SQLProAI turns a question typed in plain English into read-only SQL against your own database. You decide which tables are askable; secrets and PII are hard-denied; every query is bounded before it runs. Ask, refine, and share the result.
| region | revenue |
|---|---|
| West | $412,800 |
| Northeast | $388,140 |
| South | $301,560 |
| Midwest | $264,900 |
The number you need is sitting in the database. Getting it out shouldn't mean filing a ticket, wrestling a dashboard that was built for last quarter's question, or handing someone raw query access.
Ad-hoc questions queue behind a person who has to write the SQL. A ten-second answer takes two days — so most questions never get asked.
BI tools answer the questions someone anticipated. The moment you want to slice it a new way, you're back in the queue or exporting to a spreadsheet.
Handing a chatbot — or a person — a live connection can leak PII, expose secrets, or drop a query that scans your whole warehouse. Convenience shouldn't cost you control.
You stay in control at every step. SQLProAI connects with read-only credentials, learns only the tables you expose, and never writes anything back.
Add a data source with read-only credentials — including databases behind a firewall or in a private network, via a secure connector. SQLProAI introspects the schema — tables, columns, and how they relate — so it knows the shape of your data.
Choose what's askable. AI suggests plain-English descriptions; you approve the tables and columns to expose. Secrets and PII stay hard-denied even if you forget.
Type the question the way you'd say it. It becomes SQL, gets validated, and runs read-only. Not quite right? Refine in plain English — "now only last month."
See the result as a clean table with an auto-generated chart, and the exact SQL for anyone who wants it. Share a link, or export the full report as Excel, PDF, CSV, or an image — and have it delivered by email, text, or an agent. Reproducible, without re-running the ask.
Safety isn't a checkbox you can forget to tick. Every query passes the same gauntlet before a single row is read — and there's no path around it. Built for regulated data: banking, healthcare, anywhere a leaked value or a wrong number isn't an option.
Queries are validated to be plain SELECTs, then executed on a read-only connection. Nothing can write, alter, or delete — by construction, not by policy.
Nothing is askable until you allow-list it. A question can only touch the tables and columns you've deliberately exposed.
Password hashes, tokens, and sensitive fields are blocked at a layer that overrides exposure — so an over-eager question can never surface them.
An automatic row limit, a cap on joins, a scan-cost estimate, and a per-query timeout stop a runaway query before it runs — no accidental full-warehouse scans.
Expose a table that has an email, phone, or card number without leaking it — the value comes back masked while filters and joins still run on the real data underneath. A masked or PII column's raw value is never sent to the AI, on any path.
When someone asks a question, the model writes SQL from your table and column names — it never sees the result rows. People get the full answer; the model doesn't. For regulated data, one switch puts it in names-only mode, where no real values reach the model at all.
Point it at a shared, multi-tenant database and lock a workspace to its own rows. The filter is injected into every query after the model writes it — there's no way around it.
Queries run against your database and return only the preview rows. Your tables aren't copied, mirrored, or shipped anywhere — and what the model does see is never used to train it.
Each question is metered per workspace, so usage and cost are always visible — no surprise bills, no black box.
Aggregates can leak a person. Counts below a threshold you set are suppressed, and any group smaller than it is dropped when a query groups by a quasi-identifier like ZIP or birthdate — so a narrow question can never single someone out.
Every query is scored low, medium, or high from its shape and scan size. You decide what a high-risk one does on each source — allow, warn, require confirmation, or block — enforced identically for the app, the API, and agents.
A shared report link expires and can be revoked instantly; cap the rows it may contain, watermark it, make it open once, or send the link with no file attached. Exports are the real exfiltration surface, so they get the tightest controls.
Pointing an LLM at a database is easy; getting the number right is the hard part. SQLProAI is built to answer what you actually meant.
Define "revenue" once as an approved formula, and every question uses it — no two answers that disagree because the model invented a different calculation.
SQLProAI knows each table's grain and detects fan-out and chasm traps — the classic joins that silently multiply a total. It fixes the query, or asks you to confirm, instead of returning an inflated number.
It reads real sample values, so it filters on washer, not a guess at the spelling — and it learns from the stored procedures and views already in your database, the business logic your team encoded, so it understands how a metric is really computed. Every answer shows the exact SQL, so a number is always checkable.
The model is a replaceable component behind the governed layer — swap it and the guarantees still hold, because they live in the layer, not the model. And correctness isn't a promise: in our evaluation suites it measures 96% on a real, deliberately-misleading multi-tenant schema and 100% on an independent public dataset — with the exact SQL shown every time, so any number is checkable.
Access is defined by capability, not a shared admin login. An auditor can read every query and change nothing; a curator can refine definitions but never touch security controls or who has access. Separation of duties, enforced.
An append-only audit trail the application can't rewrite logs each question, its exact SQL, the outcome, and a risk rating — while result rows are never stored. When someone asks how a number was produced, there's an answer.
Change what “revenue” means and it's tracked field-by-field and reversible; publish a reviewed version of the model and roll back to it. The governance an auditor expects — versioned, attributable, and yours to hand over.
Need prompts to stay inside your own tenancy? Inference can run through Amazon Bedrock or Google Vertex, chosen per data source, so nothing leaves your AWS or GCP boundary.
The model is a component you pick, per source — a faster one for routine queries, a stronger one for setup — from an approved list. When a better model ships you switch to it; the guarantees don't change.
A built-in evaluation harness scores accuracy against answers computed independently, with run-over-run history — so a regression is visible, not a surprise. Governance you can prove, not just claim.
A built-in certification ladder takes anyone from their first question to platform architect — Certified User, Curator, Security Admin, Architect. Adoption doesn't hinge on one expert, and everyone learns the guardrails, not just the queries.
No column names, no joins to remember. Ask like you'd ask a colleague who happens to know the schema cold.
Ask in the app, from your own code, or through an agent — every path runs the same read-only, allow-listed, scoped, masked pipeline.
Ask on the web and refine in plain English, or describe a report — "revenue KPI, sales by type as a bar, the monthly trend as a line" — and get a multi-chart report that saves and re-runs against live data. Download any report as Excel, PDF, CSV, or an image.
A governed REST API and a standard MCP server. Give an app — or an AI agent like Claude — a scoped key and it can ask your data in plain English, with the same guardrails and no direct database access.
Attach a data source to a phone, SMS, or web-chat agent so your customers can ask out loud — answers shaped for speech, and scoped to each caller's own records. It can text or email them the full report, chart included.
Connect with a read-only role — ideally a read replica. SQLProAI never needs write access.
Flat monthly pricing — no per-seat surprises. Every plan includes the full governed engine, the correctness controls, API & agent access, and the certification Academy.
Need higher volume, or an in-cloud / on-prem deployment? Talk to us.
Connect a source, expose the tables you trust, and start asking in plain English — with the guardrails already on.