Card insights: explain, analyze, summarize

Chart and table cards built on a data model carry insight tools in their Card actions (the three dots that appear when you point at the card or tab to it, and that a touch screen always shows) and in their right-click menu, which offers the same list. Summarize with AI is also offered on KPI cards, to people who can edit the dashboard. Explain this change and Analyze are statistical — the math runs on the server, uses no AI and spends no AI credits. AI writes only when you choose it: Explain with AI, Analyze’s Phrase it and Summarize with AI each add AI-written text, need AI set up, and use AI credits, as each one says. View access is enough for Explain this change and Analyze; Summarize with AI needs edit access to the dashboard, because the summary you show goes on the card for everyone. All three run under your data permissions: when a row-level security policy on the card’s source applies to you, they refuse and say so.

Explain this change

When a number moved and you want to know why, open the card’s Card actions (or right-click the card) and choose Explain this change.

Choose the Point to explain; it is compared with the point just before it. When you choose the first point, pick another point as the baseline.
In Break down by, pick the dimension to break the change down by — region, product, channel, whatever the card’s model offers.
Choose Explain. You get the overall delta decomposed by that dimension: which members drove the change up, which pulled it down, and by how much. This is statistics only: no AI runs and no AI credits are spent.
Or choose Explain with AI for the same numbers plus a short written summary of them, labelled AI-written with the date. It uses AI credits, as the panel says. When the workspace’s AI controls leave the summary out (AI or Explain and Analyze turned off, your role, a used-up budget), the panel says No AI summary: and why, as Analyze does. If the AI writes nothing (or AI isn’t set up on your server), it says “No AI summary this time”. Either way the numbers stand on their own. A metric’s Explain with AI works the same way.

It’s a fast first answer to “what happened here?” before you drill into the underlying data.

Analyze

Analyze in the same Card actions opens a panel with three tabs. All three are deterministic statistics on the card’s own query results.

At the foot of the panel, Ask about this card takes any other question (“which region is furthest below target?”): it is answered in the AI chat panel, with the card and the filters it is showing as the question’s context. That part needs AI configured.

Anomalies

Flags points that deviate from the series’ expected range. A sensitivity slider (2–4σ) controls how strict the detection is — lower values flag more points, higher values only the clear outliers. Each flagged point shows how far it deviated.

Key influencers

Pick up to 8 breakdown dimensions and the tool ranks them by how much of the measure’s variance each one explains — a quick read on which dimension actually matters for the number you’re looking at.

Forecast

Projects the series 3, 6, or 12 periods ahead with a 95% confidence band, so you see both the expected path and the uncertainty around it.

Phrase it

Each tab has an optional Phrase it toggle, off until you tick it, that asks the AI to write a short narrative of the findings; it uses AI credits, as its label says. This is the only part of Analyze that touches AI — the anomaly detection, influencer ranking, and forecast all work without it. The narrative is labelled AI-written with the date, so it is never read as one of the computed numbers, and Analyze shows the AI sparkle only while Phrase it is ticked. If AI isn’t configured on your server, no narrative comes back; the numbers still do.

The Android app gives the numbers without an AI summary: Explain and Analyze on a card, and the explanation of a metric or an alert, run the statistics only and spend no AI credits. An AI summary in the web app has Report this summary: choose a reason, add a note if you like, and send it with the text you saw. DataSquares reviews it on the hosted service; on a self-hosted server, the people who run it do.

AI card summaries

Summarize with AI on a card’s menu writes a short summary of what the card shows. It’s offered on chart, table and KPI cards built on a data model, and the menu item says it uses AI credits.

Its numbers are computed, not guessed. The summary reads every row the card shows, up to 5,000, under your permissions, and DataSquares works out the numbers it may quote: the change on the period before, the leaders and the last, what stands out from the rest. The AI writes only the words around them. A card that offers no such number (a table, a scatter) gives the AI its first 50 rows to describe instead. A card with more rows than were read offers no computed number either: leaders or outliers of part of the rows would not be the card’s. A card that shows only its top rows by its own row limit is summarized from the rows it shows: its leaders and their values, never a total, a share of the whole, a last place or what stands out from data the card leaves out.
You see a preview first, and only you. It shows the summary labelled AI-written, when it was written and how many of the card’s rows it was written from — and, when the card has more (over 5,000, or over 50 for a card that offers no computed number), that the rest were left out — and any figure the AI typed that no number computed from the card matches, so you can check it. A figure typed with a sign, minus or plus, is checked with its sign. Choose Show to everyone to put exactly that summary on the card, Summarize again for a new one, or Cancel to leave the card as it was. A preview more than a day old, or of a card that has changed since, can’t be shown: summarize again. Each summary uses AI credits, shown or not.
On the card, it sits below the chart, never over it: the chart gives up the room. It is labelled AI-written with its date and how many rows it read. Its ✕ hides it for you only.
Summarize again whenever the data has moved on, or choose Remove AI summary to delete it from the card.

Summaries require AI to be configured on the server by your administrator, and they use the workspace’s AI credits.

Who sees a summary

A summary is written from the card’s data with no row-level security narrowing it (Summarize refuses an editor a policy applies to), so it isn’t shown where someone would see less of that data:

Row-level security. Anyone a row-level security policy on the card’s source applies to, or whose policies can’t be checked, sees “This summary was written from data you can’t see in full, so it is hidden for you.” in its place. The dashboard’s version history, version comparison and export leave it out for them too.
Sensitivity labels. Once a label is added to or removed from a column the card reads (in its model, in another model of the same source, or in a Warehouse copy of its table), anyone who sees labelled values masked (everyone but an admin) sees “This summary was written against different sensitivity labels than apply now, so it is hidden for you. An editor can summarize the card again.” in its place, until an editor does (field sensitivity).
Shared links and embeds never show a card’s summary.

A chart saved from an AI answer as a snapshot is hidden the same way, from anyone a policy anywhere in the workspace applies to: see From answer to dashboard.

What needs AI — and what doesn’t

Tool	Works without AI	AI credits
Explain this change → Explain	✓ — pure statistics	none
Explain this change → Explain with AI	✗ — needs AI configured	uses AI credits
Analyze (anomalies, influencers, forecast)	✓ — pure statistics	none
Analyze → Phrase it	✗ — needs AI configured	uses AI credits
Summarize with AI	✗ — needs AI configured	uses AI credits

AI never runs on its own here: only the three AI choices above call it, and only when you pick them.

Either way, nothing escapes your permissions: every computation and every narrative is based only on data you can already see.

Exploring dashboards — the rest of the viewer toolkit.
AI chat — the same anomaly, influencer, and forecast tools, conversationally.
Alert rules — get notified when the number crosses a threshold.