Ask a question

Type the question the way you'd ask a person. It reads what it knows about your business first (your words, your definitions, what happened last month), then the data, and answers with the number and the table it came from. The query it ran is folded under the answer if you want it.

If a word in your question could mean two things you have both told it about, it asks which before answering. Otherwise it takes the plainest reading and says which one it took. Correct it once and your correction is kept. A word written down as one metric's other name (GMV for net sales) is that metric, so it is never asked about, and the answer says "GMV is another name for net_sales".

Ask in the box on Home, which starts a chat with your question as its first turn, or on Chats: one question or a piece of work that takes a few steps and builds something, the same chat either way. On Free the chat is the welcome, a new company's first three messages, on us. After them your own Claude Code does that work, through ovs and MCP (see Limits and plans).

A chat opened with Ask about it on a card in your Inbox starts with the card's own message, already knows the run behind it, and presses the card's button for you when you say go ahead. It is the button, not a judgment, pressed once, as you.

Beside every page

Ask, the first thing on the left under your organization (or ⌘J, Ctrl+J on Windows), opens a chat beside the page you are on, and closes it again. The page stays where it was, beside it, and the left side folds to its icons while the chat is open. At its top are the chat's name (click it to rename the chat), then three buttons: + starts a new chat, the arrows open it in Chats at full width, and × puts it away. Ask on the same page brings that chat back as you left it; on any other page it opens a new chat, looking at that page. Under the name it says what it is looking at, the page beside it: a table's name with its columns and rows, a query, a dashboard. A new chat lists your last three chats to pick one up, and the box is ready for your question. About this table and the chat take turns on the right side: opening one puts the other away.

It is one chat, and a chat in Chats like any other, listed there with the rest. On a dashboard's page it is that dashboard's own chat, the one Build with the agent opens: a change it makes is a draft, and the draft's card opens the builder, where it is looked at and published. Anywhere else, while it is open, it goes with you from page to page, and + starts another; under a new one are your last three chats, to go back to one. A chat that was beside a page has Open beside and the page's name at its top in Chats (Open beside Query 1): it opens that page with the chat beside it.

It asks back when it needs you to go on: which table to join, which date to count by. The reply names what it found and ends on the question, and its answers are buttons under it, beside Not now; press one, or type. A reply that asks has no Right or Wrong under it. In Warehouse, a query it writes for you goes into your editor, in the query you have open when its editor is empty, else in a new query tab, and the chat says which in a quiet line (Wrote the query into Query 2); press Run when you want it run.

Each time you send, the chat is handed what the page beside it holds, so you can say "this": a table's description, rows and columns with what each holds; a query's SQL and the rows it gave; a dashboard's figures as its page drew them; a model's SQL and its last run; what is connected, on Sources; the Inbox card you have open, with its run and the end of its log when something broke; Home's line. What it reads for this, it reads as you, so it is handed nothing you could not open yourself. Each message keeps what it was looking at: read again later, in Chats or beside a page, the chat says Looking at and the page over a message wherever the page changed.

Someone in a group (a client looking at the dashboards shared with them) has the same Ask, above Dashboards, and the same chat beside every page. On a dashboard it is handed the cards they see and their numbers, and it answers only from what is shared with their group: asked about a table or a page that is not, it says so in words and what they can ask about instead. Ask about this number, on a number's receipt, opens the chat beside the page already asking where the number comes from; on full screen, the chat over the page.

Beside a table nobody has written anything about (it says No description yet), the chat's first offer is to write it down: Draft it has it look at the rows and file what the table is, what one row is, what to use it for and what each column means, marked Not confirmed yet until a person says so. It asks you about the two or three meanings a wrong guess would cost most (is the price before or after discounts?), with They're right under its question: pressed, those words are filed as yours, confirmed. Tell it what is right instead and it puts your words to you the same way.

Ask is also a round bubble in the bottom right-hand corner of every page, on a computer and on a phone: it opens the same chat, looking at the page you are on, and stays clear of the page's own buttons. On a dashboard's full screen the bubble floats over the page and opens the chat over it; Esc puts the chat away, and Esc again leaves full screen. On a phone, which has no left side, the corner is Ask, and the chat is the whole screen, with the arrow at the top back to the page.

Things to ask

What it can and can't do

• It changes nothing you did not ask for. A quick answer only reads. A chat builds or changes something (a dashboard, a metric, a model, your fields on a table's rows, a table you fill in) only when you ask, or say yes to what it offers, and it never changes what a source sent. Nothing leaves Overscore until you press Send on its card. What it hears you say about the business is kept as a fact, or waits in your Inbox for a yes.
• It sees what you see. Every answer runs under your own key, so a person who can't read the books can't get the books out of it. Someone in a group that was shared a few dashboards gets a chat that answers from those dashboards and their metrics, and says plainly what it can't see.
• It shows its work. Every number carries a line that says where it came from: the metric it used and who confirmed it, or the tables it read. If the answer looks wrong, that line tells you which part to fix.
• Name a table, or a file you just added, and the answer comes from it: "what were total sales in owner-first-source.csv?" works the moment the file is in. If the quick answer can't reach what you named (a document kept as a file, a table that isn't there), it says so and offers to carry the question into a chat. It never answers from a different table instead.
• It writes a metric's number the way the metric says: dollars with a dollar sign, a share as a percent. It never guesses what the rows it added up were: the count stays out of the answer, and the line under it says how many rows of which table.
• In a chat it says two things a metric's number can't say for itself. When your word is looser than the metric's name, it says in a line what was counted and by which date ("sold" to wholesale, answered by net sales: counted by invoice date). When the data stops before the period ends, it gives the data's last day (this year to 4 October), not today.
• A quick answer shows up to 200 rows. In a chat, a table it draws keeps its first 500 rows, and Download as Excel or CSV under it saves every row. For a whole table, open it in Warehouse and press Download.
• Your recent quick answers are kept in this browser for you, in this company. Someone else on the same browser, or you in another of your companies, sees none of them, and signing out clears them.

A table or a chart in the answer

A chat answer draws a table or a chart under its words whenever the answer is a split (by channel, by store), a ranked list (top products), a trend (by week or by month), a count by day, or one figure with its change, whether you asked for one or not. It draws a figure with its change and its trend, a table you can sort (one wider than the answer scrolls sideways, with a shadow on the edge where more columns are), or a bar or a line chart, with Table and Chart to switch between them. It is drawn the way a dashboard draws it, from the numbers a dashboard would ask for, read as you, so it shows only what you may see. The words say what it means and the headline figure once; the rest of the figures are in the picture. A table keeps its first 500 rows, and its foot says so when there are more. Its title says what it shows and, for a defined metric, its period in words ("Net sales by channel, this month"), so a page, a file or a weekly email made from it names the right days every time it is worked out again.

Under it, Put this on a dashboard makes a page of it, named after it, or adds it to one of your pages. The builder opens on that page, looks it over and says in a line how it looks; it changes nothing until you ask, and nothing is published until you say so. Saying it works too: "turn this into a dashboard", "add this to the sales page".

Keep or share holds the rest, and starts with Save as an analysis (below). Download as Excel or Download as CSV saves every row, headed as the table is, and the file says what it is: as of when its figures were worked out, the period, the definition used and its version, and the SQL (on a sheet of its own in Excel, in the first row of a CSV). Email it sends it to the people you pick, now or every Monday at 7:00. Now is one email, and nothing is kept; every Monday is on Dashboards, Emails to change or stop, the people you pick are added to the ones it already goes to, and if your company has no time zone yet it asks which one first: an admin's answer becomes the company's time zone, as in Settings, Company. Each email works it out again and ends with its as-of and a link to where it came from. Where from shows the definition, the SQL as it ran, the tables it read, and when it was shown to you. Each is on the record like a number on a page, and each works in words: "send this to Nina every Monday" sets the email up for you to turn on.

Ask "how are our models running?" and it draws your models' runs a day for the last two weeks, the failed ones in amber; a source's pulls and your sends work the same way, each counted on its own. Prefer words? Turn off Draw tables and charts in answers in Settings, Account, and your answers are words only.

Keep it as an analysis

When a chat has found something worth keeping, keep it whole. Save as an analysis is first in every table's Keep or share, and in the chat's ··· at the top for a chat without one. It asks a title, filled in with your first question, and Save keeps the whole chat as it is now: every question and answer, every query with its rows and time, every table and chart, and the last answer as the conclusion. Saying it works too: "save this as an analysis called September by rep". It is kept under Dashboards, on the Reports tab, frozen: opening it later shows exactly what it showed. See Dashboards, Reports, for what you can do with one.

The line under a number

Every figure in an answer, here or in a chat, carries a line the software writes from what actually ran. The AI never writes it, so it can't claim trust it doesn't have. There are three kinds:

• Trusted. The figure is a defined metric's own calculation, and a person confirmed the definition. The line names the metric, its version, its owner and who confirmed it.
• Worked out by hand. The AI wrote a query for it. The line names the tables it read and whether each is signed off, a model, or raw. Honest and traceable, but nobody has agreed that this is the right way to count it. If it's a number you argue about in meetings, define it as a metric and have an authorized person confirm its definition.
• Not found. The figure isn't in anything the AI looked up. It was sent back once to be redone; if it still doesn't match, the line says so. Check it before you use it.

A defined metric whose definition nobody has confirmed yet, or one asked with an extra filter, says exactly that instead of Trusted.

Under an answer with figures in it there is one line that says how far to trust them: a green dot and Trusted when every figure comes from a metric a person confirmed, an amber dot and Not confirmed when any was worked out by hand or from a metric nobody confirmed, and a red dot with Check before using when a figure is in no result. See how opens a line for each figure: the metric and its version, who confirmed it, or the numbers a calculation used. A calculation is only as trusted as the numbers it used: a total of figures worked out by hand is Not confirmed. For a calculation, the line says only that its arithmetic checks out, and lists each number it used with its dates: a metric's period and the day it was measured, the posting date on the rows a sum came from. Whether that calculation was the right one to make is yours to judge: a late bill posted before the metric was measured is already in it. The steps it took fold into "3 steps"; click that to see each one.

If figures still do not match the results after the answer is redone, the chat shows a separate Number check notice beneath the answer. Overscore writes that notice; it tells you which figures to check before using them.

From a document

When a chat reads a document, it adds a link such as "From vendor-contract.pdf, page 2" under the answer. Click it to open the file version and page it read. The links come from successful file reads and search results for that answer, even if there are no numbers in it. The AI cannot invent these links by writing a filename; they come from the recorded pages.

When the answer is outside your data

A chat reads only your data until you say otherwise. Before it reads a website you did not give it, it asks on a card: Allow once (for this chat), Always for this site, or Deny. When a question needs the public web (how another system works, a public figure to compare against), it writes the questions on a card, with the sites it may read and what it costs, and waits for Run it; a separate researcher, which sees none of your data, searches and reads, and the chat answers from its note. An admin chooses how research works under Settings, Plan and billing. See Agents and keys, Research on the web.

Anything that sends your data out of Overscore (rows to Meta or HubSpot, a new place to send them) is a card too, and nothing goes until you press Send on it. A key or a password you paste into a chat is handed to the step that needs it, never to the AI.

Find an older chat

The new-chat page starts with the question, four ways in and the box to ask in, then your recent chats as rows, with a search beside them: the name, the first words of the last answer, and when. Chats shows 25 conversations at a time, with the total and controls for the next and previous pages. Search by words in its title, your questions or its answers, even for a chat from months ago. Open one and keep talking; Back to chats returns to the same search and page.

Ask someone to check

Under any answer in a chat, press Ask someone to check and pick a teammate. Choose how much they see: this answer, the chat from this message on, today's chat, or the whole chat. They get an email and a row in their Inbox, and a page showing exactly that much: each answer with the line under its figures, and what it rested on: the metrics and their versions, the queries it ran and the tables they read, the notes and facts it used. They can say they're looking into it, then whether it looks right, is right with a nuance, or is wrong, and what was off. The verdict shows under your answer ("Checked by Nina, Sep 27") and on the record. A nuance or a wrong goes to your admins as a suggested fix, so the mistake isn't repeated.

Before sending a check, Overscore refuses if it knows the teammate lacks access to a table behind the answer, including a table used by its metrics. A teammate can still be invited if they have never signed in here, or their current access cannot be verified. Opening or answering the check requires signing in with current access to every original table, without masking. If an older answer did not record what a metric read, ask the question again before sending it for review. If your admins have marked someone as able to confirm (Settings, Company, Who can confirm a metric), their "looks right" also confirms the defined metrics the answer used.

When it gets it wrong

Say so, in the same box. "No, net revenue is after refunds." The correction is kept as a fact with your name on it, and every later answer, dashboard and agent reads it.