What the AI can and cannot do

The DataSquares AI assistant answers from your data, but it can still be wrong, and some limits decide how good an answer is. This page lists them as they are today, with what to do about each.

Check an answer before you use it

An answer can be wrong even when it reads well. Before you repeat a number, open its SQL and its Evidence: the sources and metrics it used, the filters and time range, whether the data was live or cached, and what it assumed. Every chart says how much of the data it covers, or why it is partial.
Governed metrics match your dashboards; the AI’s own queries may not. When a governed metric covers the question, the AI is told to answer with it, and the number then matches every dashboard that shows the metric: the AI gives the number the metric’s card shows, its latest whole period, and names that period, unless you ask for another (last quarter, this month so far) or for an all-time total. A number from a query the AI wrote on your tables can differ (another filter, another join, another date column), so name the metric for the numbers that matter, or check the SQL.
The same question can get a different answer. Asked twice, the AI can word the answer differently, and sometimes take another route to the numbers.
It says what the numbers show, not why. Explain this change shows which members drove a change and by how much; the AI doesn’t know the reasons behind it (a promotion, an outage), and it is told not to guess them.

What it works from

Your data models first. Each conversation starts with the names of the workspace’s ten newest data models and up to twenty governed metrics, the ones you follow first, then the most used, and the AI finds others by looking them up. Where the server turns on the compact model listing, it starts instead with a line for every data model, the most used first, as many as fit, and is told when there are more. Name the model, the metric or the dashboard you mean.
Large models. The AI reads a whole model at once up to about 100 to 120 fields, or about 80 to 90 when tables and fields have descriptions. On a larger model, which can be one of a dozen tables, it first gets the fields of the first tables, about as many in all, and the names of a few more tables at most, and finds the rest by asking for tables by name or by searching. With the compact model listing, it reads a model whole up to the size the answer’s effort reads at once, and a larger one most used tables first (How the assistant reads a model). Name the table or field when you know it, and give tables and fields clear descriptions.
Your glossary, within limits. A question gets the definitions of at most eight glossary terms it names, the longest first. A question names a term by its name or one of its synonyms, as whole words, its simple plural or singular too: active customers names Active customer, and carry doesn’t name ARR. A term kept with text hidden in invisible characters is never used.
English words for building. The assistant recognises a request to build or change a dashboard or a report from English words such as build, add, dashboard and report. It answers questions in other languages too, but reads a request in another language as a question first: if it doesn’t offer to build, name the job in English.
No internet. The AI works only with what DataSquares gives it from your workspace. The one exception is a theme draft from a brand page you paste: DataSquares fetches the public page itself and passes the AI a few colour hints from it (Brand generator).

Building with AI chat

Nothing is written until you approve it, and an approval card expires ten minutes after it was made (Approving a change).
The card shows what will change, not the data. It names the cards, blocks or steps and their settings, not the numbers they will show. After you apply a change, open the dashboard or report and check it.
One approval has a size. A finished dashboard brings up to six KPI cards, four trends, six breakdowns and four filter cards in one approval; a report up to twelve blocks. Ask for more in a follow-up.
The check after applying is about the change. DataSquares re-opens the item and confirms the change landed and nothing else moved; it does not check that every new card shows the numbers you expected.
An applied change is not undone from the chat. Edit the dashboard or report to change it back.
A long job can stop before it finishes. One answer can take only so many steps: a request that needs more gets an answer from what was found so far, which says what it could not finish. An answer also stops after two minutes. Ask the AI to continue, or split the request into smaller ones. On a server that turns plans on, the number of rounds depends on the effort, each step has a time limit of its own instead of the two minutes, and Answer now asks for the answer from what was found at any point (While it works).
Plans show on the web only, for now. The DataSquares Android app does not show a question’s plan, Answer now or the build plan card. A plan still runs for a question asked on a phone; while a build plan waits for you, the reply on the phone says to open the conversation on the web.
A very long conversation can become too long for the AI to read in one go: start a new conversation for a new topic.

Text the AI writes

Its words don’t update themselves. Text the AI writes into a report or a Story is the wording of the day it was written. Its numbers can: commentary the assistant writes into a report from computed facts has its numbers computed again every time the report is sent, and in a Story the numbers it quotes become live numbers when exactly one value of the result matches (every number, with Write from the data). Other numbers stay plain text, and the preview names them before you accept. A word such as rose stays as written: in a Story, the section says when its numbers have moved, and Reword with AI suggests new words.
Card summaries are dated and say how many of the card’s rows they were written from: up to 5,000, or the first 50 for a card that offers no computed number, with the rest left out. Summarize again when the data has moved on (AI card summaries).
Rewording a selection in a Story keeps its live numbers, mentions and dates where they were. When the rewrite removes or adds a number you typed, you see which before anything is replaced. A selection with a chart, an image or another block in it can’t be reworded: select the text around it (Write with AI).
Generated SQL in the SQL editor is a draft: read it before you run it.
Measure and theme drafts are checked before you can use them. A measure draft is offered only once the SquareX analyzer accepts it, and a theme draft a reader could not make out cannot be saved.

Ask AI on shares and embeds

Ask AI for viewers answers only from the cards of that one share, keeps a conversation for 30 minutes after its last answer, and answers a limited number of questions each hour for the whole share (Ask AI on a share).

On the phone

The Android app has AI Chat without the effort choice: its questions run at the effort you chose on the web, saved to your profile, or at your company’s default until you choose one (On mobile).

What it can never do

See data your permissions hide. It works as the person asking. Row-level security filters its queries, and it refuses to run SQL of its own on a source where a row-level security policy applies to you (with the new tool set it answers from your data model instead, filtered to your rows). A field labelled sensitive stays masked for everyone but company admins.
Change something without your approval in AI chat.
Be used with an API key. AI features are available in the DataSquares app only (API keys can’t use them).
Carry notes between workspaces. What it keeps for you belongs to one workspace (What DataSquares keeps).

AI in DataSquares — where the AI appears, effort and AI credits.
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.
AI chat — the full guide to asking and approving.