a team deliverables plan for July." · "Create my work plan for July."
"Log the decision: we prioritized X because Y." That becomes reusable precedent — the more you log the why, the more its recommendations look like yours.

Important — two kinds of "task"

Work task: belongs to a deliverable. "Add task X to deliverable Y."
Personal list (to-do): things from your own life. "Buy milk", "call the accountant".

If you do not mention a deliverable, it takes it as an item on your personal list. When in doubt, it asks.

"Am I the bottleneck?"

Ask "what is waiting on me?" and it lists everything that is stuck on you. There are six categories, and the empty ones do not show up:

work plans waiting for your signature;
cycles waiting for your assessment;
assessment appeals waiting for your decision;
edit requests waiting for your approval;
deliverables plans waiting for your signature;
deliverables with 100% of tasks waiting for you to confirm completion.

When there is nothing, it says that too: "Nothing is stuck waiting on you right now." And the reverse works: "which of these can you do for me?" — it points out the tasks that are a case for AI and offers to run them or draft them (in an organization without a subscription, it doesn't offer).

Without a subscription. In an organization without a subscription, the list's deadlines show as deadline paused — never overdue —, and the answer says they are paused; when whoever had to act has no access, the deadline shows as deadline on hold. And, if you have no subscription, whatever asks for signing, assessing, deciding a review or submitting a report comes with a 🔒 and the reason: Orion does not offer to do it for you. See No subscription.

Search, files and audio

"Search the internet for the average price of X" — it searches and summarizes with sources.
Send audio instead of typing: it transcribes and answers normally (great in traffic). Each plan has a cap on transcription minutes per day and per month, and audio over 20 MB is not transcribed — nor is audio longer than what is still left for the day (or the month): it tells you and suggests a shorter audio or text. When the cap runs out, it says so instead of pretending it did not hear. In the dashboard's Ask Orion, tap 🎙 (it shows when the text box is empty): the audio goes as it does on WhatsApp, and the bubble shows what Orion understood.
Send a PDF, photo or spreadsheet in the chat and ask about it: "summarize this contract", "what does this photo of the whiteboard say?". In the dashboard's Ask Orion, use the 📎: up to 5 files per message. Of documents — PDF, Word and Excel —, it reads up to 50 pages per file and 100 per message: from a larger file, it reads the first 50 and says so; whatever goes past the message's 100 is left out, and it says what it did not read. There is also a limit of 200 pages per day and 1,000 per month (on the free trial, for the length of the trial); past that, it says it did not read the file and why. You can also reply to (quote) a photo, file or voice note already in the conversation and say "review this": it fetches exactly that item, even if other messages arrived afterwards — and it never pretends to have seen what it did not receive: when it cannot recover the item, it says so. This works on WhatsApp connected from the dashboard, your own number or a number dedicated to Orion, and also on the Starting without connecting anything path (the Orion contact saved in your phone) — on that last path, for messages from the last 30 days.
Contract review before you sign or renew. Send the contract and ask "review this contract before I sign" or "can I renew this?". Send the PDF or Word file on WhatsApp, on Telegram or in the dashboard's Ask Orion (the 📎 next to the text box). He replies right away that he is reviewing it and sends the review shortly after, in this same conversation and in the language you asked in: the risks, from most to least serious, each with the clause, the passage from the contract and what to ask to change; what is missing from the contract (for example a limitation of liability or data protection); and the automatic renewal, with how many days ahead you must give notice. Before it reaches you, the review is checked: every passage it quotes has to be in the contract, and no important missing clause can be left out. It is a triage, not legal advice: he never says you can sign, and when there is a serious risk he recommends a lawyer. It only works in your direct conversation with Orion — not in a group. For a very long contract (more than about 25 pages), he asks for the part that matters most, or for the contract in parts. The list of clauses he checks is for contracts in Portuguese; for a contract in another language, the review comes from reading the text alone, and he says so.
Financial analysis of a file. Send the receivables spreadsheet, the sales report or the bank statement (CSV, Excel or a PDF table) and ask "who owes me money?", "how was the month?" or "does what I received match what I was owed?". To reconcile, send both files, together or one after the other, and ask for the reconciliation: he uses the last two you sent in this conversation in the past two hours, and they need a column in common (document, invoice or number). If only one arrived, he asks for the other. He replies right away that he is looking at it and sends the analysis shortly after, in this same conversation and in the language you asked in. The calculations come from code, row by row: how much is overdue and for how long, who owes the most, revenue, margin and month by month, and what appears in only one of the two files. He says what the numbers mean and what to do first. Every number in the answer is checked against the calculation before it reaches you, and anything the file did not allow him to calculate is said. If the answer arrives with a warning that it did not go through an Orion specialist, its numbers were not calculated or checked: ask for the file analysis again. He only reads and proposes: he never charges, pays or books anything, and reminds you to check the entry before charging anyone. If the file has no amount column he can recognize, he asks which one it is instead of guessing. It only works in your direct conversation with Orion, not in a group.
A question with numbers about a spreadsheet. Send the spreadsheet (CSV or Excel) and ask "how much did each salesperson sell?", "what is the average ticket per month?" or "which 5 products sold the most?". The calculation runs over every row, and he answers with the numbers, what they show, what you need to know before trusting them and questions to go further. Numbers written the Brazilian way, like R$ 1.234,56, are read correctly. The answer writes the numbers in the format of the language you asked in and, if the spreadsheet shows the currency (like R$), with the currency and the cents; with no currency in the spreadsheet, the number comes without one. Every number in the answer is checked against the calculation. The calculation runs on its own: when the question asks for a number, it is done over every row before the answer is written — the file's rows are never added up in the head. So the warning that the numbers did not come from the calculation becomes rare; if it still appears, ask for the spreadsheet analysis. When the question is not a calculation, for example summarizing customer comments, he says so and asks for the question in terms of the columns. For a finished spreadsheet in Google or Excel, ask for the spreadsheet. Only in your direct conversation with Orion.
A data file or document: he confirms he got it. When you send a PDF, Word file, spreadsheet or .csv and ask for analysis, the answer takes time — a large dataset takes over a minute. In that case he replies right away saying he received it and naming the file, and the analysis follows in the same conversation. If the confirmation arrives and the analysis does not, he tells you himself that the processing dropped and asks you to send the file again — he does not leave you in silence. When he splits a large piece of work to run in parallel (research on several fronts, a large spreadsheet) and the parts are not ready in time, he tells you he did not finish, says that some steps may already have been done, and asks you to check before asking again — in parts, if it is something large; the result does not arrive later on its own. If you ask for two of these in a row, he does both and delivers each one — the result or the notice — as soon as it is ready, without the second one wiping out the first. He runs up to three at the same time; if you ask for a fourth before any of them finishes, he replies right away that he did not start it and asks you to request it again once one of the others arrives. Each of these jobs splits into at most 8 parts in parallel. Before starting, he checks whether the AI allowance that pays for the request has room for the whole job — your plan's allowance or, in the Ask Orion chat of someone without a channel plan, the organization's AI allowance —, also counting the requests already in progress: if it has room for only part of it, he groups the work into fewer parts or tells you what fits now; if it has no room at all, he does not start and tells you when the allowance renews. These notices arrive on WhatsApp and Telegram. In the dashboard, while the Ask Orion chat is open, it shows by itself the result or the notice that Orion records in the conversation, with "Working in the background…" while it waits; if you close the chat, it checks again when you reopen it — even after 10 minutes —, and if 10 minutes go by without the work finishing it stops waiting and tells you the result will be in your conversation with Orion. Reloading the page starts the chat empty, and it no longer follows the work — ask Orion for the result. In the app, the result of background work does not show up on the screen by itself yet: ask him for it. Photos get no such confirmation, because reading an image comes back in seconds.
While he works, you know it. When a request needs him to look something up — your deliverables, your calendar, your emails, the internet — a short notice arrives before the answer, such as "⏳ Looking at your calendar…": only one per request, however many lookups he makes, on WhatsApp and on Telegram. If Orion lives on your own number, where the phone does not show "typing…", he also sends "⏳ One moment…" if nothing has arrived after a few seconds, and "⏳ Still on it — it's a big one." if half a minute goes by with nothing but these notices. The confirmation that a file was received counts as one of them. Never more than two of these notices per request, and none of them promises a time: the notice is not the answer — the answer follows in the same conversation. If a lookup gets truly stuck, on the plans that run on our AI key Orion does not wait indefinitely: he tries once more and, if that still fails, tells you plainly that he did not finish, says whether anything may already have been done, and asks you to send it again — if it is something large, like a long document, it is best to ask for it in parts.
When the answer fails without delay. Sometimes the answer does not come through because the AI service Orion uses returned an error or a blank answer, not because it is taking long. When it comes back blank, Orion tries once more on his own before telling you. In either case, in your direct conversation with Orion — WhatsApp, Telegram, the Ask Orion chat in the dashboard and the mobile app, voice included — he does not show you the technical message: he tells you plainly that he could not answer, says whether anything may already have been done, and what the next step is. If the failure was temporary, ask again later; if it was an error that sending the same message again will not fix, ask in a different way — if it is something large, like a long document, in parts. He does not come back with the answer on his own later: you are the one who asks again. This applies on every plan, including when your organization uses its own AI key.
The same long text message four times in a row. In your direct conversation with Orion by message — not in the dashboard or the app —, if the same long text message — 40 characters or more and at least 6 words, counting only what you typed: the size and form of an automatic reply, such as an away message — arrives from you four times in a row within ten minutes, each one after a reply of his, on the fourth he does not answer the request: he tells you, once, that it looks like an automatic reply and that he will not answer it again while it keeps arriving unchanged. Identical messages that arrive after that get no reply; the pause ends ten minutes after the last one, and writing anything else ends it right away. The same sentence three times is still answered, and a short message never counts, however many times it is repeated — next, more, continue, yes, ok, send the next one, a number, an emoji: each one is answered. Voice notes, photos and files never count either — not even when you resend a file or an audio that did not come through, as Orion asks: each one is answered.
Many messages in under a minute. In your direct conversation with Orion by message (WhatsApp or Telegram) — not in the dashboard or the app —, he answers up to five messages a minute. From the sixth in the same minute on, they are left out: on the first one left out he tells you, once, that he got many messages in a row and asks you to wait a minute and send it again; the ones that arrive right after, in the same minute, are left out too, with no further notice. Once the minute is over, everything is back to normal.
A file Orion refuses for safety. A PDF with an embedded script and an Office file with macros are blocked before being read — it tells you which file it was and why, and does not open it. This is the same check that already applied to email attachments; now it applies to what arrives over WhatsApp too. If the file is legitimate, export it again (for example, save the spreadsheet as .xlsx without macros) and resend.
Ask for your audio notes. "What did I dictate yesterday?" · "read me the note from that meeting with the client" — it lists the dictations and the meeting notes from Audio to Notes, most recent first, and reads out the one you pick; if you want the text with no treatment, ask for the raw transcript. It is honest about what it has: if the transcription was in doubt about a name or a number, it says so instead of asserting it; if the note is still being processed, it tells you that instead of inventing the content. Recording still happens in the dashboard — here it reads, it does not record. And this reading only happens in your direct conversation with it: in a group, or in work run by an AI worker, the request is refused with the reason.

Relationships and CRM

"Add the contact So-and-so, phone X." · "Summarize my sales pipeline."
"Who am I losing touch with?" — it keeps track of the people who matter and tells you, gently, who is going cold.
"Remind me to talk to Maria on the 30th about the proposal." · "Where do things stand with Ricardo?" — the CRM chapter explains where all of it is kept, and what it can and cannot move on its own.

Personal organization

To-do list: "add to my list: book the dentist by Friday." The deadline shows up in the morning brief as a visual reminder — it does not act on its own on that date. A new item always lands at the top of the list — you just decided it — and from there you drag it wherever you want. What Orion adds on its own (a commitment it overheard, a client follow-up) goes to the bottom: a suggestion does not outrank what you wrote.

In Orion → List, every row answers to three gestures, and only three: the checkbox marks it done — and nothing else does; dragging by the ⠿ handle changes its position — on the computer and on the phone; double-click opens it for editing (the Edit button does the same). Clicking the item text does nothing, on purpose: a glancing click used to mark the task as finished.

Reminders with a time: "remind me every day at 8am to review the pipeline." Here it acts at that time. Use "remind me / every day / at X" when you want action at the right moment.
Memory: "remember that my company focuses on SMBs of 50 to 500 people." It stores that and uses it later. What it stored is visible at Orion → Orion's Notes.

That memory is yours and only opens up in your direct conversation with it. Since August 15, 2026, inside a group — including a group declared as internal work — reading and writing in your Notes is out of its reach: nobody there can see what you stored, and nobody can plant a "fact" that it would later repeat back to you in private. The same goes for an AI worker: even if you tick "memory" in its Skills, those capabilities stay out. It is not a permission anyone can switch on; it is an absence.

Knowledge: saving and deleting

Orion → Knowledge holds the notes and files Orion looks at before answering, each one 🔒 Personal (only you see it), 🏢 Organization (the whole team sees it) or 🌐 Public (Orion also uses it with outsiders). Create one with the new-note button or attach a file; open an item to read it, ✎ Edit it or 🗑 Delete it. In the answer to Ask, each cited source carries the same label. Audit only reads the company's Knowledge: the organization's items and the deliverables' items show up for it as read only, and it reads, searches and asks the library. Its own 🔒 Personal notes are allowed: the new-note and attach buttons always save a 🔒 Personal item, which it can edit and delete. Its Orion and an AI worker working at its request don't save anything to the company library — they hand over the material and say why.

Without a subscription. Saving, opening, editing, deleting and searching continue — notes and files too. Asking the library, which puts the answer together with AI and cites the sources, gets a 🔒, and so does Link to a deliverable: linking a note or a file to a deliverable is creating in Y Managers. The click states the reason and the way back.

Deleting an item. Open the item and tap 🗑 Delete, next to ✎ Edit. The button shows for whoever can delete the item: a 🔒 Personal item, only for the person who saved it; an organization item, for anyone in the organization except Audit; a deliverable's item, for whoever works on the deliverable, whoever manages its team, and the Admin. Audit only deletes its own 🔒 Personal notes: not an organization item, and not a deliverable's. Whoever can't change the item sees it as read only. The confirmation has two steps: tick I understand this can't be undone and tap Delete for good. There is no undo on screen.

What happens. The item leaves Knowledge and Orion's search right away: the next searches, in public service too, no longer find it. An uploaded file is deleted for good after 30 days. What someone already read, or what Orion already answered with the item, can't be taken back.

A survey's document. While the survey is turned on in a line, its document can't be deleted: the screen says which line it is on and, if the line belongs to someone else, whose it is. Turn it off first — the person in charge of the line turns off Run on this line in Orion → Public role, in the Run a survey block. Once it is off, the document can be deleted: the answers stay in Survey responses, with the questions of the approved version, and the survey can no longer be turned on — to run it again, choose another document in the block. See Run a survey.