Text-to-SQL

How it works

Business data lives in databases, and databases answer queries written in SQL. Text-to-SQL uses a language model to translate a question — "what did we sell to Acme in March?" — into that query, run it, and turn the result back into a sentence.

Where it goes wrong

The hard part isn't the syntax; it's the business rules. What counts as a sale? Are credit notes subtracted? Are cancelled documents excluded? Which company in a group? A query that ignores those rules returns a real-looking number that's simply wrong.

The safeguards that matter

• Encoded business rules, so every query counts sales the same way your finance team does.
• Guards on what can run — reading only, never changing data.
• An honest empty result. If the query returns nothing, the answer is "I don't know", not a guess.
• Verification against the ERP's own reports.

Where emily fits

Whatever the technique underneath, emily's answers come from queries on your data, follow your ERP's rules for what counts, and say so when the data isn't there. See reliable numbers and connecting AI to your ERP safely.