Chat With Your Data: Ask Business Questions in Plain English

Chatting with your data means typing a business question such as ‘Which region grew fastest last quarter?’ and getting an answer calculated from your spreadsheet or database, without writing formulas or SQL. It works best when your question is specific, your data is clean, and you check how the answer was produced. This guide explains what happens behind the scenes, how to phrase questions that get useful answers, and how to verify results before you act on them.

How natural-language data analysis actually works

When you ask a question in plain English, a well-built tool does not simply let a language model ‘read’ your file and guess. The sensible pattern has three steps:

1. Interpret the question. The AI maps your words to the columns in your data. ‘Revenue’ might map to a column called amount_incl_vat; ‘last quarter’ becomes a date filter.
2. Compute the answer with code. The actual numbers come from a query or calculation run against your data: a sum, a group-by, a filter, a comparison. This is deterministic, so the same question on the same data gives the same figure.
3. Explain the result. The AI writes a short narrative around the computed figures: what changed, by how much, and what might be worth a closer look.

The key point is that numbers should be calculated, not generated. Language models are good at language and weak at arithmetic over thousands of rows. If a tool asks the model to produce totals from memory, you will eventually get a confident, wrong figure. We cover this in more depth in how to stop AI hallucinating numbers.

Good questions versus bad questions

The quality of the answer depends heavily on the question. Good questions name a measure, a grouping and a time frame. Vague questions force the tool to guess what you meant.

Notice that ‘why’ questions are the hardest. Your data can show what changed and where, but it usually cannot prove the cause. A drop in margin may line up with a supplier price increase, but unless the price increase is in your data, the tool can only point you at the category that moved.

A simple question formula

Use this pattern: [measure] by [grouping] for [time period], [sorted or filtered how]. For example: ‘Average order value by sales channel for Q2, highest first.’ It is not magic phrasing, but it removes most ambiguity.

Follow-up questions are where the value is

The real benefit of a chat interface is the follow-up. Say your first answer shows that Gauteng sales rose from R420,000 in April to R510,000 in May. Natural follow-ups might be:

• ‘Which products drove the Gauteng increase?’
• ‘Was the increase from more orders or larger orders?’
• ‘Did the same thing happen in May last year?’

Each question narrows the story. In a spreadsheet, each of those would be a new pivot table or filter. In a chat, it takes seconds, which means you actually ask them. If you are weighing the two approaches, see pivot tables vs AI analysis.

How to verify an answer before you trust it

Treat a chat answer like a figure from a new colleague: probably right, but worth a quick check the first few times, and always before it goes into a board pack.

1. Check the definition. Did ‘sales’ include VAT? Returns? Cancelled orders? Ask the tool which column and filters it used.
2. Reconcile one total. Compare the grand total to a number you already know, such as the revenue figure from your accounting system for the same month.
3. Spot-check one row. Pick a single customer or product and confirm its figure by hand.
4. Look at the row count. If you expected about 12,000 orders and the answer is based on 3,400, a filter is probably wrong.
5. Watch for data quality issues. Duplicate rows, blank dates and inconsistent category names quietly distort results. Our data cleaning checklist covers the common fixes.

Privacy: what the AI should and should not see

Before you connect customer data to any AI tool, find out what gets sent to the language model. Ideally the AI sees column names, summary statistics and a small, masked sample, not your full customer list with names, ID numbers and phone numbers. If you process personal information in South Africa, POPIA applies, so data minimisation is a sensible default. (This is general information, not legal advice.)

In Summarix, personal information such as emails, phone numbers and SA ID numbers is masked before any AI call, and the AI works from a statistical profile plus a masked sample of at most 15 rows. The figures themselves are computed by Summarix's own code. You can read more on the security page.

When chat is not the right tool

• Regulated figures such as VAT returns or audited financials should come from your accounting system, not an ad-hoc question.
• Complex models such as cohort retention curves or forecasting with many variables usually need a proper analyst or purpose-built tooling.
• Messy, unlabelled data will give messy answers. Clean first, then ask.

Used well, chatting with your data turns ‘I will ask the analyst on Monday’ into a two-minute check. Ask specific questions, follow up, verify the first few answers, and keep an eye on data quality. Chat is available on paid Summarix plans; see pricing for details.