How the AI uses your data

This page says what the DataSquares AI assistant is sent when it answers, what never leaves DataSquares, and what DataSquares keeps afterwards. It describes what the product does today. The legal terms are in the privacy policy and your subscription agreement.

Where a request goes

When you ask the AI something, the app sends your request to DataSquares. DataSquares puts together what the AI needs, under your permissions, and sends it from its own servers to the AI service it uses. Your browser and the phone app never talk to an AI service directly, and nobody picks a model: DataSquares chooses the AI behind each feature.

API keys can’t use AI at all: every AI feature refuses one (Scopes).

What the AI is sent

In AI chat

Your message and the conversation so far: earlier messages as text, and for a long conversation a summary of its oldest part.
The page you asked from, when you ask from one: its name and, for a dashboard, its pages, cards, filters and what you selected. Remove the context chip from the message box to leave it out. A dashboard, report or data source you can’t open is never sent.
Where to look: the names of the workspace’s ten newest data models, up to twenty governed metrics, and up to ten data sources that have no model yet. Where the server turns on the compact model listing, every data model instead, as many as fit, each with its connection’s name and type, its number of tables, its main tables’ names and its description.
Definitions: the glossary terms your message names, with their definitions.
Notes: the notes DataSquares keeps that match your message or the page, yours and the workspace’s; and what the AI looks up in its memory as it works: your notes, the workspace’s shared notes, the notes about the tables of the data models you can open, and your drafts and proposals waiting for review (none of your own while your memory is paused).
Your own good answers: up to three answers you rated thumbs-up, reduced to their approach with the values taken out (none while your memory is paused).
The workspace AI brief: the brief an admin published, whole, as the workspace’s reference data (an admin trying a draft is sent the draft).
Design preferences, when the AI builds a dashboard: yours and the workspace’s defaults, as settings (a theme, a layout, chart and number habits, a title style); and in every chat, how you like numbers shown (your own number display), so its answers write them that way (none while your memory is paused).
What it reads as it works: table and field names and descriptions, a field’s counts and up to ten of its values, and the rows its queries return. Each result is cut to a fixed size before the AI reads it.
Its own reasoning, within one answer. As the AI works through an answer’s steps, it is sent back what it reasoned in that answer’s earlier steps, as it wrote it, so it builds on that instead of working it out again at every step. That reasoning came from the AI service in the first place: nothing more of yours leaves DataSquares with it. A later message is sent the earlier answers’ text, never their reasoning. In a long build, the AI may be sent its earlier steps once more to sum them up, and goes on from that summary and its latest steps.

In the other AI features

Feature	What the AI is sent
Explain with AI, on a card or a metric	The measure, the two points compared and the breakdown’s numbers for each member
Analyze → Phrase it	The statistics on the tab you are on
Summarize with AI	The numbers computed from up to 5,000 of the card’s rows (its change, leaders and what stands out, with their labels), or the card’s first 50 rows when it offers none; read under the editor’s permissions. A preview nobody shows is deleted after a day
Write with AI in a Story, and the SQL editor’s Explain and Generate	An AI chat request, as above
Rewording a Story selection	The selected text, with each live number, mention or date as a placeholder, never its value
Write from the data in a Story	The numbers computed from the section’s live data (its values, changes and leaders, with their labels), read under your permissions; never its rows
Commentary the assistant writes into a report	An AI chat request, as above; the numbers DataSquares computes for the commentary, read under your permissions, come back to it with the approval card. Each send of the report computes them again without the AI
A Story outline	The goal you type
A Story’s grounded proposal	Bounded facts from the section’s checked numbers, not SQL or instructions from source labels (Review numerical claims)
AI layout suggestions	Your request and up to 100 layout settings and column labels; no queries, rows, filters, formulas, source identifiers or delivery settings (details)
A report comparison’s AI summary	Calculated aggregate facts and group labels; no raw rows, SQL, access policies or parameter filters (details)
A theme’s Draft with AI	Your description, or a few colour hints DataSquares takes from a public page you paste; never the page itself (details)
A measure’s Draft with AI	The model’s table and field names, its measures and glossary terms, and your request; no data values
A metric digest’s AI sentences	One pace sentence’s numbers and verdict (details)
A profiling scan’s AI findings	The scan’s findings and statistics, never a column labelled sensitive (details)
Describe the check in Data quality	The sentence you type, and the table’s fields by name and type with the counts of its latest data-quality profile; a field labelled sensitive by its name and type only; never a stored value
Suggest mapping on an activation	The dataset’s column names and types, and the destination keys you type or paste; never a value (a sample you paste is cut down to its keys first)
A lineage explanation in the Catalog	The lineage chain DataSquares builds by rules for the dashboard, card or metric, and nothing else: the names of the items in it you may open, a count of the others, and where a row-level security policy applies, never its rule; never a value
Ask AI for viewers on a share	The viewer’s question, the conversation’s last four questions and answers, and that share’s own cards with the rows the viewer may see (details)
AI search, where the server turns it on	To keep the AI catalog up to date: each model’s, table’s, field’s, relationship’s, measure’s and governed metric’s name, description and glossary synonyms; never a value from your sources, nor a measure’s or metric’s formula, nor the questions you mark as answered well. To search it: the question and, at Balanced and above, a few search phrases a quick AI step writes from it with some of the model’s table names. The few tables it finds are then read as in AI chat, with the value hints the question names (details). The schema finder, where the server turns it on: within the answer, a one-line overview of each table in the AI scope (its name, description, keys and a few column names), the tables it opens, with every value of each short code list in them (a field of up to 20 different values), the stored values the question names, and what its checks of your data found (counts, a date range, the stored spelling of a value), and, where the server turns them on, the notes about your tables it keeps for the tables it opens (yours, and the workspace’s shared facts) and those of earlier questions like yours; DataSquares keeps the names of the tables and columns it gave each answer for 30 days, never a value (details)
Explore this model in Prep for AI, where the server turns it on	What the schema finder is sent, for every table of the model in the AI scope of the person who started it, a few tables at a time: a one-line overview of each table (its name, description, keys and a few column names), the tables it opens with every value of each short code list in them, the notes about those tables that person may use, and what its checks of your data found (counts, whether a key repeats, whether a join matches), run as that person; never the values of a field labelled sensitive they may not read, and no check where row-level security applies to them. What it infers is kept as notes about your tables for the model owner’s review (details)
Draft with AI in Prep for AI	For each table drafted: its name, schema and number of rows, its joins to the other tables in the AI scope, the descriptions it has, and each field’s name and type with how many values a short code list holds and the counts of the table’s latest data-quality profile; never a stored value, and nothing of a field labelled sensitive, or of a field that reads a labelled column, but its name and type (details)
Search by meaning, where the server turns it on	What you type in the Ctrl+K palette; and, to keep its index up to date, each dashboard’s, metric’s, model field’s and saved query’s name and description, a metric’s expression and time grain, a field’s type, table and model, and the first 400 characters of a saved query’s SQL (details)

What never leaves

Connection passwords, keys and tokens. DataSquares connects to your sources itself and gives the AI only what its reads return. An answer in AI chat (Write with AI and the SQL editor’s Explain and Generate included) or in Ask AI on a share that carries something shaped like a password, a key or a token is withheld: you see I’m not able to share that response. instead, and the answer isn’t kept. What the other AI features write, such as card summaries, Story rewording and theme or measure drafts, isn’t checked this way.
Data your permissions hide. The AI works as the person asking: row-level security filters its queries, it refuses to run SQL of its own on a source where a policy applies to you, and a field labelled sensitive stays masked for everyone but company admins.
Other people’s conversations. The AI reads the conversation you are in, the topics of your own earlier chats in the workspace when it needs them, your own notes and good answers, and notes shared with the whole workspace. In an incognito chat, or while your memory is paused, it looks only at the chat you are in.
Drafts. Notes are made from published Stories only, never from drafts.

What DataSquares keeps

Conversations

Each conversation is kept with its answers, SQL and charts until it is deleted. A workspace admin can have conversations deleted 30, 90, 180, 365 or 730 days after their last activity (Keep conversations under Administration → AI, the AI settings); the default, Until deleted, keeps them.
The AI’s reasoning on an answer is kept, encrypted, with the answer’s steps so far, only while the answer is being made, so that an answer interrupted by a restart can go on. It is deleted when the answer ends (one left behind by a failure, within seven days) and is never stored with the conversation. The same goes for a long build’s summary of its earlier steps.
A conversation is private to the person who had it: nobody else in your organization can open it, administrators included. When the AI hands a conversation to a person (AI escalations), the people who may handle escalations (company admins, and a role given the AI escalations permission) see it in a queue: who asked (their name and email), the priority and the reason the AI wrote, and when it was opened, but not the conversation. One of them can answer: the answer appears in the conversation under their name, as from a person and never as the AI’s, and the AI doesn’t read it. It is kept with the conversation and deleted with it.
You can rename, pin, export (Markdown or JSON) and delete your conversations. Deleting one removes its messages.

Notes the AI keeps

The AI keeps short notes so that later answers fit your workspace. They live in your workspace’s DataSquares database and stay in that workspace.

Notes you ask for. Tell the AI to remember a preference, a definition or a rule, and it writes it in its memory as a note of yours, for your own chats.
Notes it learns from you. It also keeps what you type that reads like a preference, a definition or a rule (“by revenue I mean net revenue”, “always use the fiscal year”), and a correction you type with a thumbs-down. These are yours alone at first. It doesn’t announce them.
Answers you rated well. A thumbs-up on an answer that used your data keeps the question and how it was answered, with the values taken out, for your later questions.
Topic summaries. When a conversation moves on to a new topic, a short summary of the old one is kept for you, with numbers and quotes taken out.
Notes shared with the workspace. A note reaches other people only once a workspace admin, or the data model’s owner for a model’s notes, accepts it: a correction you type, the same note several people keep word for word, and a note you asked the AI to share, once you have sent it for review yourself (how a note reaches other people). Until then, that note is your private draft: nobody else sees it, admins included.
Notes from what people can already see. Each night, DataSquares writes short notes about the workspace’s dashboards, data models, governed metrics, glossary terms and published Stories: their names, descriptions and fields, and for a Story its summary and the opening sentence of each section. A glossary term’s note is also rewritten as soon as the term changes, and retired when it is deleted. A note about a dashboard, a model or a Story reaches only people who can open it.

What a note can hold

One rule decides what a note can hold, whichever way the note is made: a note you ask for, a note it learns from you, a correction you type, an answer you rated well, a topic summary, a note shared with the workspace, and each name, description, definition and sentence in a nightly note. The rule turns away text that holds:

The rule turns away	For example
An email address	jane.doe@example.com
A phone number	+44 20 7946 0958
A long number: five digits or more in a row, or a number with thousands separators	88123456, 1,204,331
An amount of money	$5k, 250 EUR
A link	https://…, s3://…, www.example.com
A site address with one of the endings listed below that has three parts or more, or a path after it (with a path, .co too); with .dev, .app or a country ending, three parts only as described below	docs.example.com, portal.example.in, example.de/preise, example.co/pricing
A site address ending in .co, .com, .org, .net, .gov, .edu or .ac and a two-letter country	example.co.uk, example.co.in, example.ac.in
A file share	`\\server\share`
An IP address	192.0.2.10, 2001:db8::1
Something that looks like a password, a key or a token	the password is …, api_key: …, db_password=…, clientSecret: …, a cloud access key, a private key, a database address with a password in it
Something that reads like an instruction to the AI	ignore previous instructions, you must call …

It keeps what the AI needs to know your workspace: a preference, a definition or a rule in your own words; field, table and join names, such as orders.customer_id → customers.id; and a metric’s definition, such as SUM(net_amount) - SUM(refund_amount). A name that only mentions one of these things is kept: a field called api_key or ip_address, a field described or filtered on, such as password_reset_at: when the password was last reset, or a name such as Password issues by week. A dotted name in two parts, such as users.org, could be a field, so it is kept, and so is a company written as its site, such as Amazon.com. In a formula a dotted name is a field: one inside a function’s brackets, such as SUM(finance.invoices.net), and one divided by a field, a number or a function, such as invoices.net/invoices.gross, are kept. Outside a function’s brackets, a three-part name such as finance.invoices.net reads as a site address and is turned away.

The site endings the rule recognises are .com, .net, .org, .edu, .gov, .mil, .io and .ai; and .dev, .app and the country endings .in, .de, .uk, .eu, .fr, .nl, .es, .jp, .cn, .sg, .au, .nz, .ae, .ie, .br, .mx and .za. The second group are also the names of fields (a language column such as cms.pages.es, a field such as inventory.movements.in or returns.std.dev), so a three-part name ending in one of them is read as a site address only when one of its parts has a hyphen (files.example-corp.de), its first part names a site, such as www, portal, docs, help, status, shop, wiki, intranet, mail, api, files or support (portal.example.in, status.acme.dev), or a path or a port follows it (ops.acme.de/runbooks, grafana.acme.de:3000). Otherwise it is read as a field and kept, and so is one followed by a bracket, such as orders.status.in(‘open’). An ending that is also the name of a field or a word, such as .id, .no, .it or .at, is always read as a field, so sales.orders.id is kept. A site address with an ending not listed, such as files.example.it or example.ch/preise, may not be recognised, and nor may a three-part one in the second group that matches none of the above, such as sales.acme.de, one written inside a function’s brackets, a two-part one with a port, such as example.com:8443/admin, or one whose path is only a number, such as example.com/2024, so don’t ask the AI to remember one.

The rule reads a note as a person reads it: a word written in full-width letters, such as ｈｔｔｐｓ://, or with an invisible character inside it (a zero-width space, a joiner, a direction mark, a variation selector, a filler or any other character drawn as nothing), such as a zero-width space in password, is read as the plain word, so neither gets text past the rule. Text hidden in invisible characters is turned away whole, whatever the visible text says: text written in tag characters, and variation selectors where written text never has them (two in a row, or one after a Latin letter or a space, or at the start). An emoji, a keycap, a flag such as Scotland’s and an ideograph’s variant use them as written text does, and are kept. What is kept is the text as written.

The workspace AI brief an admin writes has a rule of its own: it may hold amounts, thresholds and dates, but not an email address, a phone number, a link, an IP address, something that looks like a password, a key or a token, an instruction to the AI, or hidden text. A brief with any of these, or longer than 10,000 characters, is refused whole, with the reason, and never cut.

What happens to text the rule turns away depends on the note:

Note	When the rule turns text away
A note you ask for	It isn’t saved, and the AI is told why.
A note it learns from you, a correction you type	It isn’t kept.
A topic summary	The sentence or name is left out, and nothing is kept when it is the topic’s title. A topic summary also leaves out every sentence with a number or a quotation in it.
An answer you rated well	Nothing is kept when it is the question. How it was answered is cut to a sentence or two with no numbers, quotation marks or names in them, and the shape of its SQL with the values taken out.
A note shared with the workspace	Only a note whose title and text the rule keeps is shared.
A nightly note	The name, description or sentence is left out. A table, a field or a join whose name is turned away isn’t listed. A metric’s definition is copied whole or left out, never cut, and it is also left out when it is too long for the note. A metric whose name is turned away, or a glossary term whose term or definition is, has no note, and the note it had is retired; a dashboard, a data model or a Story whose name is turned away is noted as untitled.
A note about your tables	That piece (a grain, a filter, a gotcha, a code’s meaning, the question) is left out whole; a note you rewrite isn’t saved, and you’re told why.
A note of the AI’s memory kept before the rule read it as it does now (October 2026)	The notes kept before are read with the rule once more, and each one it turns away is retired: the note no longer applies and is shown nowhere (not under Settings → Memory, in a review, in the list of what you proposed or under an answer). Its text isn’t changed. A nightly note comes back the next night without the piece the rule turns away. This covers the AI’s memory notes, not the notes about your tables.

Once a day, a note of yours that nothing has read, written or changed for 180 days is retired, except while your memory is paused. Each person keeps at most 200 notes: past that, the AI writes no new one until some are deleted or combined, and each night the least used past 200 are retired. Settings → Memory lists, searches, edits and deletes your notes, and its Table notes tab the notes about your tables. Ask the AI what it remembers (“what do you remember about revenue?”) and it looks in its memory: it can list all your notes and read any one whole.

How the AI works with its memory

The AI reads and writes its notes itself as it works (where the AI keeps it):

Your notes: it can list them, read one whole, and write, change, rename or delete them. A note it deletes is forgotten: it no longer applies, and its text is kept as a forgotten note’s is (removing what is kept).
Shared notes and notes about tables: it only reads the workspace’s shared notes and the notes about the tables of the data models you can open. It writes for the workspace only when you ask it to share a note: the note, or a change to a shared note, is kept as your private draft, which nobody else sees, admins included, until you send it for review under Settings → Memory → Proposals. It applies to no one until it is accepted, and a change replaces the note only then.
Its own work: what it keeps about how far it got, the state of a task or anything about the conversation goes in your own notes only.
Your earlier chats: it recalls the topic of one when it needs it.
Nobody else’s: it never reads anyone else’s notes, drafts, topics or proposals. In an incognito chat it writes nothing, though it still forgets a note you ask it to; while your memory is paused it can’t open your own notes and writes nothing.
Every write is checked by the rule above and against the rows the chat read (what the AI never writes in a note), and refused whole, with the reason. DataSquares records each write, without its text.

With the memory off

An operator can turn the memory off for a whole server. Then the AI has no memory to look in or write to, there is no Settings → Memory page, and no screen lists your notes (notes about your tables are under Settings → AI notes); there are no incognito chats, no AI brief, no design preferences and no review. The notes it learns from you, your rated answers, topic summaries and the nightly notes are kept as above, and a note three people in a workspace hold word for word (a preference, a definition, a fact or a caveat) becomes a note for everyone in that workspace, with no review, until fewer than three hold it. Rules, rated answers and topic summaries are never shared this way.

Besides the notes

For the memory, DataSquares also keeps, in its database, in your workspace:

Your drafts: the notes the AI kept for you to share, seen by you alone until you send them for review or delete them. They are erased with your AI data.
Who reviewed a shared note, and when; for a note waiting for review, the people who stand behind it, counted on the review screens but never named there; and for a change to a shared note, which note it would replace, which the review shows beside it.
Fingerprints of what each chat read: for each text value in the rows a chat’s queries returned, and each value it looked up in a field, a fingerprint that is never the value itself, so the AI can be stopped from writing those values into a note (what the AI never writes in a note). They are kept with the conversation and deleted with it.
The workspace AI brief: its draft and every version that was published, with who changed or published each and when, until the workspace is deleted.
Design preferences: yours, kept only when you say so, and the workspace’s defaults with its theme lock.
Whether your memory is paused, for each workspace.
Incognito chats, only until a day after their last message (looked for every hour); nothing is learned from them, and they are in no list, search or other chat’s recall.
Each answer’s receipts: kept with your message, the notes, glossary terms and rated answers its turn was given; with the governed metrics the answer worked out, read from the conversation, until the conversation is deleted.

Ratings and reports

A thumbs rating, with the reason and the note you give, is kept with the conversation, and also with a copy of the question and the answer you rated: that copy is what lets the AI follow your well-rated answers later (it is sent only their approach, with the values taken out). Company admins see how many answers were rated up or down, the reasons, and the corrections people wrote, but not the conversations. A report of an AI answer or summary is kept with the text you reported, for review.

Records of AI use

Usage: each AI request is recorded with the person who made it, when there is one, its feature, whether it ended in an error, how long it took and the credits it used. Usage by feature is on Administration → Usage, and each debit is in the credit activity of AI credits on Administration → Billing.
The AI ledger: company admins can list each AI call of a month, with its feature, the person it ran for, how it ended and the credits it was charged, under Administration → AI, and export it with each person’s email on Business (read the AI ledger).
The AI audit log: for company admins, through GET /api/ai/audit, who had the AI run which tool and when, what it touched and whether it succeeded. What the tool read and wrote is kept only as a fingerprint, not as text.

The AI catalog and value hints

Where the server turns on AI search, the AI catalog made from the text above is kept by the AI service DataSquares uses, apart for each workspace, and deleted the same day the workspace or its company is deleted, or within minutes of an admin turning AI search off. Value hints, the most frequent stored values of text fields with how many rows hold each, and the key values joins are matched from are kept in DataSquares’ database only, never in the catalog, and are deleted as soon as an admin turns value hints off.

The schema finder’s record

Where the server turns on the schema finder, DataSquares keeps a record of what it gave each answer, the names of the tables and columns and never a value, for 30 days or until the conversation is deleted, to measure how well it finds what questions need.

Notes about your tables

Where the server turns them on with the schema finder, DataSquares keeps short notes about the tables of your data models in its database, in your workspace: what one row of a table means, its key and date column, the filters it usually needs, what a column’s codes mean, whether a join holds or repeats rows, and questions like yours (their values taken out) with the tables that answered them.

Only after an answer you were happy with: a thumbs-up, the answer saved to a dashboard or a report, or your next question asked without correcting it. What an answer would teach is otherwise deleted after 7 days.
No values from your data, except what a code means for the codes value hints keep, and nothing about what your AI scope or labels hide from the person reading. Each note passes the rule for notes, or is not kept.
Yours alone at first; the facts a check of your data proved (a unique key, a join that repeats rows or holds) are shared with your workspace at once, with no value in them.
The model’s owner reviews them. The data model’s owner and workspace admins see the notes waiting for review in Prep for AI: what each says, its table, and whether it was learned from an answer, found by Explore this model or written by a person, never whose answer it came from. A note they approve is the workspace’s; one they reject, or remove once shared, or replace with a newer note about the same thing, is no longer used, and deleted 30 days later. You see what they decided under Settings → AI notes.
Explore this model keeps the notes it infers as the notes of the person who started it, for the owner’s review; the record of each exploration (when, by whom, how many tables, notes and credits, never a note’s text) is kept with the model.
Each answer records the notes it used, for the list under the answer, until the conversation or the note is deleted.
How long: a note goes back in doubt when its table changes, is retired after 180 days unused (never your own words) and deleted 30 days after it is retired; removing a table or a field removes its notes.
See and delete yours under Settings → AI notes. They are erased with your AI data; the workspace’s shared facts stay, without your name.

Presentations: Draft from data

Draft from data sends the AI the figures DataSquares computed for you (each figure’s name, its value as you see it, its period and its change), the title of the dashboard or story it starts from, the names of the dashboard’s charts it may place, and the words you gave it (who the deck is for, your question, the outline you approved, its title included), always as data, never as instructions. Drafting from a story, a figure’s name is story text: its section’s title and up to six words of the story just before the live number. It never sends rows. The outline is not stored: only the slides you insert are kept, in the deck, with every figure stored as its binding rather than its value and each text box marked as AI-written.

Presentations: headlines and speaker notes

Suggest a headline and Write speaker notes send the AI the figures DataSquares computed for you from the slide (each figure’s name, its value as you see it, its period and its change) and the words of the slide’s title, always as data, never as instructions. They never send rows. What the AI wrote is not put on the slide until you use it; the deck keeps a fingerprint of its words (never the words themselves) so that, once used, the headline or the notes stay marked as AI-written until you say you have rewritten them. Only what you use is kept, in the deck, with every figure stored as its binding rather than its value.

Description drafts

A description the AI drafted in Prep for AI is kept in DataSquares’ database until an editor approves or rejects it, then 90 days. An approved one becomes part of the model, like a description typed by hand.

A dashboard’s or a metric’s description drafted with Describe with AI is kept the same way, and an approved one becomes the asset’s description. The AI is sent what the asset is made of (names, card titles, formulas without their quoted text, the descriptions its fields and metrics already have), never its data, a filter’s values or a labelled field’s description.

The search index

Where the server turns on search by meaning, an index of the workspace’s dashboards, metrics, model fields and saved queries, made from the text above, is kept in DataSquares’ database. An asset’s entry is removed within minutes of the asset being deleted.

Removing what is kept

Forget. Tell the AI to forget something and the note stops being used, but DataSquares keeps its text until your AI data is erased, as it does for a note retired after 180 days unused.
The memory page. Delete under Settings → Memory removes a note, or a draft, for good, Reset removes your design preferences, and Withdraw takes back a proposal that still waits. An incognito chat is deleted a day after its last message.
Deleting a conversation removes its messages and the fingerprints of what it read, but not the notes, the copies kept with your ratings or the topic summaries made from it.
Notes about your tables. Delete one under Settings → Memory → Table notes, and the AI stops using it at once.
Deactivating someone on the Users page also erases that person’s own notes, in every workspace. If that erase fails, the deactivation still goes ahead and the notes stay until the person’s data is erased (below). Suspending or removing someone through SCIM, or setting them inactive through the users API, keeps their notes.
Erasing everything. To have your conversations, ratings and notes erased, ask for your data to be deleted as the privacy policy describes. Erasing your AI data also erases your design preferences, your memory pause, your drafts and your support of proposals: a proposal only you stood behind is deleted, and your name is taken off the notes you reviewed. The record of AI use and the AI audit log are kept: they hold your user id, but no text from your conversations.

Training

DataSquares does not train or fine-tune AI models. What the AI learns from your workspace is the notes and rated answers above, kept in your workspace’s database. The privacy policy also states that DataSquares does not use your connected business data to train machine-learning models.

AI in DataSquares — where the AI appears, effort and AI credits.
Memory — your memory page, reviewed notes, incognito chats, the AI brief and design preferences.
What the AI can and cannot do — what it works from, and its limits.
Field sensitivity & masking — labelled fields, in AI answers too.
Workspace settings — AI settings, tools and escalations.