AI dashboard: what it is, with examples of what it adds to an ordinary dashboard

An AI dashboard is a dashboard that adds three things to its charts: plain-language questions answered from the same data, automatic flags on the measures that moved beyond their normal range, and a written explanation of what changed. The flags and figures should come from statistical rules computed on your data, such as a z-score against the last twelve months. The AI model writes the explanation from those computed figures.

What makes a dashboard an AI dashboard

Questions. A reader types "which region slowed?" and gets an answer computed from the same data and definitions as the tiles.

Flags. Each measure is compared with its own history, and the ones outside their normal range are marked, so the reader starts with what moved.

Explanations. A short written paragraph says what changed and what drove it, with every figure traceable to the data.

BI vendors have added versions of these. Microsoft's anomaly detection for Power BI line charts marks values outside an expected range and gives "a natural language explanation of the anomaly." Tableau's Explain Data "builds statistical models and proposes possible explanations for individual marks in a viz." Product names identify the products only; Covirage is not affiliated with Microsoft or Salesforce.

Three examples

Finance month-end. Gross margin, operating costs by line and receivables. The flag that matters is usually the quiet one: days sales outstanding creeping up while revenue looks fine.

Weekly sales. Orders, pipeline coverage and dormant accounts. Questions do the most work here: "which reps are below last quarter's run rate?"

Customer service. Ticket volumes, repeat contacts and time to resolve. A flag on repeat contacts often points to a product or billing problem before the complaints reach sales.

Worked: six measures, which moved beyond normal

A monthly finance and sales page. Each measure is compared with the mean and standard deviation of its last twelve months.

Two measures are flagged: average order value and days sales outstanding. Revenue, margin and orders all moved, but within their normal range.

The check: 4,120 orders × 446 average order value = 1,837,520, consistent with revenue of 1,840 thousand (average order value is rounded to the dollar). If orders times order value did not land near revenue, one of the three tiles would be using a different definition.

The explanation: fewer but larger orders kept revenue up, and customers are taking five days longer to pay.

How the flag is computed

z = (this month − 12-month mean) / 12-month standard deviation

Flag when |z| ≥ 2

With the last twelve months in C2:N2 and this month in B2, the z-score in O2 and the flag in P2 are:

=(B2-AVERAGE(C2:N2))/STDEV.S(C2:N2)
=IF(ABS(O2)>=2,"Flag","")

A threshold of 2 is deliberately more sensitive than the three-sigma limits that the NIST/SEMATECH e-Handbook calls "an accepted standard in industry" for control charts: a monthly page should prompt a look, not only an alarm. The threshold is a choice, and setting thresholds so the weekly sales digest is not noise covers how to set it.

The cost of that sensitivity is false flags. If the measures were independent and roughly normal, each would stay inside ±2 by chance 95.45% of the time, so:

Chance of at least one false flag across k measures = 1 − 0.9545^k

For six measures, 1 − 0.9545^6 = 24.4%: about one month in four shows a flag by chance alone. Real measures are correlated (revenue, orders and order value move together), so treat this as a rough guide. Seasonal measures need one more rule: compare them with the same month last year, or every December is flagged.

What the AI model writes, and what it must cite

The AI model writes from the computed figures only. Every number in the explanation is one the tools produced, with its source: "average order value 446 against a twelve-month mean of 410 (z = 2.57), from 4,120 orders." It may connect flags ("fewer but larger orders") and suggest a question to ask next. It may not introduce a figure the tools did not compute, round a figure into a different one, or explain a measure that was not flagged as if it were. What makes a change an insight covers when a movement is worth a sentence at all.

AI dashboard vs BI dashboard vs report vs list

A BI dashboard shows the charts someone built. An AI dashboard adds questions, flags and explanations on top of the same kind of charts. A report is a fixed document for a date, and a list is a ranked set of rows for someone to work: the twenty customers whose payments slowed most, with an owner for each. When the job is action rather than awareness, the list usually wins; dashboard vs report vs list sets out which gets used.

Questions to ask before you buy one

• Where do the figures come from? Computed from every row by code, or summarized in generated text.
• How are flags set? A rule you can read, per measure, with seasonality handled.
• What happens when data is missing? A late file or a missing column should be stated, not filled in.
• Who can see what? Questions must respect the same access rules as the tiles.
• Do answers use the tiles' definitions? One definition of revenue across the whole page.

Where it goes wrong

• Too many measures. With twenty measures on the page, 1 − 0.9545^20 = 60.6%: a flag most months is noise.
• Seasonal measures judged against a twelve-month average, so every December is flagged.
• An explanation that quotes figures the tools did not compute.
• Flags with no owner or action, so the dashboard is glanced at and closed.
• Questions answered from a different definition than the tiles on the same page.

An AI dashboard on your own files

Covirage's tools set each measure's threshold from its own history, compute what crossed it and cite the rows behind it. The external AI model writes the explanation from those computed figures and answers follow-up questions from the same definitions; it never does the arithmetic. See what moved, why, and the rows behind it, from your own files, and AI data analysis for the method underneath. For the spreadsheet build, see how to build an Excel dashboard; for a ready layout, see the KPI dashboard template; and for the platforms, see dashboard software.