Dashboards · Trust the numbers

Every number on the board comes from your data.

Mammoth uses AI to design a dashboard. It never uses AI to write the figures on it. Each number is computed by a query on your data, checked before it is shown, and traceable back to the rows it came from.

Try free Try the live example

Live example

Built from one sentence. Every figure computed.

A real Mammoth dashboard, live on this page. Click a bar or a slice to filter the whole board: every card, chart and total recomputes from the data.

Sample data for a fictional coffee chain. Made from one sentence: “revenue, gross margin %, orders and units by region, store, channel, category and month, with key insights”. Gross margin is total margin divided by total revenue, not an average of per-order percentages.

Open the full board

How a number gets onto the board

The AI designs. The database computes.

Three things happen between your question and a figure on the screen, and none of them lets a language model invent a value.

The AI never writes a number

The AI picks fields and designs; every number is computed by a database query compiled from those choices.

No made-up columns

Every field the AI names is checked against your real columns, and a name that doesn't exist is dropped and reported.

Words checked against the numbers

Generated summary lines are checked against computed facts, and a line citing a figure the data doesn't contain is dropped.

Checks before anything is shown

What Mammoth refuses to put on a board

A dashboard that looks right and is wrong is worse than no dashboard. These run on every board, whoever builds it.

Chart titles match what they plot

If a chart's title names a measure the chart doesn't draw, the title is replaced rather than shipped.

Refusals based on your data

Q&A declines a question only when it can't be answered from your fields, not because of how it was worded.

Comparison pages without holes

A side-by-side comparison page only carries charts that can be drawn side by side; others are declined, or hidden and never deleted.

A computed head-to-head

A comparison opens on two real subjects and states the gap in computed numbers; a gap under 2% is called close.

Balances are not added up over time

Balances like cash, headcount or inventory on a snapshot table show the latest period instead of being summed across months.

Sample data is labelled

A board on sample data says so in a ribbon or chip, and boards on sample data can't be published.

Check it yourself

Every figure can be traced back to its rows

You don't have to take the board's word for anything.

How this was made

A chart's menu shows which columns it uses and how, the filters on it, and for editors the exact SQL.

The rows add up to the number

When you open the rows behind a number, their count and total equal the number you clicked.

Drilled detail adds up

When you drill into a mark, the groups underneath sum to the value you clicked.

Assumptions in one place

Editors see the judgement calls behind a board in one place, such as an assumed currency, a total inflated by a join, or charts that couldn't be built.

Limits

Limits worth knowing

Better to know before you choose than after.

Read the commentary before it goes out: The check confirms that a figure a sentence cites exists in the computed results, not that the sentence attaches it to the right thing, so read generated commentary before it goes out. Text blocks carry no data, so they can never hold a computed figure.
Receipts don't cover every tile yet: Shown on added charts and most generated base charts; some combined base charts have none. SQL is editor-only and hidden while previewing as another viewer. It does not report row counts used or dropped.
Rows behind a number: up to 500, signed in: Up to 500 rows, for signed-in readers, within their row-level access. Contact-like columns are left out. The author can turn it off for a chart.
Q&A won't forecast or advise: No medians, forecasts or regressions in Q&A answers, and no investment-style advice.
A published board can lag a moment: A published board can briefly show the previous version after an edit or new data.
The AI can pick a measure you didn't mean: An average where you wanted a total is still computed correctly from your data, and still the wrong number for your question. The receipt shows what each chart computes, so review a board before you share it and ask in plain words to change it.
Clean data is the pipeline's job: The board shows what the data says. Fixing the data happens in the data engine underneath, as recorded steps that re-run on new data.

Checked against the product on 2026-10-23. Every dashboard capability, with its plans and limits, is also published as machine-readable JSON.

Questions people ask first

Can the AI make up a number?

No. The AI decides what to show and how to lay it out. Every figure on the board is the result of a query that runs on your data, and summary sentences are checked against those results before they appear.

So is every board right?

Every number is computed from your data, but the AI can still choose a measure you didn't mean, such as an average where you wanted a total. Each chart's receipt shows exactly what it computes, so you can check before you share, and ask in plain words to change it.

Can I see where a number came from?

Yes. Open a chart's receipt to see its columns, calculation and filters; editors can see the SQL. Signed-in readers can click a number to see the rows behind it, and those rows add up to the number.

What happens if my data is wrong?

The board shows what the data says. Cleaning and checking the data happens upstream, in the pipeline that feeds the board, where each step is recorded and re-runs on new data.

What if a refresh fails?

Readers keep seeing the last good version, flagged as out of date, instead of a broken board.