What should an AI recommendation contain?
A reviewable recommendation exposes its evidence and limitations. It identifies the affected account or campaign, reporting period, freshness, observed change, rationale, proposed action, expected learning window, and required permission. A recommendation that cannot link back to source evidence should not be approved.
The system should also state uncertainty. Marketing data can be delayed, modeled, duplicated, misconfigured, or influenced by factors outside the connected platforms. Fluent language does not make a causal claim reliable.
Which checks belong in the human approval gate?
The approval gate should scale with potential harm. Reading a report and pausing a high-spend campaign do not require the same control. Management-capable provider permissions must be visible before a user approves an execution action.
| Risk area | Reviewer question | Safe response when uncertain |
|---|---|---|
| Evidence | Can I verify the source, date, scope, and metric definition? | Refresh or inspect source data |
| Causality | Is the explanation proven or only plausible? | Convert the claim into a hypothesis |
| Business context | Could margin, inventory, capacity, or policy change the decision? | Request the missing context |
| Permission | Will this read data or change a provider account? | Require explicit action approval |
| Reversibility | Can the team undo and observe the change safely? | Reduce scope or use an experiment |
How should recommendations be classified?
Classification prevents a low-risk observation from being confused with a material account action. It also determines which role can approve the next step and what audit record is needed.
- Informational: explains a metric or change; no provider action.
- Investigative: requests a diagnostic check or missing context.
- Experimental: proposes a bounded test with a comparison and decision rule.
- Operational: proposes a reversible workflow or configuration change.
- Material: changes budget, targeting, publishing, access, billing, or another high-impact state.
What is the review workflow?
A consistent workflow is more important than a long checklist. The reviewer should be able to stop the process at any point, ask for evidence, narrow the action, or record a reason for rejection.
- 1
Confirm data freshness, connector health, and metric definitions.
- 2
Read the observation separately from the model explanation.
- 3
Check alternative explanations and business constraints.
- 4
Choose approve, edit, test, defer, or reject.
- 5
For execution, confirm account, scope, exact change, and rollback.
- 6
Review the outcome after the agreed measurement window.
Which failure states should block approval?
Block approval when source data is stale, the affected account is ambiguous, the recommendation exceeds the user role, required context is absent, or rollback is undefined. Also block any recommendation that claims guaranteed performance or conceals that it is based on demo or synthetic data.
Evidence
Official sources
These primary references support the platform-specific definitions used in this Guide. Growomo commentary and workflows remain distinct from provider documentation.