The inbox

The inbox is where the system asks you things. Everything waiting is one list, one row per thing: what it is about with the kind of thing it is after it (a model, a metric, a fact), the one line that says what is asked under that, and at the right when it arrived and where it stands in a word: Broke, Behind, Needs a key, Moved, Needs a yes. Each row has an icon for its kind, with a red mark on it when the thing broke. Its three tabs carry their counts: how many things wait on you, how many improvements are still open on Later, and how many were handled for you this month, which the line beside the title says too: "173 were handled for you this month." Click a row and it opens in a panel on the right with its buttons under it; a fact's Details there say whom it is about, what it says (change the words if they need it) and since when, already filled in from the card, with where it will be saved and its key folded under Advanced; Details opens by itself only when the card is missing something. Yes, that's right keeps the fact as the card says it. Empty a box there and that part is left out: a new metric is saved without it, and one you already have keeps what it has. Later sets it aside. Answering is the single most useful thing you can do, because your answer becomes something it knows for good.

Find what needs you

The count is of rows: the things that need you. It is the same number beside Inbox in the menu and on your phone. Home counts only the rows that ask you something (a metric, a fact, a number that moved); what broke is said on its own there. A row can stand for several things: the eighteen things heard in one interview are one row, not eighteen. Rows come 25 at a time. A list longer than that has a search box, a menu of sections and next and previous pages; on a shorter list they are behind Search, under the table. Choose Feedback on answers there to see only the answers people marked wrong or right with a nuance. Check for problems now, also under the table, looks for a failed sync or a source gone quiet right away; it looks on its own every few minutes. A row you answer leaves the table at once. If the list cannot load, it says what went wrong and offers Retry.

The kinds of rows

• Things it heard. From the interview, a chat, a note, or an agent. Everything heard in one interview or one chat is one row: "Your interview: 18 things it heard, waiting for a yes". Its panel lists each thing with a tick, the first six and a link to the rest. Yes to the 17 ticked saves the ticked ones, each exactly as approving it on its own would; Not these leaves the ticked ones out; anything unticked stays waiting on the same row. One that needs more than a yes (which Sam is meant, a detail it could not fill in) starts unticked: click its words to open it on its own, with its Details, and use the link at the top to come back. Later sets the whole row aside as one item. Something heard on its own is a row of its own, and says whether it is new, confirms what it knew, changes it, or contradicts it. Approved, it becomes a confirmed fact with your name on it: a yes to its sentence is enough, and nothing has to be filled in to agree with it. Only a person can confirm; nothing a model writes is ever confirmed on its own.
• A number that moved. Every night it goes through every definition and works out again each one whose data has changed since it was last measured. A value that moved more than 2% from last time is one row for that metric, however many nights it happens, with what it was and what it is: Use the new figures accepts it, Ask about it starts a chat about it, and Later sets it aside. If the number goes back to what it was, the row closes itself. A metric whose query no longer runs is a row of its own until it runs again, its line saying what is wrong in plain words where it can ("Can't be worked out: what it reads has no column called credit_memo_total") and its panel keeping the database's own words for whoever fixes the query: Open the metric goes to it, and Seen closes the row until the next night's check, which raises it again if it still does not run.
• A metric nobody has confirmed. Each is a row of its own, named for the metric, with the first question about it: "Does net sales include wholesale?" (what its data shows that its definition does not settle), or "Is this how it's worked out?" when there is none, or "The model it reads, westside.daily_sales, changed on Oct 20. Is it still right?" when a model under a confirmed metric was rewritten (the mark comes off until someone confirms it again). Its panel is the metric itself: what it is in a sentence, the question with its answers as buttons, one at a time, and Confirm for an admin or a named confirmer. Nobody is emailed about them. Confirmed, the row goes.
• A metric a source proposes. When a source's first sync lands, each of its obvious metrics is a row ("Mrr metric, From Stripe. Trialing and past-due subscriptions are left out. Is that right?") whose panel has its definition, last month's figure with Where from, the choice it makes when it makes one, Confirm for an admin or a named confirmer, and Not this. Confirm makes it a metric with a receipt, and Not this removes it for good.
• Something broke. A sync that failed, a key that expired, a table whose shape changed. The card is the agent's own message to you, written from what it found. It says what failed and when, what it read, what looks wrong, and what it offers to do ("I saw the Stripe sync fail at 6:02 am on Oct 21. I read the error, and it looks like the source refused our key... I need a new key. Paste it into the card and I will try again."). A failure it already ran again says so ("I ran it again and it failed again"). Its button does what it offers, and the card closes itself when the next run works. When AI on its own is on (Settings, Company), the AI reads the run's error and its log to say what looks wrong. When it is off, or at its ceiling for the month, a rule of thumb picks the fix and the card says so ("a rule of thumb, since AI on its own is off"). Ask about it opens a chat that already knows the run, its log and the fix it offers, so "what else could it be?" is answered from the run itself. Say go ahead there and it presses the card's button for you. It is the button, not a judgment, pressed once, as you, and only when you say so. A new key still goes in through the card's Replace the key, never into the chat. Not this closes it for that failure, and a run that fails after your answer is a new row. A source or a model that falls behind its schedule is one row however many times it happens, and closes itself when it catches up. A sync, alert or email that stopped because the key or person that set it up is gone says so too; Noted closes it. A model whose checks held its new rows (a month already reported that changed in this week's file, months that no longer add up to the year to date) says which check and what it found, and that the model was not updated, so it still shows the numbers it had. Those rows are fine, update it, for a person who may change that data, puts the new rows in and checks the next file against them. Not this leaves the table as it was. The card says whom it reached: "Sent to Rachel Moore by email, 1 phone and Slack."
• A query people keep running. The same query run many times in a month is a card to keep it under one name, so everyone reads the same rows: Keep it as a model makes it a model (a view, on Free).
• An alert. A number you asked to be watched crossed its line, once per incident, under Broke or went quiet: the alert's name as the Alerts list says it, and what happened ("Below its line: $56,702"). Open it for the value, the line and when, and, for a defined metric, how it was worked out. Under them is what it looked into before it told you, a few lines with the fix a behind source's card offers, and above them a few sentences written by AI from those lines when AI on its own is on. Open the alert goes to it, Ask about it opens a chat that already has what it found, and Noted closes the card. Under its words, whom it reached ("Sent to Rachel Moore by email and 1 phone"). Nothing about an alert is ever saved into Context, and the card closes itself when the alert recovers; the next time this month it reopens the same card.
• Still right? When a finished answer has used something it only heard and nobody confirmed, it asks once, for that version: "... was used in an answer. Still right?" Something it heard over a year ago that no answer has used is asked the same way, ten a night. Still right, keep it confirms it. On Never ask (Settings, Company) it asks nothing.
• A finished month that changed. Every night every metric whose data has changed is measured again month by month. A month that has ended and settled, and then moves (rows landed late, a model changed, a sync dropped rows), goes on that metric's one row, addressed to the metric's owner: every changed month with what it was, what it is now and the change. Use the new figures accepts every month on the row; Ask about it starts a chat about it; Later sets it aside. A month that moved by less than half a percent of its own figure is accepted without asking and listed under Handled for you. The figures show to people who can read the data behind the metric; anyone else sees which months changed.
• A page number close to a new metric. A metric defined after a page works something out its own way, close but not the same: one card with both figures. Use the metric switches the page's number to the metric; Keep the page's number keeps the page's own calculation for now (about that version of the metric only). A number that is the same calculation over fewer rows gets the same card and the same two buttons, and a number whose link a new version undid gets a card with Open the dashboard. To make the page's calculation a metric of its own, press Define them on the dashboard's Where the numbers come from tab: the builder's chat reads it back to you and defines it on your yes.
• An old answer close to a new metric. A chat answer's own SQL that nearly computes a metric defined later (a filter left out, another date): "was it net sales?" Use the metric counts those answers as the metric's; Leave the answers leaves them as they were.
• Facts from a document. When a deck, a contract or meeting notes you keep in Warehouse, Files is read, it suggests up to five facts, each said as the document's: "Ben's plan sets Riverfront's target: full run-rate by Q6", "Ben's plan requires...", never a plan's target as if it had already happened. They are one row, "4 facts from Ben's plan". Its panel says first what Overscore read and what each answer does, then lists each fact with its page and a tick. Open one on its own for the sentence, the words on the page it rests on, and a link to the file at that page; where it is filed is under Advanced. Only people who can open the file see them. Approve one and it's a fact about the company: The company in Context lists it, the chat answers from it, and its source is the file and the page.

Feedback on an answer keeps any Number check notice below the saved answer, so the caution is still there when you review the card.

Handled for you

The Inbox's third tab: what closed itself this month, so nothing vanishes. One row per thing and what happened to it, newest first: a table that fell behind 26 times and caught up each time, a failed run that worked the next time, past months that moved by a hair and were accepted, a dashboard's checks of older versions replaced by a newer one, a draft's checks, an alert that came back inside its line, a number added up the same way in several places with no metric for it yet. That last one is a note: the chat that builds pages reads it, so a new page adds the number up the same way, and nobody is asked about it. Each row says at its right, in a word, what happened: Worked, Caught up, Accepted, Replaced, Came back, Settled, Kept or Noted. Click a row to see what was said each time, or the months with what each was and is. There is nothing to press: none of it needed you.

The count on the Handled for you tab, and the line beside the title, count every time, so a table that fell behind 26 times is 26. The month is your company's own, by its time zone.

Checks from teammates

Answers a teammate sent you to check (Ask someone to check, under any chat answer) sit at the top. A check can be sent to a teammate who has never signed in here or whose current access cannot be verified; known insufficient or revoked access is refused. Opening and answering always require current access to all the original tables, without masking, including those behind its metrics. Open one here or with the email's Open review button, see exactly what it rested on, and say looks right, right with a nuance, or wrong and why.

Later

The Inbox's second tab: the improvements people noted and nobody has made yet, grouped by what they're about (tables, models, metrics, dashboards, sources). Each is a row: what to improve, then where it lives, who noted it and why, with the day it was noted and its status (Open, In a session, Done) at the right; the tab's count is how many are not done yet. A row answered "Later", including feedback on an answer, lands there instead of vanishing (a row of several things as one item that lists them), and keeps who may read it: someone who couldn't read the card doesn't see the item; the assistant and agents add what they spot but aren't fixing now; you can add your own.

Click one to see where it lives, where it came from ("from Sarah's chat, Sep 29"), and what happened to it. Work on it in a chat starts a chat with it; only once that chat exists does the item say "in a session". The assistant looks into it and says what it would change before changing anything. A failure you answered Later stays there: the same thing failing again raises no new row while the item is open, and the item is marked done when it next runs fine. It is filed as about the model that failed, and its chat opens already holding that model: its saved SQL, what its last save changed and the failed run's error. The first reply says what broke and proposes the fix; nobody is asked to find or paste the SQL. Come back next week and choose Resume the chat to pick up that conversation. Start a new chat is a separate choice when you want a fresh conversation. Put it back returns the item to open and keeps its chat for later. With several open in one area, Work on these in a chat starts one new chat for those items. Mark it done with a line on what changed, and it stays as history. One about a table or a model shows only to people who can read that data.

On a phone, each item stacks what needs doing, who noted it and when, and its full status, so you can read it without sliding across a table.

Cards by email

A card can be addressed to whoever owns that number. They get one email with the question and a link that opens that card in the app, where they answer it. A metric whose finished months changed emails its owner once, when its row is made, not once for each month.

What to say to it