A decision rule starts with data the metric can use
An agent should establish what the dataset measures before calculating efficiency or proposing a bid change. Advertising spend and attributed sales support ACoS. Organic totals do not become ad-attributed results because they appear in the same spreadsheet. The data-mode router labels ads, organic, and mixed inputs before selecting a metric.
ACoS is ad spend divided by attributed ad sales, multiplied by 100. Spend of $250 against $1,000 in attributed sales is 25% ACoS. The denominator must remain attributed sales; replacing it with total business revenue answers a different question.
How do data checks become a bounded bid decision?
The bidding rulebook applies three gates: sufficient evidence, a meaningful distance from the target, and a limit on the resulting change. A deadband around the target prevents small fluctuations from triggering repeated bid movements. A daily clamp bounds the proposed adjustment even when the performance signal is strong.
| Gate | What it prevents | Result when it fails |
|---|---|---|
| Data mode and metric validity | Calculating ad efficiency from unrelated totals | Reject the calculation or route to the appropriate metric |
| Data sufficiency | Treating sparse observations as a reliable signal | Hold and identify the missing evidence |
| Distance from target | Reacting to small variation around the goal | Hold within the deadband |
| Precedence and change limit | Conflicting actions or an oversized adjustment | Apply the governing rule and bound the proposal |
The thresholds are part of the team's rulebook. They should be visible and versioned, not invented afresh for each answer. Mixed datasets keep ads-backed and organic-backed records distinct rather than blending them into one apparently precise efficiency number.
Does a rule firing authorize a live change?
No. A decision plan records the evidence, governing rule, proposed action, and applicable limits. Execution is a separate step under the account's approval policy. A hold is a useful result when data is thin; it should explain what would need to change before the decision can be reconsidered.
These two guides establish data routing and bidding controls, not a universal rulebook for every Amazon operation. Amazon Agent Skills packages repeatable methods, and the Skills reference distinguishes their jobs and approval gates. For campaign strategy, continue to Amazon PPC; for the data and tool setup behind a recurring decision, use agent architecture.