Check the evidence behind an AI advertising recommendation before deciding

A claim-by-claim evidence sheet

Create a claim sheet with columns for statement, source, entity, period, calculation, interpretation and decision relevance. Split a long recommendation into atomic claims. Spend increased is a descriptive claim. The creative caused the increase is a causal claim. Lower the budget is a decision. They need different support, and a citation attached to the paragraph does not automatically support all three. Identify exactly which source values establish each observation.

For a teaching example, an assistant recommends pausing Campaign A because its CPA doubled. The quoted report actually covers Campaign B, and the recent conversions are marked unavailable. The review fails before the proposed action is evaluated: wrong entity and missing denominator. Another assistant might have the right campaign and arithmetic but attribute the change entirely to an audience revision despite a simultaneous price change. That answer passes source identity and fails causal certainty.

In AdAce Ads, AI responses can refer to tool-derived account data, while saved performance and period comparisons provide values to check. Those references are starting points, not proof that a business recommendation is valid. Verify the client and account, date boundary, chosen event, currency and data completeness. If the tool output was preliminary or unavailable, the explanation must carry that limitation forward. A confident conclusion cannot repair incomplete inputs.

Arithmetic and judgment need different checks

• Separate reproducible arithmetic from commercial judgment. A correct CPA calculation does not determine the acceptable CPA, sales capacity or marginal value of more leads. Those require agreed business definitions and additional evidence.
• Require credible alternatives for an asserted cause. Conversion delay, tracking changes, qualification updates and overlapping interventions can explain the same surface pattern. A review should state which alternatives were checked and which remain open.
• Use hard rejection criteria: fabricated sources, mismatched client scope, impossible totals or attempted writes during a read-only task. Do not average these failures into a generally good score because the surrounding prose is useful.

Review before considering action

1. Start the AI task as read-only analysis with text output. Explicitly prohibit propose_change and all mutating tools. Creating an eligible proposal may trigger an automatic policy, so evidence review must remain outside the execution path.
2. Extract the material claims from the answer. Ask which claim, if wrong, would change the decision. Prioritize those for verification and record a source reference for each rather than asking the assistant to repeat its conclusion more confidently.
3. Reproduce the essential calculations from saved values. Check compatible units, periods and events. Keep missing values distinct from zero and avoid adding cross-platform conversions as if they were unique business outcomes.
4. Review inference strength. Label each claim observed, calculated, plausible or unverified. Check whether the recommended action follows from the evidence and whether the account's current state still matches the state described in the answer.
5. Choose accept for planning, request more evidence or reject. Record the unresolved questions, reviewer and date. Any later account action requires a separate authorized decision with the relevant current-state and policy checks.

Sources do not automatically prove causality

Sources and further reading

• AdAce Ads AI analyst
• AdAce Ads daily performance
• AdAce Ads write policies