AI for data development: from question to proven change
The wholesaler's nightly processing keeps finishing later. Noor suspects the procedure that merges stock movements, but has no time to measure for three evenings. An AI model can propose a better query. The question is whether it really is better.
Plainwater AI gives agents a data engineer's tools: metadata, quality checks, performance figures and a disposable copy to prove proposals on. A proposal without a measurement is an opinion; here it gets a verdict.
Updated .
What can an agent look at in your database?
An agent reads metadata, quality figures and performance data through more than fifty fixed tools. Every tool has a fixed template and returns no rows in the default mode.
- Metadata: objects, definitions and dependencies between tables, views and procedures.
- Quality: missing values, uniqueness, referential integrity and column profiles.
- Performance: the heaviest queries, missing indexes, index usage, statistics age and procedure history.
- ETL: the log profiles of your loads, so the agent sees what happened overnight.
For Noor this means the agent finds the five heaviest queries of last week by itself, and sees the stock procedure appears twice.
How much may an agent do per assignment?
Per assignment you choose a work mode, and the gateway enforces it on every tool call. The mode sits on the assignment, not in the prompt.
| Work mode | What the agent does | What the source needs |
|---|---|---|
| Look only | Read metadata, counts and history | A read-only login |
| Measure on source | On top: the estimated plan and real execution counters | Extra rights to show plans |
| Prove on a disposable copy | Read the source, run and measure the proposal on a copy | Consent for what the copy may contain |
How does an agent prove its proposal is better?
The agent runs its proposal in a verification sandbox: a disposable copy of the database where it measures and compares against the original. The outcome is a verdict, not an impression.
- Create a copy. The platform builds a copy with the schema and the data the customer allows.
- Apply the proposal. The script comes from the assignment itself, with a fixed fingerprint.
- Measure. Duration, reads and result of the old and the new version.
- Compare. Are the results identical and is the new version faster? Then a verdict follows, with the figures.
- Clean up. The copy is always removed, also when the assignment stops or fails.
The customer chooses what goes into the copy: synthetic data, synthetic data shaped to the real statistics, or real data.
May an agent write its own SQL?
On the disposable copy yes, and on a customer source only if it has been explicitly released for it. That release is a decision with a reason and a named owner.
On the disposable copy fixed limits apply: one SELECT, a transaction that always rolls back, a cap on columns, rows and time. When in doubt the platform refuses the query.
Writing to and executing on the real source is possible too, as a permission per source. The agent then makes a proposal, and a person approves it before anything changes.
Which sources does it work on?
SQL Server and Azure SQL, PostgreSQL and SAP HANA Cloud are connected natively. Connectors cover the rest of a Microsoft and data landscape.
Connectors: Power BI, Azure Data Factory, Microsoft Fabric, Databricks, dbt Cloud, Matillion, SAP Analytics Cloud, SAP Datasphere, SharePoint, OneDrive, Google Drive, Azure DevOps and GitLab.
If your organisation runs its own MCP server, you connect that as well. Every tool is approved first, and write tools go through a proposal. An agent can split work into subtasks and have partial results reviewed.
Where it stops
A disposable copy with synthetic data proves a query is correct and roughly how it behaves. It does not prove exact production run times; that takes measuring on the source or real data in the copy.
Plainwater AI does not replace your warehouse or ETL. It works on top of them, and the change itself goes to production through your own process.
A verdict is only as good as the measurement. The platform shows the figures with it, so Noor sees for herself what the verdict rests on.
Frequently asked questions
Can an agent change my production database?
Only if that permission is switched on for that source, and then through a proposal a person approves. By default an agent only reads.
What is a verification sandbox?
A disposable copy of the database where the agent runs its proposal, measures and compares it with the original. The copy is always removed afterwards.
Which data goes into the disposable copy?
The customer decides per source: synthetic, synthetic shaped to the real statistics, or real data. Without consent there is no copy.
Does it work with Power BI and Data Factory?
Yes, through connectors. An agent can, for example, read semantic models in Power BI and edit definitions.
Can I connect my own tools?
Yes, through your own MCP server. Every tool is approved first, and write actions go through a proposal.