Asking the data in plain language means typing a business question in ordinary words and getting the answer from the company's data on the spot. An AI model turns the question into a database query, runs it on the data warehouse and explains the result. The industry calls this conversational analytics or natural-language analytics. When the AI takes several steps on its own, it is called agentic analytics. Older tools of this kind were called search-driven analytics, because the question was typed into a search box.
Without such a tool, a manager emails an analyst and waits. The analyst writes a query in SQL, the language that databases understand, runs it, makes a chart and replies, sometimes days later.
Here is a typical case. The head of credit cards at a private bank in Mumbai sees that card spending in Pune fell 9% in October, while every other city grew. She types: "Why did card spending in Pune fall in October?"
First, the tool finds the right tables. A warehouse can hold thousands of tables, with names such as txn_fct_v2. So the tool reads the bank's data catalog, a searchable list of every table and what each column means. Second, it sends her question and the table descriptions to a large language model. The request usually passes through an LLM gateway, which sends it to the chosen model and records its cost. The model writes a SQL query. Third, the tool runs the query and reads the result. It sees that one kind of shop stands out, so it writes a second query that splits the fall by type of shop. Writing a query, running it, reading the result and trying again is how AI coding agents work.
The answer comes back in about a minute. Spending at electronics shops in Pune fell by ₹14 crore, which is almost the whole fall. In September, a large electronics chain in Pune had run a cashback offer on the bank's cards, and the offer ended on 30 September. The tool shows a chart, a short explanation and the queries it ran.
Showing the queries matters. A language model can write a query that runs correctly but answers a slightly different question. For example, it may count refunded purchases as spending. The model then states a wrong number with confidence. This is a form of AI hallucination. So the tool shows its work, and an analyst checks any answer that will be used for a decision. The tool must also apply the warehouse's permissions, so that each manager sees only the data that her login allows.
How is this different from a BI dashboard? A dashboard answers questions that an analyst planned in advance, and many people read it again and again. A plain-language tool answers a new question, once, for the person who asked it. Whether these tools will replace dashboards, or sit beside them, is still an open question. The general idea is explained on the page on the AI data analyst.