AI in DataSquares

DataSquares has one AI, the DataSquares AI assistant. It answers questions about your data with the numbers, the chart and the SQL behind them, builds dashboards, reports, alerts and pipelines once you approve exactly what it will do, and writes summaries and drafts across the product. It works with your data permissions, it runs when someone asks for it, and it is paid for in AI credits.

DataSquares chooses the AI behind each feature, so there is no model to pick: you choose how much effort an answer gets. Asked what powers it, the assistant says it is the DataSquares AI assistant and that the details of the technology behind it are not shared; it can tell you the effort you chose.

Where the AI appears

Where	What the AI does	Read more
AI Chat: the AI page, the panel on any page, Ask AI and Build a dashboard with AI… in Ctrl+K, Build with AI on the new-dashboard page, Ask about this card, and Ask AI in the dashboard editor	Answers with a chart or table, the SQL and the evidence; builds or changes dashboards, reports, cards, filters, alerts, email schedules, calculated fields and pipelines once you approve; explains a failed pipeline run	AI chat
Dashboard cards	Summarize with AI writes a summary from numbers computed from the card, which you preview and then show on the card; Analyze’s Phrase it words what Analyze found; Explain with AI adds a short summary to the breakdown Explain this change computes	Card insights
Governed metrics	Explain with AI adds a one-line summary to a metric’s breakdown; a digest’s AI sentences, once its creator turns them on, reword its pace sentences	Why did it move? · Digests
Stories	From a question when you start a story (the AI writes the first section from your data), From a goal (an outline of the sections), Write with AI, Write from the data (a section’s narrative from numbers computed from its data), What should come next?, rewording a selection, Reword with AI for stale numbers, a grounded proposal in a claim review, and Ask AI on a comment thread	Write with AI
Reports	AI layout suggestions for a paginated report, and an optional AI summary of a comparison between two runs	AI layout suggestions · Compare changes
SQL editor	Explain a query, and Generate SQL from a sentence	AI assistance
Presentations	Draft from data writes an outline, then the slides, from figures read with your access, and every figure stays live; Draft slides with AI does the same in an open deck	Draft from data
Data models	Draft with AI in Prep for AI drafts missing table and field descriptions, published only once an editor approves them; Explore this model walks every table of a model with the schema finder, on an editor’s click, where the server turns it on	Prep for AI
Measures and themes	Draft with AI in the SquareX measure editor; Draft with AI for a theme	SquareX · Brand generator
Pipelines	Suggest recipe drafts a transform from plain language in the pipeline editor; Suggest mapping matches an activation’s columns to the keys its destination takes, which you check and save; in AI chat, a pipeline is built, changed or run once you approve	AI authoring · Activations · Pipelines from chat
Data quality	Add AI findings on a profiling scan: the AI ranks and words the scan’s suggestions, never from a column labelled sensitive; Describe the check drafts a check from a sentence, saved only when you save it	Data quality
Catalog	A paragraph that explains the lineage of a dashboard, a card or a metric, from its source on, written on a click from the lineage the Catalog draws by rules; on Business and Enterprise	Lineage
Share links and embeds	Ask AI for viewers answers questions about that share’s cards only, once the publisher turns it on	Ask AI on a share
Large data models	The assistant reads as much of a model at once as its size allows, all of it up to a size its effort sets where the server turns on the compact model listing; on a model of several hundred tables, AI search finds the tables a question needs from a catalog of its tables and fields, or, where the server turns it on, the schema finder walks a model larger than the effort reads whole for the question, within the AI scope an editor sets	How the assistant reads a model · AI search on large models · The schema finder
Search	Ask AI in the Ctrl+K palette sends your question to AI chat; the Catalog results are a search, by meaning where the server turns that on	Search and Ask AI
Android app	AI Chat, from Home, from a search or from Ask about this	AI chat on mobile
Browser extension	Answers in Chrome’s or Edge’s side panel, beside a CRM, an ERP or any web app, on the data your administrators set for each site; a page’s text joins a question only when you add it	Ask AI beside any web app

When the AI runs

Apart from two features below, the AI runs only when a person asks for it: a message in AI chat, including a question sent there from Ctrl+K or a card’s Ask about this card; a click on an AI action, such as Summarize with AI, whose label says it uses AI credits; or an AI option someone turned on, such as Ask AI for viewers on a share.

Two features use AI with no AI click. On a server that turns on search by meaning, each search in the Ctrl+K palette, and keeping the search index up to date as your work changes, use the workspace’s AI credits (Semantic search). On a server that turns on AI search, keeping the AI catalog of your models up to date as they change uses them too (AI search on large models). A company administrator can turn either off under Administration → AI (AI controls).

Some features do their main work without AI, spend no AI credits doing it, and add AI only when you choose it:

Explain this change and Analyze compute their numbers as statistics on the server. A written summary comes from the AI only when you choose Explain with AI or Analyze’s Phrase it.
A metric digest sends sentences written by rules. The AI rewords them only when the digest’s creator has turned on its AI sentences.
A data quality profiling scan finds its suggestions with rules. The AI reviews them only when Add AI findings was ticked for that scan.

Applying a change you already approved in AI chat runs no AI at all.

Effort

In AI chat on the web, Chat options, the button in the chat’s header, names how much work the AI puts into your next message. Open it for the four efforts, each named by the effort it gives; Ctrl+Shift+. (⇧⌘. on a Mac) while you type moves to the next one. A higher effort does more work and uses more AI credits. Every plan has all four efforts, a trial included.

Effort	What it does
Quick	Light effort, for lookups, single numbers and a quick chart.
Balanced	Standard effort, for most questions.
Deep	High effort, for dashboards, reports, Stories and hard questions. It used to be called Thorough.
Max	Maximum effort, for the hardest questions.

Until you choose one, Chat options names your company’s default effort: Quick on Team and Balanced on Business and Enterprise, unless a company admin chose another under Administration → AI → Effort (Set the company’s effort), or a workspace admin set that workspace’s Answer depth (AI settings).

Your choice stays. It is saved to your profile, so it applies to every new conversation and every AI chat box, on any device you sign in on, until you change it: in the chat, in Write with AI in a Story, in the SQL editor’s Explain and Generate, and in Draft with AI for SquareX measures. Choosing your default effort counts as a choice too (see Effort in a conversation). A choice you made in one browser before your choice was saved to your profile moves there the first time you open the chat. The Android app has no effort choice yet: its questions run at your saved effort.

Your company may:

hide the effort choice: every conversation then runs at the company’s default effort, and no choice of yours is used. The default can still move a conversation up one effort by itself, as it does with the choice shown (see Effort in a conversation);
set the highest effort people can choose: the efforts above it are shown in Chat options, marked Not offered here, a choice saved before it was set runs at the highest, and no conversation moves above it by itself;
set the viewers’ effort: viewers, Ask AI on shared links and embedded chat run at it (Quick, unless an admin allows more on Business or Enterprise). It is also the highest effort viewers can choose, and their conversations never move above it by themselves.

The other AI actions (summaries, rewording, drafts in an editor) each run at an effort DataSquares sets for that action.

Effort in a conversation

A conversation keeps the effort its first message ran at, which keeps its follow-ups quick. It changes only like this:

You choose a higher effort: it applies from your next message.
You choose a lower one between Balanced, Deep and Max: it applies from your next message too. The three differ only in how much effort goes into each step, so a move between them, either way, keeps the conversation’s follow-ups quick.
You choose Quick in a conversation that runs at a higher effort: the conversation moves to Quick once the AI summarizes the oldest part of a long conversation, once the conversation has gone unused for a while (five minutes after its last answer, or an hour from its third message on), or from the first message of a new conversation, whichever comes first. Until then each step runs at Balanced’s effort, and on a server that turns plans on, the AI already does Quick’s lighter work from your next message: no more rounds of work than Quick allows, and no plan. A message that would have needed more rounds can then use fewer AI credits.
Until you choose an effort, a follow-up that asks for much more than the conversation so far, such as building or changing a dashboard or report, or an analysis of why something changed after a few lookups, moves the conversation up one effort: from Quick to Balanced, or from Balanced to Deep. A conversation whose first message asks to build a dashboard or report starts one effort above your default. This happens with the effort choice hidden too. The default never moves a conversation to Max by itself, nor above the highest effort your company lets you choose (for viewers, the viewers’ effort), and never lowers its effort for a simpler message.
An effort you chose yourself is never changed by the AI, even when it is the same as the default.
Your company lowers the highest effort below the one a conversation runs at: the conversation comes down to it as if you had chosen it, as above.

A higher or lower effort is measured against the effort the conversation runs at, not against the one you chose last: in a conversation at Deep, choosing Quick and then Balanced runs Balanced from your next message.

Questions answered at once

Some questions don’t need the AI to work anything out, at any effort:

A verified answer. When you ask a question your workspace admins have verified, in its own words, in one of the other phrasings they added, or in other words that ask exactly the same for the same period, the AI runs its governed metric or dashboard card again on fresh data, over that period, and answers in a few seconds, marked Verified answer. The same question for another period, or about a product, customer or place the verified question doesn’t name, is worked out.
One number from a governed metric, such as What was revenue last month?: the AI evaluates the governed metric directly, for the period you name when it is one of the date filter’s periods (today, yesterday, the last 7, 30 or 90 days, this or last month, quarter or year, the year to date, the last 12 months), read in the workspace’s time zone. Asked with no period (What’s our revenue?), it gives the number the metric’s cards show: its latest whole period, named, with its change and, when it has a target, how it stands against it and its pace over the period running now. The answer says the period it covers and is marked From the governed metric, followed by its name. A question about part of the metric’s data, such as one product, customer or region, or orders over an amount, is worked out, and so is one for another period, such as a fiscal year or a weekend. It needs the Metrics view permission, as does a verified answer that runs a metric: for a role without it, the question is worked out from the data model.
Again: asking for the same answer again (again, refresh) runs it again on fresh data. The numbers are never kept from an earlier answer. Asking for it again for one product, customer or place (again, but for Germany) is a new question, and is worked out.
An unclear question gets one question back, with each meaning it could have as a complete question you can choose. It never asks twice in a row. Choosing an option, or replying with its number, answers it as that question; a reply that adds to it (the second, but only for Germany) is worked out. It asks which measure you mean only between certified metrics that each fit your question. A question that doesn’t say what to measure (which categories grew fastest this year?) is otherwise worked out: the AI looks up your governed definitions first, uses the certified metric that fits (or, when none does, the one governed measure that fits), and the answer says which one it used. It asks only when several fit equally well.

They use far fewer AI credits than a worked-out answer. When the AI is unsure that a question is one of these, or the metric or card can’t answer it as asked, it works the question out instead.

AI credits

Every AI feature draws on your workspace’s one balance of AI credits. Allowances, packs and auto top-up are on Seats, plans & AI credits.

Counted to the thousandth, shown in whole credits. Each run is charged exactly for its work, with no minimum; every number of credits you see is rounded to the nearest whole credit, and an amount under half a credit reads under 1. See How AI credits are counted.
Every AI run is charged for the AI work it did, including a run that stops or fails partway. A request the AI never started costs nothing: one refused at its start because AI is turned off, because the credits or a budget are used up, because a limit was reached, or because AI isn’t set up.
A step cut off partway is counted once, a little later. When you stop an answer or it breaks off, the AI work of the step it was on may not be known yet. DataSquares asks the AI service for it afterwards, usually within a few minutes, and charges it once, as its own line in the AI ledger under the same feature. Only work the AI service reports is charged, never an estimate.
No numbers on an answer. An answer doesn’t show what it used: its details name the effort it ran at, never its credits. Billing administrators see the balance on Administration → Billing → AI credits. Usage by person, workspace and feature is on Administration → Usage.
Running low, one banner under the page’s header says so, once: members get Ask an admin, which tells the billing administrators, and people who can buy credits get Buy credits. Dismissed, it stays away until the credits run out or the month turns (Running low).
At zero, AI keeps working below zero up to your company’s limit, if a billing administrator allows it, and that goes on the next invoice. Otherwise, or at the limit, the next AI request is refused with AI credits used up: members are asked to tell a workspace admin, and admins get a link to add credits. Nothing else in the product is affected, and a verified answer asked word for word, or asking for the last quick answer again, still answers: neither needs the AI, and the answer says so. Editors use the included credits first, everyone the bought ones after (which credits a request uses).
Ask AI on a share or an embed is paid for by the workspace that published it, from the credits it bought, and stops at zero.

What the AI never does

Changes your work without your approval. In AI chat it shows an approval card with exactly what will change, and only that change runs, once you approve it. See Approving a change.
Sees data you can’t. It works as the person asking: row-level security filters what its queries return, and fields labelled sensitive stay masked for everyone but company admins, in AI answers too.
Runs for an API key. AI features are available in the DataSquares app only; API keys can’t use them.
Tells you which AI is behind it. It is the DataSquares AI assistant. It is instructed never to name a model or the company behind one, and a sentence in an answer that claims otherwise is replaced before the answer is kept.

Before you rely on an answer

The AI can be wrong. Check the SQL and the Evidence of an answer before you repeat a number, prefer your governed metrics for the numbers your dashboards show, and read what the AI can and cannot do. When the AI refuses or stops, When the AI can’t answer says what each message means and who can fix it.

AI chat — asking, approving changes and conversations.
What the AI can and cannot do — the limits worth knowing.
When the AI can’t answer — every refusal, and what to do.
How the AI uses your data — what it is sent, what DataSquares keeps.
Memory — what the AI remembers about you and your workspace, and how you control it.
Seats, plans & AI credits — allowances, packs and auto top-up.
Workspace settings — the AI settings and tools admins control.