How to Write SQL Queries with AI

Large language models learned SQL from huge amounts of public code, documentation and Q&A forums, so they are good at producing syntactically valid queries. What they cannot know is your database. Without your schema, a model guesses table names like "users" or "orders" and column names like "created_at", and the query may fail or, worse, run and return the wrong answer. The single biggest improvement you can make is to paste the schema: CREATE TABLE statements, or a list of tables, columns, data types and how tables relate through foreign keys. Adding a few sample rows and notes on business rules, such as "cancelled orders have status = 'X'" or "amounts are stored in cents", prevents many silent errors. State the SQL dialect too. PostgreSQL, MySQL, SQL Server, SQLite, BigQuery and Snowflake differ in date functions, string concatenation, LIMIT versus TOP, and other details. A query written for one may fail or behave differently on another. A reliable workflow has four steps: give context, ask the question in plain English, ask the model to explain its query in words, then test. The explanation step matters because it exposes misunderstandings, such as counting rows instead of distinct customers, or using an INNER JOIN that silently drops customers who have no orders. Common misconceptions include believing that a query that runs is correct (a wrong join can double-count sums), that AI output is safe to run on production (an UPDATE or DELETE without a WHERE clause changes every row), and that you must share real data. Usually the schema alone is enough, which also keeps customer information out of the chat. General assistants such as ChatGPT, Claude and Gemini, coding tools such as GitHub Copilot, and the AI helpers built into many database editors all work best with this same pattern.