How do we avoid being surprised?

Establish three things while nothing is urgent: what the rate above the allowance is, whether the product can alert somebody at a threshold, and whether it can attribute usage to a person or process. All three are available in most products and none is useful for the first time after an invoice has arrived.

Why is the rate worse above the line?

Because included capacity is bundled and additional capacity is sold singly, which is ordinary commercial practice rather than a penalty. The practical consequence is that sustained overage costs more than the larger plan would have, so a pattern of exceeding the allowance is a purchasing signal rather than a usage problem.

Should we set a hard cap?

It depends what stops. For exploratory or optional work a cap is sensible protection; for something people depend on it means the tool fails mid-task with nobody warned. The question worth answering is what the tool is doing, and the choice belongs to you rather than to whatever the default happens to be.

Whose budget does it come out of?

Usually whoever holds the plan, which is frequently not whoever consumed it. A shared allowance drained by one team's automation arrives against somebody else's budget, and without usage attribution the conversation has no facts in it. Asking whether the product can break usage down is worth doing before you need the answer.

Glossary

Overage

Overage is what you pay once a plan's included allowance is used up, which is where most unexpected AI invoices come from.

In plain terms

The plan included a certain amount, you used more than that, and the extra is charged separately. Sometimes at a stated rate, sometimes by moving you to a larger plan, sometimes by simply stopping. What makes it the commonest surprise is that going over rarely feels like an event: nobody chose to exceed anything, the work just continued as usual.

Why it matters

Because it is the one charge that arrives without a decision behind it. Every other cost was agreed by somebody at some point; overage accrues from ordinary work continuing past a line nobody was watching. That makes it hard to attribute afterwards and easy to prevent beforehand, and the preventive step is a question rather than a control.

How it works

Ask what the rate above the allowance is, and compare it against the effective rate inside. Included capacity is sold as a bundle and additional capacity is sold singly, so the second is often the less favourable of the two, which is the shape people find counter-intuitive. That is ordinary commercial practice rather than a penalty, and where it holds it makes a plan change cheaper than sustained overage. Ask also whether the charge is capped, because that answer decides your worst case.

Automated work is where it accumulates fastest and is noticed slowest. A process running unattended over incoming records does not stop to consider its budget, and a change in volume upstream can multiply consumption without anybody touching the automation. Attended use is self-limiting because people get bored; unattended use is not.

Alerting is usually available and usually not configured, which is the practical gap. Most products can notify somebody at a threshold, and the notification is worth setting up during calm rather than after the first surprise. Where a product offers a hard cap as well, the choice between stopping and continuing to charge is a real decision that belongs to you rather than to a default.

The first response should be a plan review rather than a cost control. Sustained overage usually means the plan is the wrong size, and paying single-unit rates every month to stay on a cheaper plan is the expensive way to do it. Occasional overage is different and is often the correct outcome, since sizing a plan for a rare peak means paying for that peak all year.

Who is charged and who caused it are frequently different people, and that gap does the organisational damage. A shared allowance consumed by one team's automation arrives as a bill against a budget somebody else holds, and without a breakdown by user or by process the conversation becomes an argument. Asking whether the product can attribute usage is worth doing before it matters.

Two ways to reach the same overage

It is worth separating these two because organisations tend to respond to overage with approval processes, which address the left-hand path exclusively. Requiring sign-off for new AI work does nothing about a running automation whose input volume increased, and the second is where the surprises actually come from. What does work is unglamorous: a threshold alert configured while nothing is wrong, a known answer to what the rate above the line is, and the ability to attribute usage to a process so the conversation afterwards has facts in it. None of those constrains anybody's decisions, which is precisely why they suit a cost that nobody decided to incur.

Seen in the wild

An automation whose volume rose upstream, consuming a shared allowance without anybody changing the automation itself.
Make
A workflow calling a model on every record, where a quiet increase in records is invisible until the invoice.
n8n
Paying per request against a balance, where there is no allowance to exceed and the cost is simply continuous.
OpenRouter

Common misconceptions

People assume

We will notice before it matters.

In fact

Attended use is self-limiting and unattended use is not, and most overage comes from the second. A process running over incoming records reflects a change in volume rather than a change in behaviour, so there is nothing for anybody to notice until the invoice arrives with the period already over.

People assume

Overage means we should use less.

In fact

Sometimes, and more often it means the plan is the wrong size. Paying single-unit rates every month to stay on a cheaper plan costs more than moving up, so persistent overage is a purchasing signal. Occasional overage is a different thing and is frequently the right outcome, since sizing for a rare peak means paying for it all year.

People assume

A hard cap is the safe default.

In fact

It is safe for the budget and not for the work: a cap on something people depend on means the tool stops mid-task with no warning to whoever was relying on it. Whether stopping or charging is worse depends entirely on what the tool is doing, which is why it is a decision rather than a default.

Telling them apart

Overage vs Allowance

Overage

What happens past the line. Accrues from ordinary work continuing.

Allowance

Where the line is. Chosen when the plan was bought.

Nobody decides to go into overage, which is exactly why it is the surprising one.

Questions

How do we avoid being surprised?: Establish three things while nothing is urgent: what the rate above the allowance is, whether the product can alert somebody at a threshold, and whether it can attribute usage to a person or process. All three are available in most products and none is useful for the first time after an invoice has arrived.
Why is the rate worse above the line?: Because included capacity is bundled and additional capacity is sold singly, which is ordinary commercial practice rather than a penalty. The practical consequence is that sustained overage costs more than the larger plan would have, so a pattern of exceeding the allowance is a purchasing signal rather than a usage problem.
Should we set a hard cap?: It depends what stops. For exploratory or optional work a cap is sensible protection; for something people depend on it means the tool fails mid-task with nobody warned. The question worth answering is what the tool is doing, and the choice belongs to you rather than to whatever the default happens to be.
Whose budget does it come out of?: Usually whoever holds the plan, which is frequently not whoever consumed it. A shared allowance drained by one team's automation arrives against somebody else's budget, and without usage attribution the conversation has no facts in it. Asking whether the product can break usage down is worth doing before you need the answer.

Key takeaways

It is the one charge that arrives without anybody having decided anything.
Unattended work accumulates it fastest and reveals it slowest.
Sustained overage is a purchasing signal; occasional overage is often correct.
A hard cap protects the budget by letting the work fail, which is a real choice.