Report Builders Are Dead
Several recent projects have come with a request for a report builder. I don't build them anymore. Give an AI your database structure and tell it how to query the data, and the user can just ask for the report they want.
A few of the projects I have worked on recently came with the same request. The client wanted a report builder.
You know the kind of thing. A screen where you pick a table, tick the fields you want, add some filters, choose how to group it, and export the result to a spreadsheet. Almost every business system ends up with someone asking for one, because the reports that shipped with the system never quite cover the question they have this week.
In the age of AI there are much better ways to do this.
What a Report Builder Actually Is
Underneath all the dropdowns, a report builder is a way for someone who doesn’t write SQL to write SQL.
Every field you tick becomes a column in a select. Every filter becomes part of a where clause. Every “group by” box is exactly what it says. The interface exists to translate what the user wants into a query, and to stop them from writing one that breaks something.
That translation is the part AI is now good at. The user doesn’t need a set of dropdowns to describe what they want. They can just say it.
Send the AI Your Database Structure
The approach I use instead is simple. Send the AI your database structure, then let it write the query.
The structure can be live. Rather than maintaining a document describing the database, the schema is read from the database itself when the request is made, so when a new table or column is added it is available straight away. Nobody has to go back and update a report builder to include it.
You also control what the AI gets to see, and there are two ways to do that.
Exclude specific fields
Send everything except the fields you name, such as password hashes, tokens, or personal details nobody should be reporting on.
Include specific fields
Send nothing except the tables and fields you name. Anything new stays hidden until someone decides it should be reportable.
Which one suits depends on the system. Exclusion keeps up with the database on its own. Inclusion means nothing gets exposed by accident.
You Tell the AI How to Query the Data
The schema tells the AI what exists. It doesn’t tell it what anything means.
So along with the structure, you tell the AI how to query the data. Which table holds the real order totals and which one is a cache. That a cancelled order still exists in the table and has to be filtered out. What the business means by an active customer. Which field to use for dates, and what time zone they are in.
This is the knowledge that would otherwise be baked into a report builder’s options, or more often, sitting in the head of whoever built the system. Written down once as instructions, it applies to every question anyone asks.
Then the user asks for what they want, in plain English.
SELECT c.state,
SUM(CASE WHEN o.placed_at >= '2026-07-01' THEN o.total END) AS this_year,
SUM(CASE WHEN o.placed_at < '2025-10-01' THEN o.total END) AS last_year
FROM orders o
JOIN customers c ON c.id = o.customer_id
WHERE o.status <> 'cancelled'
AND (o.placed_at >= '2026-07-01' AND o.placed_at < '2026-10-01'
OR o.placed_at >= '2025-07-01' AND o.placed_at < '2025-10-01')
GROUP BY c.state
ORDER BY this_year DESC;
What comes back. The cancelled-order filter is there because the instructions said it had to be.
The query runs, and the user gets their report. If it isn’t quite what they wanted, they say so and it gets adjusted. No new screen, no feature request, no waiting for a developer to add a field to a dropdown.
Other People Have Built This Too
I am not the only one doing this. There are open source projects on GitHub built around the same idea, and they are worth a look if you want to see how far it can be taken or would rather start from something that already exists.
WrenAI
Turns plain English questions into SQL, charts and dashboards, with a layer where you describe what the data means so the answers stay consistent.
Vanna
You give it your schema, documentation and example queries, and it writes SQL from questions. The repository is now archived, but it is still a clear example of the approach.
DB-GPT
A larger framework for building AI data assistants, with text to SQL as one part of it.
Chat2DB
A database client with AI built in. Connect it to a database and ask it to write, explain or tune queries.
SQL Chat
A chat based SQL client. You talk to the database instead of writing the query yourself.
Why I Stopped Building Them
A report builder is a large piece of work. It has to understand joins, handle every field type, validate filters, and stay in step with the database every time the database changes. It is often one of the most complicated screens in the whole system, and it is built for a handful of users.
The AI approach is a fraction of that. A connection to the database, a schema read with the fields you have chosen to hide or show, a set of instructions about how the data works, and a box to type into.
So when a client asks me for a report builder now, that is what they get instead.