Select a product:
Introduction to LLM AI Charts
Table of Contents
- Introduction
- Licensing Setup
- Source Data Concepts
- Data Wells & the Virtual Table
- Chart Catalog
- Useful Client Settings
- Interactivity & Cross-Filtering
- Account Settings
- An Example of Iterative Development
- History Search
- Keys and Credits
- Unlocking & Offline / Air-Gapped Use
- Models and Privacy
- Setup Bring Your Own Model
- Frequently Asked Questions
Introduction
Most traditional Power BI visuals work by following rigid layout rules that determine how to draw lines, shapes, and text. These visuals expect you to provide source data in very specific ways (for example, explicitly mapping fields to the X or Y axis).
LLM AI Charts is different. Instead of rigid rules, it allows you to provide a list of fields related to what you want to render, and the AI can self-identify:
- What type of chart makes sense, given the data
- How to best format the chart
This flexibility lets you explore potentially many different ways of rendering your data.
Three ideas run through the whole experience:
- Interactive – when the AI renders with one of the live interactive engines, the chart is not a flat picture. Clicking a bar, slice, point, node, or cell cross-filters the rest of your report, the selected marks highlight while the others dim, hovering reveals tooltips, and legends can cross-filter too. (A static Python image is also available when you'd rather have a fixed picture.)
- Insightful – every version carries a short, AI-written description of what it chose and why (chart type, derived measures, non-default aggregations). That summary shows in the hover information panel, can be pinned as a one-line caption under the title, and is searchable across your version history. The AI can also add derived measures that it thinks are useful, and you can ask it to add more or remove them. The AI can also add annotations and other chart elements like trend lines, forecast ranges, and reference lines. Settings can control much of this behavior, or you can ask for changes in plain English.
- Configurable – beyond plain-English prompting, a rich set of settings lets you steer results repeatably: favor a chart family, cap to a "top N", fix a date grain, pick the render engine, tune cross-filter dimming and the selected-mark glyph, choose the tooltip source, and more. And for the deepest control, advanced users can edit the generated chart code by hand — an Edit code-behind button opens your chart's Python or JavaScript in your account for true fine-tuning. This is entirely optional and aimed at those comfortable with code; you never need to touch it to use the visual.
You can then take the output generated by LLM AI Charts and ask the AI to make adjustments using plain English. From simple requests like "make the legend smaller" to larger changes like "add a forecast range", you can quickly refine the chart toward your ideal layout.
How it works: LLM AI Charts uses Large Language Models (LLMs). You can connect in one of three ways:
- Freemium – the lowest-friction starting point. No account, no payment, no setup. Each visual instance gets a per-instance usage budget (a finite credit grant); once exhausted, you'd start a fresh instance in another report or move to a paid license. Designed for evaluation.
- Use our hosted LLM service (routed through Microsoft Azure AI Foundry, with some models accessed via the provider's native API) – fast to start, no LLM setup required. Hosted models span a range of cost and quality tiers; the live list (with the current default highlighted) appears under Models and Privacy below, and the visual's own model picker always shows what's available to you. Usage credits from your organizational license cover the hosting cost, billed per token consumed (so cost scales with model choice, request size, and complexity).
- "Bring-Your-Own" (BYO) – use your own API key against your own provider account. Supported providers: Anthropic, OpenAI, and Azure OpenAI. In BYO mode you pay your provider directly for tokens consumed; we charge a small flat operating fee per request from your BIC credit balance.
Privacy: LLM AI Charts transmits dataset shape (column names, types, and aggregate statistics like distinct count, min/max, mean), any commentary or per-field descriptions you have entered, and a limited sample of dataset rows used for chart-code execution. The dataset shape is safe-by-construction (aggregates only, no raw values), but commentary, descriptions, and the row sample are transmitted in the clear. Avoid placing personally identifiable information (PII) or other sensitive data into the dataset, commentary, or field descriptions. You can preview the exact payload before sending it via the "See what's sent" link on the visual's landing page (see Models and Privacy). See the LLM AI Charts Privacy Policy for the full data-flow breakdown.
Reliability: LLM AI Charts does not guarantee that every request will produce a usable chart on the first try. Results may vary, and sometimes charts may render poorly or not at all. In such cases, re-issuing the request or adjusting the instructions usually leads to improved results. Built-in settings can also help guide the LLM toward better solutions.
Versioning: Every render request is recorded in history. By signing into your bizintelligencechampions.com account, you can view past versions, revert to earlier ones, and manage your usage credits. All requests carry an associated version number.
Licensing: For sustained use, a bizintelligencechampions.com account with an active organizational license is recommended. A free account includes a 21-day trial license key and usage credits so you can test the visual in your own environment. After the trial, you can continue with a month-to-month subscription or choose a longer-term discounted plan. The freemium mode above lets you try the visual without creating an account first. LLM AI Charts requires an internet connection and uses external data access – meaning it cannot be Power BI Certified. Generating charts requires connectivity, but once you unlock a specific version it will render indefinitely, including in offline / air-gapped deployments.
Rendering: Depending on the chart type and library the LLM picks, your chart may render as a static image (matplotlib / seaborn via Python) or as an interactive chart (D3, an interactive Python engine, or a custom visual layer). Rendering is performed client-side – LLM-generated code executes in your browser via sandboxed runtimes (pyodide for Python, an inline Function constructor for JavaScript, or direct DOM rendering for HTML). Once the layout is determined, charts redraw quickly when model data changes.
Licensing Setup
Everything license-related happens on the visual's landing page — there is nothing to enter in the Power BI format pane. The first time you add the visual to a report, it asks how you want to use it:
- Try it Free — freemium mode. No account and no key needed; you get a limited AI budget per visual instance to see what it can do.
- I have a license — opens an entry panel right on the landing page where you type (or paste) your licensee name, license key, and optionally a client secret. The visual verifies the details against your account the moment you save, so a typo tells you immediately instead of failing on your first Generate.
Where does a key come from? Creating a bizintelligencechampions.com account gets you a trial key good for 21 days — a much larger allowance than freemium, with access to the full model catalog. Purchased licenses appear in the same place on your account page. Entered details are remembered by the visual; if you ever need to switch accounts or keys, use the Change license key link that appears on the landing page once a license is saved.
You do not need to change your reports to use a new license key, when a new key gets added to your account, which happens if you renew your organizational license - either through the monthly subscription, or purchasing an extension (e.g. 1 year, 2 year, etc.). As long as the key used in your reports refers to an account that has a non-expired key, your reports will work.
We encourage you to use the client secret key option, in addition to your license details — it's entered in the same landing-page panel. This increases the level of protection for your account (including usage credits).
One more thing worth knowing: our sample / demo PBIX files use pinned versions, so they render fully — interactivity included — with no license at all. To generate new versions from a sample's data, enter your own license details.
Source Data Concepts
- Values – fields that describe each row (text, dates, or numbers — categorical or continuous are both fine).
- Measures – DAX measures, evaluated within the grain set by your Values and Group by fields (calculated based on that context, page filters, etc.).
- Group by – optional. One to three fields that define the primary grouping (series / breakdown) for the chart. These become the leading hierarchy levels, enabling group-level highlighting, drill-through, and tooltips.
A key concept — granularity. The fields you place in Values and Group by together set the granularity (level of detail) of the data: each unique combination of their values becomes one row. Anything in Measures does not change the granularity — instead, each measure is calculated at that grain, producing one value per row. So if you want finer or coarser detail, add or remove fields in Values / Group by; the Measures well only adds numbers computed against whatever grain those fields define.
Every unique combination of data fields — and optionally an Identifier (configured in Settings) — is associated with a version number.
- Automatic increment: Generating a new chart while in Edit mode automatically increments the version number.
- Revert to previous versions: If you change the Version setting to a lower number, the visual will retrieve and redraw the chart logic for that version, provided it was a successful version. (Versions are always in the context of the choice of fields you are using.)
- Failed versions: A generation attempt that encounters an error still creates a version number, but failed versions cannot be used for rendering.
- Version 0: Setting the version number to "0" places the visual back into its landing page state (like it was when first added to the page).
Another important feature is the information panel. You can make this panel appear by clicking on the chart surface, both in Edit and View modes.
In the above example, we're in Edit mode where additional options are available. From here we can optionally pick an explicit chart type and generate a new version based on that. We can also set the client secret code value (see "Keys and Credits" for more). Finally, we can ask for a new chart version that applies explicit changes that are requested using a prompt. (This is entered the lower text box.)
In v2, the information panel also talks back while you're editing — this is a big part of what makes the visual insightful:
- A reply to your last request, marked ✓ (did what you asked), △ (did it partially), or ✗ (couldn't, with the reason). This is how the visual explains itself when, for example, it chooses a different chart type than you requested because your data doesn't support it.
- An improvement suggestion (💡) — forward-looking advice, such as adding a typical dimension or measure that would unlock a richer chart.
- A rotating "Did you know?" tip highlighting a feature or setting you might not have discovered yet.
These author-facing messages show in Edit mode only (a report's viewers don't see them) and can be turned off with the Hide Some Warnings and Prices setting.
Data Wells & the Virtual Table
The single most useful thing to understand about this visual: it never sees your underlying table. When you add fields to the Values and Measures wells, Power BI builds a virtual table behind the scenes — one row per distinct combination of the Values-well fields, with each Measures-well field aggregated per combination — and that is the data the visual (and the AI) works from. Which fields you put in which well therefore decides which chart types are even possible.
Why should you care? Because the virtual table is a hard boundary the AI cannot see past. The model never connects to your semantic model — it works only with what crosses the wells. If detail was aggregated away, collapsed as a duplicate, or simply never added, no amount of clever prompting can recover it: ask for a "distribution of individual amounts" when only per-category sums came through, and the request doesn't fail loudly — you just get the best chart the data can support, which may not be the chart you meant. AI adds intelligence on top of your data; it cannot add data. Get the grain right first, and everything downstream — the suggestions, the chart choice, the chart itself — gets smarter.
Let's make that concrete with a small example table — 20 rows, two categorical columns
(Region, Product), one numeric column (Amount), and an
identity column (ID):
| ID | Region | Product | Amount |
|---|---|---|---|
| 1 | North | Alpha | 120 |
| 2 | North | Alpha | 145 |
| 3 | North | Alpha | 120 |
| 4 | North | Alpha | 180 |
| 5 | North | Beta | 95 |
| 6 | North | Beta | 210 |
| 7 | North | Beta | 155 |
| 8 | North | Gamma | 300 |
| 9 | North | Gamma | 240 |
| 10 | North | Delta | 75 |
| ID | Region | Product | Amount |
|---|---|---|---|
| 11 | South | Alpha | 130 |
| 12 | South | Alpha | 160 |
| 13 | South | Beta | 88 |
| 14 | South | Beta | 120 |
| 15 | South | Beta | 175 |
| 16 | South | Beta | 96 |
| 17 | South | Gamma | 260 |
| 18 | South | Gamma | 205 |
| 19 | South | Gamma | 330 |
| 20 | South | Delta | 140 |
Scenario 1: just the two categoricals
Add only Region and Product to Values. The virtual table
collapses to the distinct combinations — eight rows, and that's all the visual sees:
| Region | Product |
|---|---|
| North | Alpha |
| North | Beta |
| North | Gamma |
| North | Delta |
| South | Alpha |
| South | Beta |
| South | Gamma |
| South | Delta |
What can you expect? Honestly — not a lot. There is no magnitude anywhere: no measure, and every combination appears exactly once, so even an inferred "count" is 1 for every row. Checking this shape against the engine's actual eligibility rules, about two dozen structural chart types remain on the table (bar/pie/donut families, Treemap, Sunburst, Sankey, Network, heatmap-style grids…) — but each can only show which combinations exist. Everything that needs a number is off the menu: no distribution charts (Box, Violin, Histogram…), no Scatter, no Line, not even a Card. If a generation from a shape like this looks thin, this is why.
Scenario 2: add the ID column
Now add ID to Values as well. Because every ID is unique, the virtual table
stops collapsing — all 20 rows survive (the two tables above, minus Amount).
That changes everything: the visual can now infer a count — four Alpha rows in the
North, one lonely Delta — and the eligible list more than doubles (23 → 52 types in a
back-check against the real rules). The distribution and raw-observation families open up
(Box plot, Violin, Beeswarm, Histogram, ECDF, Event plot…), plus Card, Waterfall, and
others that need a magnitude. The ID's job here isn't to be plotted — it's to preserve the
grain so counting means something.
Scenario 3: a numeric in the Values well (raw grain)
Remove ID and add Amount to Values (not Measures). The virtual
table is now the distinct (Region, Product, Amount) triples — 19 rows, not
20: rows 1 and 3 (North / Alpha / 120) are identical, so they collapse into one. Probably not
what you intended — but this shape genuinely has a job: it keeps the individual numeric
observations, which is exactly what distribution analysis wants. A Box plot, Violin, or
Beeswarm of raw Amounts per Region or Product is only honest at this grain — aggregated data
can't show spread. The collapse is the trap to remember: identical observations silently merge, so
if duplicates are legitimate in your data, keep an ID column alongside to protect the grain.
Scenario 4: the numeric in the Measures well
Move Amount to Measures instead. Now its nature changes: Power BI
aggregates it per combination of the Values-well fields — you pick the mechanism
(Sum, Average, Min, Max…) or supply a DAX measure. The virtual table is back to the eight
combinations, each carrying one calculated number:
| Region | Product | Sum of Amount |
|---|---|---|
| North | Alpha | 565 |
| North | Beta | 460 |
| North | Gamma | 540 |
| North | Delta | 75 |
| South | Alpha | 290 |
| South | Beta | 479 |
| South | Gamma | 795 |
| South | Delta | 140 |
This is the natural shape for comparison charts, and the AI has real magnitudes to work with: the bar families, Lollipop, Pareto, Treemap and the hierarchy family, Heatmaps with true values, Waterfall, Marimekko, Sankey with weighted flows, and so on. One nuance: some distribution types may still be offered on a small aggregated shape like this, but they'd only ever see one value per combination — the picker weighs that, and now you know why a "distribution of sums" is rarely the right ask.
A final, subtle point: aggregation also changes the implicit row count. In Scenario 2 the
visual could infer "four Alphas in the North" from the grain itself; here every combination is
one summary row, so that inference is gone. If the count matters to your story, add an explicit
DAX count measure (e.g. Rows = COUNTROWS('Sales')) to the Measures well alongside the
Sum. In our back-check that second measure opens the door further (52 → 58 types) — the
two-measure family: Scatter, Regression with confidence interval, Joint plot, Polar, Multiple
line, and the multi-KPI Card.
Eligibility lists above were verified against the engine's actual chart-type rules as of July 2026; the catalog grows regularly, so treat the counts as illustrative rather than contractual.
Chart Catalog
The engine can render the chart types below. You don't pick from this list directly — you describe what you want (or simply let the AI decide) and it chooses the best fit for your data's shape. This catalog is generated live from the engine's current chart metadata, so it always reflects what's actually available in the deployed build.
| Chart type | What it's good for | Works best with |
|---|---|---|
| Animated treemap | Treemap whose cell areas animate across an ordered time dimension, showing how a hierarchy's composition shifts period by period | Requires a 2-level categorical hierarchy plus an additive measure observed at 3 or more ordered time points |
| Animated World Choropleth | Animated World Choropleth: a filled world map whose country shading animates across an ordered time dimension, revealing how a measure shifts across countries period by period | Best for a country-level measure observed at 3 or more ordered time points — watch how sales, users, or population shift across countries year by year |
| Annotated heatmap | Heatmap with annotations for each cell | Requires two variables: categorical × categorical with values |
| Arc diagram | Nodes arranged along a single straight axis; edges drawn as semi-circular arcs above/below the axis to show pairwise relationships | Best for 5-50 nodes and a sparse-to-medium-density set of pairwise relationships |
| Area chart | Area chart to show cumulative continuous data over categories | Requires ordered categories or time with a continuous variable |
| Autocorrelation plot | Autocorrelation (ACF) and partial-autocorrelation (PACF) correlogram for time series lag-structure analysis | Requires a datetime series with one continuous variable |
| Bar chart (horizontal) | Horizontal bars; useful for long category names | Requires one categorical and one numeric variable |
| Bar chart (vertical) | Vertical bars; good for categorical vs numeric values | Requires one categorical and one numeric variable |
| Basic Sankey | Visualizes flows from one categorical group (sources) to another categorical group (targets) | Requires two categorical (nodes) and a continuous variable |
| Beeswarm chart | Displays distributions of a numeric variable across categories using jittered points to avoid overlap. | Best for showing distribution density across small categorical sets. |
| Binned bubble matrix | A grid of bubbles at the intersections of two BANDED axes — categorical/ordinal dimensions, or continuous variables binned into ordered ranges (e.g. income <30K / 30-50K / ...). Each cell holds ONE bubble whose size encodes an AGGREGATED measure (mean/sum/median) for the rows in that cell, and an optional independent color can encode a second measure; empty cells stay blank | Two banded/categorical axes (or two continuous variables split into ranges) plus an aggregatable size measure per cell, with optional independent color. Choose this over a continuous Bubble chart when raw points would overplot into a blob, or when the axes are naturally grouped |
| Box plot | Box plot showing distribution and outliers of a continuous variable | Requires one continuous variable (plus optional category) |
| Bubble chart | A continuous bubble chart: each data row is ONE mark plotted at its raw (x, y) position on two CONTINUOUS numeric axes, with marker size encoding a third continuous measure (and optional color a fourth). The axes stay continuous — values are never grouped into bands | Three continuous variables (x, y, size); one mark per row on continuous axes. If you want both axes grouped into bands with one aggregated bubble per cell, pick Binned bubble matrix instead |
| Bullet chart | A compact KPI bar showing a measure against a target marker and qualitative range bands (e.g. poor/satisfactory/good), packing actual-vs-target-vs-context into one small row. | Best for actual-vs-target KPIs with thresholds (revenue vs goal, SLA vs target) — a denser, clearer replacement for gauges. |
| Bump chart | Plots each series' RANK (not raw value) across ordered time points, with crossing lines showing how positions rise and fall over time. | Best for leaderboard/position dynamics: how a handful of regions, products, or teams change rank quarter to quarter. |
| Calendar heatmap | GitHub-style grid showing daily values arranged in week/month calendar format with color intensity | Requires daily datetime data with associated numeric values; best for activity/contribution tracking over time |
| Candlestick chart | Financial chart showing open, high, low, and close values over time. | Best for financial or trading data. |
| Card | A single headline number, optionally with a small label or sub-text underneath. The dataset is summarized to ONE value (sum / mean / count / max / latest) and rendered large enough to read at a glance — ideal for KPI tiles and very small viewports where a richer chart would be illegible. | One numeric value that summarizes the dataset; small viewports |
| Card with multiple KPI | A grid (or stack) of N labeled value tiles, each showing one headline number derived from a distinct numeric column. Ideal for dashboards that need to surface several independent metrics at once on a small tile — Revenue / Users / Conversion / etc. — and especially well-suited to long-narrow or short-wide viewports where stacking N small cards reads far better than a single stretched chart. | Multiple independent numeric metrics that should be surfaced together; KPI dashboards; long-narrow / short-wide viewports |
| Card with sparkline | A headline number paired with a tiny inline trendline (sparkline) that shows recent history of that value. Combines the at-a-glance readability of a Card with a compact visual of how the metric is trending — ideal for KPI tiles in small viewports. | One numeric metric trending over an ordered dimension; small viewports |
| Chord diagram | Circular visualization showing flows or relationships between entities arranged around a ring, connected by curved arcs | Requires matrix data showing bidirectional flows between entities; best for migration, trade, or relationship intensity data |
| Circular heatmap | Circular heatmap to visualize values in radial layout | Requires two variables: one categorical/ordered, one continuous |
| Clustered heatmap | Heatmap with clustering of rows and columns | Requires two variables for matrix values |
| Connected scatterplot | A scatterplot of two measures whose points are connected in time order, tracing the path the relationship takes over time (with an arrowhead or start/end markers showing direction). | Best for the joint trajectory of two related metrics over time (e.g. units vs revenue, unemployment vs inflation) where the path shape is the insight. |
| Custom small multiples | Repeats the same chart across facets for categorical comparison. | Best when comparing patterns across categories. |
| Cycle plot | Splits a time series into one mini-panel per seasonal sub-period (e.g. a small line for each quarter across years), so within-period trend and across-period seasonality are both visible at once. | Best for multi-year monthly/quarterly/weekly data where the question is seasonality vs trend ('is Q4 always highest, and is each Q4 growing?'). |
| Dendrogram (tree diagram) | Hierarchical tree layout: each node is a categorical entity; edges connect parents to children at orthogonal angles | Best for 2-4 hierarchy levels with ≤ ~80 leaves |
| Density contour plot | Contour plot for density estimation of two continuous variables | Requires two continuous variables |
| Donut chart | Donut chart showing category proportions | Requires one categorical variable with proportions |
| Dumbbell chart | Shows change between two numeric values per category using connected points. | Best for before/after or comparison scenarios. |
| ECDF plot | Empirical cumulative distribution function plot | Requires one continuous variable |
| Error bar plot | Plot with error bars for uncertainty in measurements | Requires one continuous variable with error ranges |
| Event plot | Event plot to show occurrence times of events | Requires one continuous variable representing event times |
| Funnel chart | Ordered, tapering horizontal bars showing how a quantity drops off from one sequential stage to the next, with each stage's width encoding its value and the narrowing communicating conversion/attrition. | Best for conversion/attrition pipelines: signup funnels, sales pipeline stages, checkout flow drop-off. |
| Grouped bar chart | Grouped bars to compare subcategories across categories | Requires one categorical, one subcategory, and one numeric variable |
| Heatmap | Heatmap showing values for two continuous variables or one continuous plus one categorical | Requires two variables: categorical × categorical or categorical × continuous |
| Hexbin plot | Hexbin plot for density of two continuous variables | Requires two continuous variables |
| Histogram | Histogram of a continuous variable | Requires one continuous variable |
| Horizon chart | Compact time series visualization using layered bands to show magnitude. | Best for dense time series dashboards. |
| Horizontal/vertical span plot | Highlight horizontal or vertical spans over a plot | Requires one continuous variable with ranges to highlight |
| Icicle plot | Rectangular flat sibling of the sunburst; nested categorical hierarchy rendered as stacked rows or columns of proportionally-sized rectangles | Best for 2-4 levels of hierarchy with ≤ ~50 leaves; an alternative to a sunburst when horizontal/vertical scanning is preferred |
| Joint plot | Joint plot for two variables combining scatter and distribution plots | Requires two continuous variables |
| KDE plot (1D & 2D) | Kernel density estimate plot | Requires one continuous variable (or two for 2D KDE) |
| Lag plot (scatter of value vs lag) | Lag plot for autocorrelation of time series | Requires a datetime series with one continuous variable |
| Line chart | Use a line chart for one or more continuous series over ordered categories; multiple measures or a series category yield multiple lines | One or more continuous series over an ordered/temporal axis |
| Lollipop chart | A cleaner alternative to a bar chart: a thin stem from the baseline to a dot that marks each category's value, reducing ink and clutter when ranking many categories. | Best as a cleaner bar replacement for ranking many categories by a single measure. |
| Marimekko / Mosaic plot | Two-dimensional proportional area chart: column widths encode one categorical total and within-column segment heights encode a second categorical's share | Best for 3-10 outer × 3-10 inner categories with a positive measure |
| Multi-level Sankey | Visualizes flows across more than two categorical stages, showing how quantities split and rejoin as they move through a sequence of categories or processes | Requires three (or more) categorical and one continuous variable |
| Network diagram | Force-directed graph visualization showing nodes and edges representing relationships or connections | Requires node and edge data (source-target pairs); best for relationship, dependency, or social network visualization |
| North America (Bubbles) | North America (Bubbles): each row plotted as a bubble at its latitude/longitude coordinates on a North America basemap - bubble size encodes one measure, bubble color a second measure or a category | Best for point-located data with latitude and longitude columns: stores, cities, facilities, or events across North America, with a measure for bubble size and optionally a second measure or category for color |
| Packed circles | Hierarchical bubble layout where every node is a circle, packed without overlap inside its parent circle; sized by a numeric measure | Best with one numeric measure plus 1-2 categorical groupings producing 20-150 leaves |
| Pair plot (scatter matrix) | Pair plot for multiple continuous variables showing pairwise relationships | Requires at least two continuous variables |
| Parallel coordinates | Multivariate visualization with one parallel vertical axis per dimension; each row is rendered as a polyline connecting its value on every axis | Best for 4-8 numeric dimensions and up to ~1000 rows (heavier overlap with more) |
| Pareto chart | Pareto chart: categories ranked descending as bars with a cumulative-percentage line on a secondary axis and an 80% reference, exposing the vital few causes behind most of the effect | Best for finding the vital few: defect/complaint/incident counts by cause, revenue concentration by product or customer (the 80/20 view) |
| Pie chart | Pie chart showing proportion of categories | Requires one categorical variable with proportions |
| Polar plot | Polar plot for circular data with continuous variables | Requires angle and radius (two continuous variables) |
| Radar (spider) chart | Radar chart comparing multiple continuous variables across categories | Requires multiple continuous variables across categories |
| Regression plot (line with CI) | Regression line plot with confidence intervals | Requires one continuous variable (with optional predictor) |
| Ridgeline plots | Overlapping density plots stacked vertically to compare distributions. | Best with many observations per category. |
| Rug plot | Rug plot to show individual data points along an axis | Requires one continuous variable |
| Sankey alluvial flow | Show how quantities move or transition between categories across multiple stages, with link widths emphasizing the relative size of each flow. | Requires at least two categorical variables |
| Scatter plot | Scatter plot of two continuous variables | Requires two continuous variables |
| Seasonal decomposition | Splits one time series into trend, seasonal and residual components in aligned stacked panels | Requires a datetime series with one continuous variable and at least two full seasonal cycles |
| Slope chart | Connects each series' value at two points with a straight line, so the slope's direction and steepness encode the magnitude of change between them. | Best for two-period change across many categories: YoY by region/segment, before/after comparisons, ranking shifts between two snapshots. |
| Spiral plot | Wraps a long time series around an Archimedean spiral so each full loop is one period (year/week), making periodic and seasonal patterns visually pop. | Best for long, finely-grained, strongly periodic series (years of daily data) where the cyclic pattern is the story. |
| Stacked area chart | Stacked area chart to show multiple continuous series over categories | Requires ordered categories with multiple continuous series |
| Stacked bar chart | Stacked bars for cumulative category comparisons | Requires one categorical, one subcategory, and one numeric variable |
| Stem plot | Stem plot for discrete continuous data visualization | Requires one continuous variable |
| Step plot | Step plot for discrete changes over ordered categories | Requires ordered categories with one continuous variable |
| Streamgraph | Smooth, organic stacked-area variant centered on a baseline that flexes around zero, used to compare flows of multiple categorical series over time | Best with 3-12 categorical series and a continuous time axis with at least 15-30 evenly-spaced points |
| Strip plot | Strip / jitter plot — categorical groups on one axis, continuous values on the other, with a small random horizontal offset on the category axis to spread overlapping points. | Requires one categorical and one continuous variable; jitter on the category axis prevents overplot at the same value. |
| Sunburst chart | Sunburst chart for hierarchical categorical data | Requires hierarchical categorical data |
| Tabular with embedded | Tabular with embedded: one row per category, attribute columns rendered as decorated value-cards and small horizontal charts (range bars, lollipops, sparklines) | Best for a small set of items compared across several attributes at once (a forecast/roster/scorecard table) |
| Time series plot | Time series plot: continuous data over datetime | Requires a datetime axis with at least one continuous variable |
| Treemap | Treemap for hierarchical categorical data visualization | Requires hierarchical categorical data with values |
| USA Choropleth (by state) | USA Choropleth (by state): US states shaded by a measure on a filled Albers-USA map, revealing state-level geographic patterns | Best for state-level US metrics: sales, users, or population by state, any measure that varies by US state (keyed to state names or USPS codes) |
| USA Choropleth (by ZIP-3) | USA Choropleth (by ZIP-3): three-digit ZIP prefix regions shaded by a measure on a filled US map - 5-digit ZIPs are grouped into their ~900 three-digit prefix areas (e.g. 90210 into region 902), revealing regional postal patterns (this is a REGIONAL zip map, not one polygon per 5-digit ZIP) | Best for regional US ZIP analysis: metrics by 3-digit ZIP prefix area, any measure that varies across US postal regions (keyed to 5-digit ZIP codes, grouped to their 3-digit prefix) |
| Violin plot | Violin plot showing distribution of continuous variable with density | Requires one continuous variable (plus optional category) |
| Voronoi treemap | Treemap variant using Voronoi cells (organic polygons) instead of axis-aligned rectangles, sized to encode a numeric measure | Best for 5-30 categories with one non-negative measure; an alternative to a plain treemap when the organic look is desired |
| Waffle chart | Grid of squares showing proportions as colored cells within a fixed grid (typically 10x10 for percentages) | Best for showing part-to-whole relationships with small number of categories; more accurate than pie charts |
| Waterfall chart | Shows cumulative effect of sequential positive and negative values, with floating bars connecting to running total | Best for financial data showing contribution to total (revenue bridges, P&L analysis, budget variance) |
| Word cloud | Text visualization where word size is proportional to frequency or importance score | Requires text data or pre-computed word-frequency pairs; best for qualitative text exploration |
| World Choropleth | World Choropleth: countries shaded by a measure on a filled world map, revealing geographic and regional patterns across the globe | Best for country-level metrics: sales, users, or population by country, any measure that varies by nation (keyed to country names or ISO codes) |
Useful Client Settings
- Consume usage credits.
- Produce a chart that is not saved. (I.e. if the page is refreshed, the saved chart will return to view.)
- Is not guaranteed to be what you prefer to present.
- Is not open to additional prompting by the viewer. (It does use your prompting through the Commentary and Field Definitions settings.)
The visual controls what's inside the visual — not its container. Container-level formatting — background, border, shadow, padding — belongs to Power BI's standard format pane (under General → Effects), and neither the visual nor the AI can change it. No prompt can reach it: it sits outside the boundary the visual is allowed to touch.
The most common surprise this causes: a solid background where you expected transparency. Rendered charts default to a transparent background, but the container's own Background toggle is a separate Power BI setting — if it's left On, Power BI paints a solid fill behind the chart, and the visual can't do anything about it on its own. Asking the AI for "a transparent background" won't help either — the render already is transparent. The fix is in the format pane: General → Effects → Background — switch it off (or raise its transparency to 100%).
The flip side is worth knowing too: container settings are also how you add polish that the AI deliberately leaves alone — a card-style border, rounded corners, a drop shadow. Our demo reports get their "card look" exactly this way.
Interactivity & Cross-Filtering
One of the things that sets LLM AI Charts apart from a typical "AI image" tool is that most of its output is live, not a flat picture. How interactive a chart is depends on the render engine the AI picks (or that you pin with the Render Type setting):
- Static Chart (Python) – a fixed image (matplotlib / seaborn). No hover, tooltips, cross-filter, or selection. Best when you want a guaranteed, pixel-stable picture.
- Interactive (Plotly), Interactive (D3) – live charts with click-to-cross-filter, hover tooltips, and selection highlighting. Plotly suits standard statistical charts; D3 covers more bespoke / custom layouts.
- Leave to visual (default) – lets the engine choose the best renderer for your data and chart type.
Click to cross-filter. On an interactive render, clicking a mark — a bar, slice, point, cell, Sankey node, etc. — filters the rest of your Power BI report to that selection, just like a native visual. The marks you selected stay vivid while the others dim back, so the focus is obvious at a glance.
Selection Dim Opacity (%). Controls how far the non-selected marks fade when you cross-filter (1 = nearly invisible, 100 = no dimming). The default of 25 applies uniformly across heatmaps, bar / scatter / line, pie / treemap / sunburst / funnel, Sankey, and D3 charts.
Selected Mark Glyph. Optionally place a small high-contrast glyph (a Dot or Check) on each selected mark. This is off by default — most charts read fine on dimming alone — but it helps when marks are pale or light-coloured (for example an already-pale heatmap cell) and the dim is hard to see. It is automatically suppressed on flow / relationship and distribution-shape charts (Sankey, Chord, Arc, line, area, box, violin), where the highlight already makes the selection clear.
Tooltips. Hovering a mark on an interactive render shows a tooltip. The Tooltip Source setting chooses where that content comes from — Auto (default), the Visual (Power BI's own tooltip), or the Chart (the tooltip the generated chart code builds).
Plotly-specific controls. For Plotly renders you can additionally:
- Plotly Legend Click Behavior – clicking a legend entry can Cross-filter (default), hide/show the trace (Plotly's own default), or be disabled.
- Plotly Sankey Node Behavior – Sankey nodes can be Click to cross-filter (default) or Drag to rearrange.
- Show Plotly Toolbar – show Plotly's hover toolbar (zoom, pan, box / lasso select, autoscale, reset). Drag-to-zoom and single-click cross-filter work either way.
Drill-through and group highlighting. Fields placed in the Group by well become the chart's leading hierarchy levels, which is what enables group-level highlighting, drill-through, and richer tooltips on grouped / hierarchical charts (e.g. treemap, sunburst).
Note: cross-filtering and selection require an interactive renderer. If a chart came back as a static Python image and you want it clickable, set Render Type to one of the interactive options (or Leave to visual) and regenerate.
Account Settings
- Low Credit Warning Level. Get an alert if your usage credits drops below a level that you can choose.
- Server-side Retries. Similar to the client-side retry setting, this can help catch certain types of errors and retry them silently. This may use additional usage credits, depending on the type of error.
- Override Upload Thumbnails. When disabled, the option to upload thumbnails is based on the client setting Save Thumbnails. When enabled, it will ignore the client setting and save thumbnails for all instances of the LLM AI Charts visual, for your account.
- Force LLM Model Type. This is a way to override any client-side selection of model type. If you have developers who will be able to change the model type setting on the client, this is a way to make an account-level decision to stick with a particular LLM model. This can make it simpler for developers but also potentially control costs better, since you can identify the cost (e.g. usage credits) across all instances of LLM AI Charts using your account.
An Example of Iterative Development
History Search
There are two sections: search criteria and results. For search criteria, you can use:
- Start Date, End Date. You can express this in local time. This compares to the date and time the version was recorded.
- Text. This text is wild-card compared against fields and the description text.
- Include Failed. Normally you shouldn't need to see failed versions, but you can optionally enable this to identify facts like when they happened.
For search results, you can see:
- Date. This is expressed in your local time.
- Identifier. If you used an identifier and at least one version has a non-blank value, this will be shown.
- Version. This is the all-important version number that can be set in the client-side settings.
- Fields. All data fields that participated in creating the listed version are shown here. The field list plus identifier form a unique "request group" that has its own version numbering, starting at "1". You can't "cross over" and start using a version for a different set of fields and/or identifiers.
- Description. This is generated by the LLM as a summary of what it did, to create the chart layout.
- Ref. If you generated the version based on another existing version, this provides a reference to the source version number. This helps paint a picture of version lineage.
- Thumbnail. If you've enabled thumbnails, this helps remind you of what the resulting chart looked like.
Keys and Credits
- Generating a new version: charged per request based on the number of tokens the language model processes — both the input (your dataset shape, any commentary and field descriptions, and a sample of rows) and the output (the generated chart code and its summary). The richer the request and the more capable (and costly) the model, the more tokens are consumed and the more credits the request uses. Each attempt is counted, so a request that retries (see the Retry Count setting) consumes the tokens for each attempt. A small per-request minimum applies, so even a tiny request costs at least one credit. Because this is token-based, there is no single fixed price per chart — cost ranges with your data size, prompting, model choice, and how many attempts are needed.
- Loading an existing version: a small flat charge per load (no language-model call is made, so no tokens are consumed). This applies only when a version must be fetched from our services — not during a report load or refresh in View mode, since versions are cached.
- Bring-Your-Own (BYO): when you use your own Anthropic, OpenAI, or Azure OpenAI / Azure AI Foundry account, your provider bills you directly for the tokens consumed. From your BIC balance we charge only a small flat operating fee per request — the token cost itself does not come out of your BIC credits.
You can find the latest rates and worked examples on the usage credits price list.
You can check your credit balance in both your Account Profile and within the visual (tooltip and information overlay). You may also set a credit warning level to receive an email alert when your balance drops below a threshold you choose.
For added security, LLM AI Charts allows you to set a Client Secret Key:
- Generate this key from Account Profile → Product and Keys → API and Download Keys.
- The key expected in LLM AI Charts is your account’s API key.
- Note: You can only set this key inside LLM AI Charts on the landing page or information overlay (not in settings) — it is never displayed for security reasons.
- If your account has an API key enabled, you must configure it as the Client Secret Key in LLM AI Charts.
Unlocking & Offline / Air-Gapped Use
Normally, loading a saved version asks our service for the rendering logic, and the visual checks in with us to confirm your license. Unlocking changes that for a specific chart version: it is a one-time action (a credit charge, taken once for that version) that permanently licenses that version to you.
Once a version is unlocked:
- It loads for free. Re-displaying an unlocked version doesn't consume further credits — you've already paid to unlock it.
- It keeps working indefinitely. The visual carries a cryptographically signed proof of the unlock, so that version continues to render even after your organizational license has lapsed. (You still need a current license to generate or unlock new versions.)
- It renders offline — including air-gapped. Because the signed proof can be verified locally, an unlocked version draws even when the visual can't reach our servers at all. Versions that haven't been unlocked are blocked when we're unreachable.
This is the supported path for environments with no outbound internet access. You can't create a chart air-gapped (generation needs to reach our LLM service), but you can do your design work where there is connectivity, unlock the specific versions you want to keep, and then deploy that report into an air-gapped or intermittently-connected setting where those unlocked versions render reliably and indefinitely.
You unlock versions from your chart history on your bizintelligencechampions.com account. The current unlock cost is credit-based and may change over time; your account always shows the charge before you confirm.
Distributing to Others — Pinning (No License Required)
Unlocking licenses a version to you. Pinning goes one step further: it prepares a finished chart to be handed to other people. When you pin a version for distribution, the visual freezes that rendering and strips your license, keys, and account credentials out of the report. The result is a self-contained chart that, technically, carries no license at all.
A pinned rendering:
- Renders for anyone, anywhere — watermark-free. Power BI Service, Desktop, embedded, even fully offline. The recipient needs no bizintelligencechampions.com account, no license, and no usage credits.
- Leaks nothing. Because the credentials are stripped out, a shared file can't be used to generate against your account or spend your credits.
- Is a frozen snapshot. The recipient can view and interact with the chart, but can't re-generate or edit it. To change a chart, you re-pin a fresh version from your licensed account and re-deliver.
Why this matters for consultants and agencies. If you build Power BI reporting for clients, you can license LLM AI Charts once, through your own consultancy, and use it to build and pin as many client deliverables as you like. Your clients run the finished reports with no license of their own — you license once and deliver to many. It's the same mechanism behind our sample PBIX files, which ship pinned so they render fully for anyone who opens them.
You pin versions from your chart history on your bizintelligencechampions.com account — the same place you unlock them. One practical note: connect once after pinning and save the report before sharing, so the stripped, share-safe copy is the one your recipients receive.
Models and Privacy
You have three choices for how requests reach an LLM: freemium (no account, evaluation mode), use our hosted LLMs (routed through Microsoft Azure AI Foundry, plus some models accessed via the provider's native API), or Bring-Your-Own (BYO) with your own API key (Anthropic, OpenAI, or Azure OpenAI).
Using our hosted LLMs means:
- You can start using the visual immediately — no model setup is required on your part.
- Information about your data is included in a prompt that goes through the model, hosted in Microsoft Azure AI Foundry or accessed via the upstream provider's native API.
- Prompts are logged for up to 90 days so we can diagnose problems and improve our algorithms.
- Your metadata is not used for training. It is not accessible to other customers.
- For security and compliance, prompts may be temporarily stored and reviewed by the upstream provider if flagged for abuse (per that provider's published terms).
Currently available hosted models (this list is generated live from our catalog — it evolves as we tune offerings, and the visual's model picker always shows what's actually available to you):
- Anthropic - claude-fable-5 – Anthropic Claude Fable 5 — is Anthropic's most capable model. Top-tier design judgment and analytical depth on the most ambitious charts, at the highest per-call cost. Trial / paid only (not available on freemium). Choose it when quality matters most and budget is generous.
- Anthropic - claude-opus-4.8 – Anthropic Claude Opus 4.8 — our deep-reasoning model for the most ambitious or complex charts. Highest quality at a premium per-call cost. Trial / paid only (not available on freemium).
- Anthropic - claude-sonnet-4-6 – Anthropic Claude Sonnet 4.6 — our balanced default. Consistently high-quality charts with strong design judgment at a mid-tier cost. Recommended for everyday use.
- Anthropic - claude-sonnet-5 (current default) – Anthropic Claude Sonnet 5 — the best combination of speed and intelligence. Fast, high-quality output with strong chart-design judgment and reliable code generation. Recommended for everyday use.
- Azure OpenAI - gpt-5-mini – OpenAI gpt-5-mini — our lowest-cost model. Fast and dependable for everyday and high-volume charts. For the most demanding or design-sensitive work, a higher-tier model will usually do better.
You can override your model selection at the account level (applies globally to all your visual instances) by signing into your bizintelligencechampions.com account.
The amount of data included in generated prompts is controllable via the Privacy Level setting. The default sends dataset shape (column names, types, and aggregate statistics like distinct count and min/max) plus any commentary or per-field descriptions you have entered. A limited sample of dataset rows is also transmitted for chart-code execution; these rows are not obfuscated. Avoid placing personally identifiable information (PII) or other sensitive data into the dataset, commentary, or field descriptions.
You don't have to take our word for it. Right on the visual's landing page, next to the privacy picker, is a "See what's sent" link. Clicking it opens a preview overlay showing the exact JSON payload that would be transmitted for the current data and settings — before you ever press Generate. Because the payload reflects your chosen Privacy Level, you can raise or lower that setting and watch precisely what changes, so there are no surprises about what leaves your report.
Setup Bring Your Own Model
Bring-Your-Own (BYO) LLM mode lets you supply your own API key against your own provider account. Three provider families are supported: Anthropic (their Messages API directly), OpenAI (api.openai.com directly), and Azure OpenAI (your own Azure OpenAI / Azure AI Foundry deployment). In BYO mode you pay your provider for tokens consumed; we charge a small flat operating fee per request from your BIC credit balance.
The walkthrough below shows the Azure AI Foundry setup path end-to-end. For Anthropic or OpenAI BYO, the setup is simpler: just generate an API key in your provider's console and paste it into the corresponding field on your bizintelligencechampions.com account's LLM AI Charts settings.
Azure AI Foundry walkthrough — prerequisites: an Azure subscription. The following steps cover basic setup, from that starting point.
1. Create an Azure AI Foundry
In this example, we accepted most defaults, outside of what is shown above.
2. Go to the Azure AI Foundry Portal
Visit the foundry you just created in the previous step...
3. Find a model and create a deployment for it
You can start with the Model Catalog:
Here we found and picked "gpt-4.1":
We clicked "Use this model" and are prompted for additional details and have the option to "create resource and deploy":
Details about deployment types and pricing are not in scope here: this example is only showing basic setup. More details can be found by visiting Microsoft. After our resource was created, we might see something like:
4. Extract three settings and copy to your bizintelligencechampions.com account
The source for these settings are shown here:
(You can also obtain the key and endpoint using Foundry, with the "Get endpoint" button.)
These values should be placed here, in your bizintelligencechampions.com account:
You can enable BYO LLM on a visual-by-visual basis, or globally at your account level (default model = BYO LLM).
FAQ
Q: Once I've set up my instance of the LLM AI Charts visual with my data, am I charged usage credits?
A: Not unless you've allowed report viewers to generate new versions. (The default setting is false.)
Credit usage is typically incurred during design.
Q: What's the difference between the "Values", "Measures", and "Group by" data wells?
A: Values does not allow DAX measures and is generally intended for categorical data (though not strictly required).
Measures allows DAX measures, which are typically continuous but do not have to be.
Group by is optional — one to three fields that define the primary grouping (series / breakdown). They become the leading hierarchy levels, which is what enables group-level highlighting, drill-through, and richer tooltips.
Q: Can I click the chart to cross-filter the rest of my report?
A: Yes, when the chart renders with an interactive engine (D3, Vega, or an interactive Python engine). Clicking a mark filters the report and highlights your selection while the rest dims back. A Static Chart (Python) render is a fixed image and is not clickable — set Render Type to an interactive option (or Leave to visual) and regenerate if you want interactivity. See Interactivity & Cross-Filtering for the full set of controls.
Q: I keep getting errors or invalid charts as output.
A: Your dataset may not be understood well enough by the LLM. Try adding more details in the Commentary setting
(a free-text area where you can provide up to 200 characters of guidance for the model).
If issues persist, please submit a support request
or email info@bizintelligencechampions.com.
We review usage details and make improvements in future releases based on real-world use cases.
Q: I make requests for certain chart types, but the visual never shows these.
A: An explicit chart type is treated as a strong preference, but it can be overridden when your data can't support it.
For example, a scatter plot requires two continuous variables — if your dataset only has categorical values,
the visual will choose a more appropriate chart type. In v2 you're no longer left guessing why: when the visual
picks something different from what you asked, it can reply in the hover information panel (while editing) with a
short note marked ✓ (did what you asked), △ (did it partially), or ✗ (couldn't — with the reason). So you might see,
for instance, "✗ A scatter plot needs two numeric measures; showed a bar chart of counts instead."
These author messages appear in Edit mode unless you enable the Hide Some Warnings and Prices setting.
Q: I know what I want in terms of "top 10" or "group by month" (instead of by day or year).
Do I need to keep requesting changes using the iteration process?
A: Not necessarily. Settings such as Favor Limit To and Favor Date Summary are designed for these cases.
You can use settings to influence results, or continue refining with adjustment requests—both approaches work.
Q: Are my results guaranteed to be accurate?
A: No — nothing is guaranteed with AI. Like any work product, results should be validated and tested.
Q: Does the visual work in every cloud environment?
A: In most standard scenarios, yes. However, some government clouds (e.g. DoD) may have restrictions that prevent communication with our services, which is necessary for the visual to function. We recommend testing in your target environment, using the trial option. If you encounter difficulties, feel free to submit a support request.
Q: Which model type is "best"?
A: This is subjective, and the catalog evolves — so rather than pin specific names here, see the live list under Models and Privacy. In general: the current default is chosen to give a strong quality-vs-cost balance; higher-tier models produce cleaner output on complex requests at proportionally higher cost; and the lowest-cost option trades some quality (and a higher retry/error likelihood) for price. The visual's model picker lists each option, and you can override the choice per visual or account-wide.
Bring-Your-Own (BYO) mode also lets you use models we don't host directly — supply your own API key for Anthropic, OpenAI, or Azure OpenAI / Azure AI Foundry.
Q: Am I charged usage credits even when the LLM returns a version that does not render properly (e.g. an error)?
A: Yes - its considered usage. Charges are token-based: each call to the language model consumes input and output tokens whether or not the result ends up rendering, and those tokens are what you're billed for. So the actual number of credits per chart is not fixed - given the iterative process of building your chart, there is a range that depends on your prompting, how much information you provide to the LLM, your model choice, and how many attempts are made.
Q: I'm low on usage credits. What options do I have?
A: You can buy a standalone usage credit pack (50,000 credits for $149). Credits are also included with each organizational license term:
- 1-year license = 60,000 credits
- 2-year license = 150,000 credits
- 5-year license = 600,000 credits
Longer terms provide more credits per dollar. Purchasing a new license while one is active extends your current end-date — you never lose days. See the price list for current pricing.
Q: Can I use LLM AI Charts in an air-gapped scenario (no internet access)?
A: For generating charts, no — creating a new version requires reaching our LLM service. But you can render previously unlocked versions with no connectivity at all: unlock the specific versions you need while you have internet, and they will render indefinitely in an air-gapped or intermittently-connected deployment. See Unlocking & Offline / Air-Gapped Use for details.
Q: Will the visual be Power BI Certified?
A: Not likely. Certification requires no external data access, but LLM AI Charts depends on our services and some public CDNs
(e.g. cdn.jsdelivr.net).
Q: I have an idea for improvement, how can I share this? (or: I'm not seeing an answer to a question I have.)
A: We’d love to hear from you. Please email info@bizintelligencechampions.com
or submit a ticket.