Limiting an AI budget: the surprise bill now has a stop button

On August 31, Vercel introduced per-user budgets that reject requests once the limit is reached, with alerts at 50%, 75%, and 100%. A spending cap is becoming something you require, just like a password.

Limiting an AI budget before the bill arrives is now a feature providers offer. On August 31, 2026, Vercel launched per-user budgets for its AI gateway, rejecting new requests once the limit is reached, with email alerts at 50%, 75%, and 100% (Vercel changelog, August 31, 2026).

What the cap does and does not do

The mechanism is simple: each user receives an amount, consumption increments, and requests beyond it are rejected instead of billed. Reset is monthly by default, configurable as daily, weekly, or never. The provider imposes no amount. A crucial detail is in the announcement: per-user budgets do not replace existing per-key, per-project, or per-team budgets. An individual cap does not prevent ten people spending ten times that limit in one day; an organisation setting only that cap would have misread the page.

Why the counter surprises even when you watch it

Agent-work billing resembles no subscription invoice. The same request can cost twice as much depending on how much the agent rereads before responding, and failed generation costs as much as success. Unit prices fall while consumption rises, detailed in token prices fall, bills rise. Then there is the expense nobody budgets: rebuilding the same feature three times because nothing fixed what it should do, whose scale we showed in $2,000 of vibe coding and zero users.

A cap does not replace scope

A spending limit stops the bleeding without curing the cause. Excessive project consumption has little to do with model prices: scope moves, undecided matters are settled afresh by an agent every morning, and missing verification allows work through that must be redone. A monthly budget on a poorly scoped project turns overspending into a service interruption on day twenty-four. Both actions belong together: a cutoff cap and an approved document stating what should be built.

How we put the number before the spending

Maestro applies that logic inside the product. Before construction starts, an estimate appears, and a cap based on it stops work when exceeded instead of silently continuing. A monthly budget covers all projects. A queue prevents several projects consuming in parallel without anyone watching. The principle fits in one sentence: a visible number before spending, never afterward. Our campaign amounts are published to the cent in our bills, line by line, with calculation methods, because an unsupported number merely asks to be believed.

Three questions for a provider

Before entrusting a project to a usage-billed tool, require three written answers. Where is the cap set, and at which level: user, project, team, key? What happens when it is reached: rejection, queueing, or billed overage? Who receives alerts and at which threshold? A provider answering all three lets you govern spending; one displaying only an undefined credit counter asks for trust, which we have seen end badly with several tools, including those examined in Airtable and Notion's limits.

Set your cap tonight

Open the billing page for every AI tool you pay for and search for 'budget' or 'limit'. Set an amount you would accept losing without argument, choose the lowest available alert threshold, and schedule a consumption review in one month. Those fifteen minutes give the only number that matters next: what your actual usage costs when nobody has constrained it.