> ## Documentation Index
> Fetch the complete documentation index at: https://docs.retailgrid.io/llms.txt
> Use this file to discover all available pages before exploring further.

# Create a dashboard with AI

> Describe the analytical view you want in plain language and Retailgrid generates the dashboard for you.

When the built-in [Dashboards](/dashboards/overview) don't answer the question you're asking, the **Dashboard Assistant** lets you describe the view you want and Retailgrid generates it from your data.

The assistant lives as a panel inside the [dashboard editor](/dashboards/editor) - AI generation and manual editing work on the same dashboard configuration, so you can prompt, then fine-tune by hand, then prompt again.

## Start from a prompt

1. Open **Dashboards** from the left sidebar.
2. Create a new dashboard (or open an existing one). The dashboard editor opens with the **Dashboard Assistant** panel on the left.
3. Describe the view you want and send the prompt.

The assistant can also generate or update the dashboard's **name** and **description** to match its content.

## Write a useful prompt

The AI does better when your prompt names four things:

1. **The metric** - revenue, gross margin, price index, sales units, competitor undercut count.
2. **The time window** - last 7 / 30 / 90 days, this quarter, year over year.
3. **The dimension** - by category, by brand, by store, by competitor.
4. **The intent** - top-N, comparison vs. last period, trend over time, distribution.

Good prompts:

* *"Show me **gross margin %** for the last **90 days**, broken down by **category**, as a **trend** with one line per category."*
* *"List the **top 50 SKUs by sales** where a **competitor is undercutting us right now**, with the gap in % and absolute terms."*
* *"Compare **price index** **this quarter vs. last quarter** by **category**, sorted by largest change."*

Weak prompts that produce vague results:

* *"Show me pricing."* - no metric, no window, no dimension.
* *"How are we doing?"* - no measurable signal.

## How the assistant applies changes

The assistant reads your prompt **and the current dashboard configuration**, then works in one of three modes:

* **Add** - extends the dashboard with additional widgets or filters. Widgets already on the dashboard are never duplicated; if everything relevant is already there, the assistant says so instead of adding noise.
* **Remove** - takes widgets or filters off the dashboard.
* **Replace** - swaps in a completely different dashboard when you ask for something new.

Changes apply to the dashboard immediately - there's no separate apply step - and every response includes a summary of what was added, removed, or updated. The dashboard **auto-saves** as you go (you'll see *Saved · Just now* under the dashboard name).

## Iterate on the result

Because the assistant and the manual builder share the same configuration, iteration is fluid:

* **Re-prompt** - ask for the change directly (*"swap the trend chart for a bar chart by brand"*). The assistant modifies the existing dashboard rather than starting over.
* **Edit by hand** - add or remove widgets from the widget picker, drag widgets between rows, adjust filters. See the [dashboard editor](/dashboards/editor) guide.

## Limitations

* Widget vocabulary is bounded by the widget catalog and the metrics in your [Metrics glossary](/reference/metrics). Asking for a metric you don't have data for won't work.
* Each dashboard holds one instance of each widget - the assistant won't add the same widget twice.
* Filters are global to the dashboard (Period, Category, Brand, Competitor); per-widget filters aren't supported.

## Related

* [Dashboard editor](/dashboards/editor)
* [Dashboards overview](/dashboards/overview)
* [Datasets](/data-requirements/datasets)
* [Metrics glossary](/reference/metrics)