AI solution design · Databases and data
Ask your database
Managers ask a question in plain English and get a table from the SQL Server database, with the query and an explanation shown.
Talk about something like thisThe situation
Managers wait for IT or an analyst every time they want a figure that is not on an existing report. "How many orders over £5,000 shipped late in September, by customer?" is a ten-minute question for someone who knows the schema and a two-day wait for everyone else.
How it works
Step by step
- 1
Describe the data properly
Not the raw tables. A small set of views with business names, descriptions and a few example questions each. This step does most of the work.
- 2
Turn the question into SQL
The model is given the question and only those view descriptions, and writes a query against them.
- 3
Check before running
The query is validated: only approved views, SELECT statements only, a row limit and a timeout. Anything else is rejected.
- 4
Run it as the person asking
The query runs through a read-only login with row-level security, so people only see the rows they are entitled to.
- 5
Show the working
The answer is a table, with the query and a plain-English description of how it was worked out. Questions that fail are reviewed and become new examples.
What it is built from
- SQL Server
- The existing database, exposed through governed views and a read-only login
- Azure OpenAI
- Writes the query and describes it
- .NET API
- Validates the SQL, applies limits, runs it and logs it
- Angular or Teams front end
- The question box and results
- Power BI (optional)
- Lets users pin a useful answer to a dashboard
Safeguards
- Read-only: the model never has a login that can change data
- Allow-list of views, with sensitive columns left out unless the user's role permits them
- Row and time limits so a careless question cannot slow the database
- Every question, query and result count is logged
What you end up with
- A question box for managers, on the web or in Teams
- A documented set of business views over your data
- A query log, and a review list of questions that failed
- Fewer ad hoc report requests reaching IT
Where to start
Pick one area, such as sales orders, and spend the first stage building the views and example questions. The AI part is quick once the data is described well.
More in this group
Cleaning up customer data
Find likely duplicates, standardise messy free-text fields and flag bad records, with a person approving every merge.
Read the design →Forecasting from your own history
Learn from years of orders or jobs to predict next month's workload, and publish the forecast back where your team already looks.
Read the design →