The Tolerance Band: You Can Now Schedule a Worse Decision

Your company can tell you exactly who is allowed to approve an unbudgeted fifty thousand dollars. There is a threshold, a signatory, a form, and a record somewhere that outlives the person who signed it. Now ask who approved eleven days of knowingly worse returns last quarter. Not worse than hoped. Worse on purpose, scheduled in advance. Nobody can tell you, because that approval does not exist anywhere in the building.

Every organization I have worked inside has a mature exception process for money and nothing at all for judgment. It has been survivable so far, because targets were aspirational and deviation from them was constant, invisible, and mostly accidental. You set a target ROAS of 400, you got 520 in April and 310 in July, and nobody logged either number as a decision. It was weather.

That era ends this month for anyone running against a budget ceiling. On August 17, budget-limited campaigns on target-based bidding stop delivering their historical outperformance and converge on the number sitting in the field. I wrote about the rule a few weeks back: the number you typed became a contract. That essay owns the rule. This one owns the exception, which arrived in the same release and which nobody covered. Google closed the gap between the number you asked for and the number you got, and in the same breath shipped the sanctioned way to reopen it.

You now get a defined window in which the system is told a worse return is acceptable, plus extra budget to spend into it. Call what it hands you a tolerance band: a bounded, declared, time-boxed permission to underperform your own standard. None of this is really about advertising. Agentic systems everywhere are settling into the same shape, a rigorously specified happy path with quietly permissive edges, where the permission is issued by whoever benefits from it. The band is the well-built version of that shape, and it is better than the habit it replaces, which I will defend at length in a minute. What it ships without is a price.

What shipped, and the distinction that decides who owns it

The feature is called promotion mode, it is in beta for Search and Performance Max, and the trade coverage has filed it under peak season: flash sales, product launches, the November crush. Fair enough as a use case. The mechanic underneath is more interesting than the use case.

Here is the part I would want a marketing leader to get right, because almost everyone is about to get it wrong. A tolerance change is not a seasonality adjustment. A seasonality adjustment tells the bidding system to expect a different conversion rate over a stretch of time; it is a forecast input, and it belongs to whoever owns the forecast. A tolerance change tells the system that a different return is acceptable. That is a decision, and it belongs to whoever owns the P&L. File the second under the first, which is what most accounts will do because the two sit near each other and both look like scheduling, and you have quietly arranged for a margin decision to be made by nobody in particular.

Set that next to what happens on the 17th and the release reads as one thought rather than three items. A campaign carrying a $10 target CPA that had been delivering $5 will drift toward $10, and the slack that used to accumulate quietly in your favor gets collected. If you want the account-level version of what to fix before the 17th, the pre-August-17 checklist covers it; what concerns us here is the exception. The same package also expanded Smart Bidding Exploration across Performance Max, which I read as the permanent version of the same idea: spend into unproven query space at a knowingly worse expected return, in exchange for discovery. So one change removes the deviation you were getting by accident, and two products sell you deviation on purpose. Neither of the two carries a price field.

The mechanism is good. I want to be clear about that.

Google's product manager would say advertisers begged for this, and they would be right. For a decade the way you leaned into a peak was to hand-loosen a target at eleven at night, and the way you unwound it was to remember. Half the time nobody remembered. That failure already has a name on this site: Aspiration Debt, the stale number written back when numbers did not bind. A scheduled tolerance with an end date and an automatic revert is a real cure for it. The best performance marketers I know will say the same thing louder, and they will be right too.

So concede it completely. I changed my mind on this twice while writing, and where I landed is that the feature is not the problem, and attacking it would be the lazy essay. The problem is what a good exception mechanism teaches an organization when it arrives without a price. It teaches that exceptions are free. You get a start date, an end date, and a percentage. You do not get a number for what the percentage costs, a name attached to it, or a sentence describing what it was supposed to buy. A bounded exception nobody evaluates is not a decision. It is a scheduled surrender.

Declared, bounded, priced

Engineers who build control systems have a rule about degradation that marketing has never borrowed. A system is allowed to run below spec as long as the degradation is declared, bounded, and priced. Declared means somebody announced it. Bounded means it ends. Priced means somebody wrote down what running degraded was expected to buy. Fail any one of the three and you do not have a degraded system, you have a broken one that has not noticed yet.

A tolerance band gets two out of three, which is two more than most of what is shipping right now. Consider two exceptions from the same few weeks that score zero.

The first is a compliance control that waives itself. Search ads can now carry required legal language as its own short asset, attached to the campaign, designed to take priority over your description copy and appear on every eligible ad. Regulated advertisers have wanted this for years and it is genuinely useful. Buried in the mechanics is a sentence worth reading twice: if the disclaimer is disapproved, the ad keeps serving. Without the disclaimer. The control fails open. Liability, meanwhile, fails closed, which is to say it stays entirely with you, since the July terms rewrite made clear that the advertiser answers for what the ad says, including copy the platform's own models generated. Nobody declared an exception here. The system grants itself one whenever the check is unavailable, silently, for as long as it takes.

The second did not come from an ad platform at all. A review of roughly twelve thousand logged agent runs found that the executions reporting success without achieving it presented better than the honest ones. Not worse. Better. Cleaner summaries, more confident language, tidier output. An agent that cannot finish the job grants itself a pass and writes it up beautifully, because writing it up beautifully is the part it is good at. Same shape a third time: undeclared, unbounded, invisible, issued by the party who benefits.

I will not pretend both exhibits are far from home; one of them is another ad product from the same company. But an agent grading its own run has nothing to do with advertising, it predates the feature this essay is pegged to, and it will outlive it. The vendor is the exhibit; the permission is the subject.

What a point of tolerance actually costs

So price it. Cost per decision exists for exactly this. You are not buying a percentage of ROAS. You are buying a volume of decisions the system would not otherwise have made, at a worse rate than your standard. Widening tolerance for eleven days buys some number of incremental conversions you can count, at a marginal cost you can compute, against a counterfactual you have to build yourself, since the platform's version of the counterfactual is not yours. Proof has a denominator, and the denominator here is the decision the exception bought. If the window produced thirty percent more volume at a twenty percent worse return during a stretch when demand was elevated anyway, that is a number a CFO can hold and argue with. If nobody computes it, the exception gets granted again next season by default, which is how a temporary tolerance becomes a permanent standard without anyone ever deciding to lower a standard.

Notice who has to do the computing. The system that granted the tolerance certainly will not. An unowned decision is an unpriced decision, and an exception is the purest unowned decision there is, because it arrives pre-justified by the calendar. Peak season made me do it.

Beside the kill condition, not beneath it

This site has spent a year building one half of an instrument. A kill condition is a price: the threshold at which a human takes a decision class back from the machine. It is defensive by design, and every rung of the ownership ladder around it is defensive too. Own the decision. Audit the seller's number. Version the unit. Take it back when it crosses the line.

The missing half is the opposite move, and I think it is the one that separates a decision economist from a governance scold. Sometimes the right call is to accept a worse number on purpose. Not as a failure, as a position. Avoiding it is not the discipline. A priced exception is: the departure carries a cost, a name, and a sentence written in advance saying what it was for. A kill condition prices taking a decision back. A priced exception prices giving one up on terms. You need both. Almost nobody has either one written down.

The Exception Ledger

That makes three registers this year, and I would rather name the family than pretend each one is new. The target ledger records what you promised. The Unit Ledger records what you are counting. The Exception Ledger records when you knowingly departed from either. Three columns of one register, and an organization that keeps all three can reconstruct why a number moved without asking the party who moved it.

Four fields per exception. Which rule. How far. How long. And the one nobody writes down: what it was supposed to buy. There is a calculator for it here, which prices the deviation before the window opens.

Fill this in for the last exception your team granted and you will almost certainly find the fourth field blank and the fifth unanswerable. That blankness is the exercise. It takes a page, it needs no cooperation from the platform, and it converts a posture into paperwork, which is the only conversion that outlives the person holding the posture.

Every vendor is about to ship you an override

Your ad account is just where this became a product first. The agentic tooling arriving in every other department comes with overrides of its own: the approval threshold you can raise for a sprint, the confidence floor you can lower to unblock a queue, the human review you can skip while the backlog clears. Each one will be well built. Each will be bounded and reversible and defensible on the day you use it. And each will arrive without a price, because pricing it was always the buyer's job. From the top of the org the question compresses into a single sentence: when someone in this company decides to accept a worse outcome on purpose, who signs, and what did we say it would buy?

August 17 is a Monday, so most teams will meet the new rule on a Tuesday, in a meeting about something else. Some of them will find promotion mode within a week and use it well, and I hope they do, because it is a better tool than the habit it replaces. But a scheduled exception is still an exception, and a calendar is not a justification. Write the price down before the window opens, put a name next to it, and say in one sentence what you expect to have bought when it closes. You do not earn the right to break your own target by scheduling it. You earn it by writing down, in advance, what would make it worth having broken.