The hidden cost of rule-based automation

Rule-based automation is popular because it is legible. You can read a rule. You can explain it to a client. If ACOS exceeds 35%, lower the bid 10%. Nobody is confused.

That legibility is exactly what hides the cost.

A rule is a frozen opinion

When you wrote that rule, you knew things. You knew your margin. You knew what a click was worth in that season. You knew which competitors were bidding. The rule compressed all of it into one number: 35%.

Then the context moved, quietly:

• Your cost of goods went up 8%. The break-even ACOS is no longer 35%.
• Q4 arrived. Conversion rates rose, so a 40% ACOS is now fine.
• A competitor exited. Clicks got cheaper and the threshold is too conservative.
• You launched a variant. It needs a bad ACOS on purpose.

The rule doesn't know about any of this. It executes with the same confidence it had on the day you wrote it — which is the problem. A rule cannot tell you it has become wrong. It has no mechanism for doubt.

Four costs that never show up in a report

1. The threshold cliff. At 34.9% ACOS nothing happens. At 35.1% you cut 10%. Two nearly identical keywords get opposite treatment because a continuous reality met a discrete rule.

2. No sense of trade-off. Lower the bid 10% — versus what? Versus raising it? Versus negating one search term underneath it? Versus doing nothing? A rule has one response. It never considers alternatives, so it cannot choose the best one.

3. Rules that fight each other. Your ACOS rule lowers a bid. Your impression-share rule raises it. Your budget rule caps the campaign. Each is individually reasonable. Together they oscillate, and the account pays for the argument.

4. No accountability. The deepest cost. A rule cannot tell you whether it was right, because it never made a claim. It had no expectation to compare against. You cannot audit a rule's judgement — only its execution.

The maintenance treadmill

The standard answer is: review the rules. Every quarter, revisit thresholds, adjust for seasonality, re-derive break-even from current margins.

Almost nobody does this. Not from laziness — because it's genuinely hard, requires data across systems, and is invisible work. Nothing breaks when you skip it. The account just quietly earns a little less, and there is no alert for slightly wrong for six months.

So rules decay. That decay is the real cost, and it never appears in any report because reports show what happened, not what a better decision would have produced.

What replaces a rule

Not a bigger rule. Not more rules with more conditions — that's the same frozen opinion with more branches.

What replaces a rule is a decision process: read the current context, model what each available action is likely to produce, compare them, pick the strongest, execute inside limits, and then check the prediction against what actually happened.

That last step is what a rule can never do. It's the difference between a system that executes your old assumptions and one that updates them.

Rules aren't evil. They're just frozen. And Amazon isn't.