Chat with your data: the complete guide to AI data analysis

What "chat with your data" means

Chatting with your data means asking a question about a dataset in ordinary language and getting a real answer back: a chart, a number, a table or a short written explanation. You skip the steps that usually sit between a question and an answer: finding the right table, learning the schema, writing SQL, choosing a chart type and formatting it.

The idea is not new. What changed is that large language models can now reliably translate a question like "Which product line grew fastest this quarter?" into a query against your actual columns, run it, and pick a sensible way to show the result.

There are three broad kinds of tools that call themselves "chat with your data":

1. General chat assistants with file upload. You attach a spreadsheet to a chatbot and ask questions. Quick for one-off checks, but answers scroll away and nothing is saved as a working analysis.
2. BI platforms with an AI add-on. Enterprise dashboards that added a question box. Powerful, but they assume a modelled warehouse and a data team to maintain it.
3. Conversational analytics workspaces. Tools built around the question from the start, where every answer becomes an object you can keep, edit and build on. DashboardGPT is in this group.

Why traditional BI slows you down

In most companies, getting an answer from data follows the same six steps: find the data, understand the schema, build the query, configure the chart, assemble a dashboard, then interpret the result. Only the last step is the one you actually care about.

That gap is why so many questions never get asked. A marketing manager who wants to know whether last month's webinar drove signups either waits in the data team's queue or exports a CSV and fights with a pivot table.

Conversational analytics collapses those steps. You ask, you see, you explore. The work of finding columns, writing the query and choosing a chart happens behind the question.

How it works, step by step

When you ask DashboardGPT a question, four things happen in order:

1. Your data is profiled. When you upload a CSV or Excel file, or connect a warehouse, each column is profiled: its type, whether it looks like a date, a currency, a category or an ID, and how it relates to other columns. This is what lets the AI understand that created_at is a date and amount_usd is money.
2. Your question becomes a query. The AI reads your question alongside that profile and writes a query against your real columns. It does not guess at numbers; it computes them.
3. The query runs and a visual is chosen. Query execution and chart rendering are built on Apache Superset, the open-source BI engine. A trend over time becomes a line, a comparison across categories becomes a bar chart, a single figure becomes a metric card.
4. The answer lands on a canvas with an explanation. You get the chart plus a short written insight that cites the numbers it is based on. The answer stays on an infinite canvas, where you can move it, group it with related answers and ask follow-ups about it.

How to ask a good question

The single biggest factor in answer quality is how the question is phrased. Good questions have three parts:

• A metric: revenue, signups, churn rate, average order value, ticket count.
• A breakdown: by region, by plan, by channel, by week.
• A window: last 30 days, this quarter, 2025 vs 2026.

Compare these:

A few more habits help:

• Use your column names when you know them. If the column is called mrr_usd, saying "MRR" works, but "mrr_usd" leaves no room for doubt.
• Say what you want to compare. "Compare Q2 and Q3 conversion rate by channel" produces a clearer chart than two separate questions.
• Ask why after you ask what. Once you see that March revenue dipped, ask "What changed in March?" The follow-up is where the insight usually is.

Follow-up questions are the real advantage

A one-shot answer is useful. A conversation is far more useful, because each question builds on the last.

On the DashboardGPT canvas, follow-ups know what you're pointing at. Select a chart and ask "Break this down by plan" or "Only show enterprise customers", and the new answer is linked to the one it came from. Context travels with the thread, so you don't have to restate the dataset, the filters or the time window each time.

A typical exploration looks like this:

1. Monthly recurring revenue for the last 12 months. You see a dip in March.
2. Break March down by plan. The dip is all in the Enterprise plan.
3. Which Enterprise accounts churned in March? Two annual contracts ended on the same day.
4. Show net new MRR excluding those two accounts. Growth was flat, not negative.

Four questions, a few minutes, and you have an explanation instead of a chart that just raises questions.

Chat with your data vs pivot tables vs SQL

Each approach has its place. Here is how they compare for everyday business questions:

The practical answer for most teams: use conversational analytics to explore and explain, and keep SQL for the pipelines and reports that must be exactly the same every time.

Can you trust the answers?

AI analysis is only useful if you can check it. Three things separate a trustworthy tool from a risky one:

• It computes, it doesn't guess. Numbers should come from a query that ran against your data, not from a model predicting what the answer probably is.
• It shows its working. You should be able to see the query and the figures behind every chart and every written insight.
• It says when a question is ambiguous. "Revenue" could mean gross, net or recognised. A good tool asks, or tells you which column it used.

Common mistakes to watch for: a date column read as text, totals that double-count because of a join, and averages taken over the wrong grain (for example, averaging daily averages). When a number looks surprising, ask "How was this calculated?" before you share it.

What data can you use?

On the free Explore plan you can upload CSV and Excel files up to 50 MB and ask unlimited questions of the demo dataset. The Team plan adds warehouse connections, scheduled refresh and shared canvases, so answers stay current and the whole team works from the same numbers.

Good starter datasets are the exports you already have: a Stripe or billing export, a CRM pipeline export, an ad-platform spend report, or an orders file from your store.

Getting started in five minutes

1. Open the playground. No account, no install. It opens straight into the product with a demo dataset loaded.
2. Ask three questions about the demo data: one trend, one comparison, one "why".
3. Try a follow-up by selecting a chart and asking it to break down or filter the result.
4. Create a free account and upload one of your own CSV or Excel files.
5. Bring your AI key from your chosen provider so model usage is billed by them directly.