AI budget recommendations are only useful if execution respects account limits. Ad Platform MCP combines natural-language optimization with server-enforced budget safety controls.
Phase 1: Analysis Only
With a read-only API key, ask:
> "Rank campaigns by ROAS last 7 days and list current daily budgets"
> "Which high-ROAS campaigns are budget-capped before 3pm?"
Validate the model's numbers against native UIs once per platform (Meta setup, Google setup).
Phase 2: Constrained Writes
Switch to write-enabled keys with conservative org settings:
Prompt with explicit caps:
> "Increase daily budget by at most 10% on campaigns with ROAS > 3.0 and capped before noon. Skip everything else."
The server rejects over-limit requests even if the model mis-parses "10%" as "10x".
Phase 3: Human Approval
Large reallocations return pending_approval responses. Managers approve in Dashboard → Approvals with full payload context.
What Not to Do
Avoid vague instructions like "maximize spend on winners." Specify metrics, caps, and platforms. Multi-platform moves should name each network to prevent wrong-tool execution.
Agency Note
Per-client workspaces and approval routing are covered in agency multi-client ad MCP.