Copilot does not apply changes to your model autonomously. Every structural change (whether triggered by reverse engineering, a bulk governance action, or a direct prompt) is presented as a proposal first. You review the proposal and approve it before anything lands. This article explains the review screen, when to expect the AI to ask a clarifying question, and how to refine a prompt if the proposal isn’t what you wanted.
Why this matters: The proposal-and-approve pattern is the core safety model underpinning every Copilot use case. It’s why an AI operating on your schema doesn’t create risk of unintended change. Understanding how the flow works makes everything else Copilot does easier to trust.
The proposal-and-approve flow
When you submit a Copilot prompt that would change the model, here’s what happens:
- Copilot reads your prompt and the project context.
- It generates the proposed change: a set of objects to create, modify, or delete.
- The review screen opens, showing the proposed change alongside the current state.
- You review each proposed change. Accept the whole batch, reject the whole batch, or accept individual changes and reject others.
- Only the accepted changes are applied to the model. Rejected changes are discarded. The original prompt remains in the conversation so you can refine and try again.
No change to your model happens without you clicking Accept. Copilot cannot commit changes in the background, cannot act on a schedule, and cannot be delegated authority to apply changes without review.
What the review screen shows
The review screen layout depends on what kind of change was proposed:
- Single object change. Before-and-after view of the object. Differences are highlighted.
- Bulk change (many objects). A list of every proposed change with a checkbox next to each. Unchecked items won’t be applied.
- New objects. A diagram preview of what will be added to the project, including new tables, columns, and relationships.
- Deletion proposals. The objects to be deleted are listed with any dependent objects called out explicitly. If you accept a deletion that breaks references, Copilot warns you before applying.
Refining a prompt that didn’t land
If the proposal doesn’t match what you wanted, you have three options:
Option 1: refine the prompt in a follow-up. Cancel the proposal, then ask a follow-up in the same conversation. Copilot retains the prior context, so you don’t have to restart. Example: the first prompt created tables but didn’t wire up foreign keys. Follow up with “add foreign key relationships between these tables based on the naming convention.”
Option 2: accept partially and refine the rest. Accept the correct changes, reject the rest, then prompt Copilot to redo the rejected parts differently. This is often faster than starting over.
Option 3: reject and restart with a better prompt. If the proposal is so far off that refinement would take longer than restating the prompt, reject everything and write a more specific prompt. Be explicit about what the prior attempt got wrong.
When Copilot will ask a clarifying question
For prompts where intent is ambiguous, Copilot will ask a clarifying question before proposing a change. Common examples:
- “There are three tables matching your description. Which one did you mean?”
- “Should I create the new column in every dimension table, or only in the ones you specified?”
- “This change would break a reference in another table. Do you want me to update the dependent table too?”
Clarifying questions are a good sign. They prevent the AI from guessing and producing a wrong proposal. Answer specifically, and Copilot will proceed to a better proposal.
What Copilot will never do
- Apply changes without review. Every structural change requires explicit acceptance.
- Exceed your permissions. If your SqlDBM role doesn’t allow editing, Copilot can propose a change but you cannot accept it. Propose-without-accept does not commit.
- Modify objects outside the current project. Copilot’s scope is bounded to the project you’re working in today. Cross-project changes are not possible; cross-project awareness is planned for a future release.
- Ignore your pre-prompt or global standards. Standards defined at the project level are sent along with every prompt. Copilot follows them by default.
Advanced: working without proposals (read-only questions)
Not every Copilot interaction triggers a proposal. Questions about the model, “explain this table,” “where is this column used,” “summarize this project”, produce read-only responses that don’t touch the schema. These conversations are safe to run regardless of your role. See “Talk to your model” for the full pattern.
Related articles
- Getting started with Copilot
- Talk to your model
- Bulk governance actions
- Reverse engineering a model with Copilot