How the AI's work gets checked

In Short: Nothing gets built until you say yes to the plan, and nothing runs until you've seen the Clean and Ready SQL.

• The Manager shows you a build plan with Approve and Reject buttons before it builds anything.
• For the Clean and Ready layers, you see the SQL that was written. You can approve it, edit it, or reject it before the pipeline runs. Reject, and the agent writes a new version.
• If you edit the SQL yourself, it gets a security check too and must be a read-only SELECT.
• The Data Engineer, Analytics Engineer, and BI Developer pause and ask a question when your request is unclear, then carry on with your answer.

In Short: Generated SQL is tested before it reaches you, and the data is tested after it runs.

Before you see generated SQL, it goes through a security check, a check that the columns and tables it names exist, and a dry run that checks the query will bind.

After the SQL runs, the Quality Engineer's data checks run on the result. A failed check on the Ready layer stops the run. On Raw and Clean, a failed check can show as a warning without stopping the run.

In Short: Every run leaves a record, and Clean-layer SQL keeps its earlier versions.

• Each pipeline run has a record with its status, the layer it reached, the result for each layer, any errors, its metrics, and its start and end times.
• Clean-layer SQL is saved as versioned files with their details, so earlier versions stay on record.

In Short: You can roll a table back to an earlier snapshot, in two steps, with a reason.

Your tables are stored as Apache Iceberg tables, which keep snapshots. Open a table's Snapshots tab, pick an earlier snapshot, and review what the rollback will change. Then confirm and enter a reason of at least 10 characters. Any member of the workspace can do this.

In Short: They make mistakes, and some of those mistakes look fine. This is why the approval steps above exist.

• SQL can run cleanly and still be wrong about what your business means. For example, which customers count as "active". A passing check can't tell you that. A person who knows the business can.
• A model can misread a vague request. Some agents pause to ask, but they can't catch every misreading. Read the plan before you approve it.
• The Analyst answers without an approval step. Treat its answers like a colleague's first pass.
• We don't publish accuracy numbers for the agents yet. Don't trust one from anyone, us included, unless they show how it was measured.
• There is no audit log page in the app today. You have run records, SQL versions, and table snapshots, but not a single browsable history of who did what.
• We aren't certified under SOC 2 or ISO 27001. See the security page.
• Not every source is supported yet. Check the connectors list to see if yours is.

In a BYOC deployment your tables stay in your own cloud account. The security page explains what we store and what you control.

Not sure this fits your team? Read Is OptimaFlo right for you? or ask us directly.