A billing alert is not a spending limit
Count the services your business pays for by usage. Cloud hosting, almost certainly. An AI model you call from your own software. Text messages, maps, email delivery, transcription. Each one has a meter, and none of them has a ceiling unless you built one.
What most people have built is an alert. When spending passes a threshold, an email goes out. If that email lands at midnight on a Saturday, the meter keeps running until somebody reads it on Monday.
That gap has always existed. Two things have made it worth closing now.
Why this got more urgent
The first is AI. The developer Simon Willison made the case in October that hard budget caps need to become the default on "pretty much everything". His reasoning is that coding assistants have made it very easy to set running a piece of software that calls paid services, and very easy not to notice what it is doing. As he puts it: "Soft caps, 'after $X/month, send me a warning email', will not cut it."
AI billing is also less predictable than what came before. Google's own announcement of its new cost controls says so directly: a short prompt can set off a long chain of work, and the old ways of estimating a bill no longer hold.
The second is that the big providers have finally started offering a real off switch.
What AWS and Google now offer
AWS spend limits. Amazon Web Services announced these in September. You set the most you will pay per month for a project. According to the AWS documentation, if usage reaches the limit, "AWS pauses that project which stops its resources so your costs stay within it". You get warnings at 50%, 75%, and 90%. The small print matters:
- AWS is still releasing the feature to a limited number of customers, so you may not have it yet.
- The lowest limit you can set is $20 or AWS's estimate of your likely spend, whichever is higher.
- AWS describes it as designed for experiments and test environments. It can be used in production only where "a brief pause of your resources" is acceptable.
Google Cloud Spend Caps. Launched in July as a public preview. You set a monthly cap on one service within one project. When spending reaches it, Google blocks further chargeable use of that service until somebody lifts the cap by hand. Nothing is deleted. Google says caps on AI services act within minutes. Fixed commitments you have already signed up to carry on billing regardless.
Both are narrower than the headline suggests. Both are still a real change. For years the honest answer to "how do I stop the bill passing a set amount" was that you had to build it yourself.
The decision behind the setting
A cap is a choice between two failures. Either the service stops, or the bill grows. Neither provider can make that choice for you, and it has a different answer for different things.
A test environment, a prototype, or an internal tool that somebody set up with an AI assistant should stop. A surprise bill there buys you nothing. Your live shop on the busiest day of the year should probably not stop, and wants an alert that reaches a human who can act.
So the useful work is sorting what you run into those two piles. Most businesses have never done it, which means everything is in the second pile by default, including the things nobody is watching.
What to do
- List everything billed by usage. Include the AI services and the small ones on somebody's company card.
- Sort each into "should stop" or "must stay up". Be honest about how little belongs in the second group.
- Cap the first group. Use the provider's hard limit where one exists.
- Where there is no cap, improvise one. Many AI providers let you buy credit in advance and turn off automatic top-up, which is a hard cap by another name. Others let you set a monthly limit on each access key. Use both.
- Keep experiments in their own account or project. A cap on a separate project cannot take the live system down with it.
- Give each automated tool its own key and its own limit. If an AI assistant is making paid calls for you, it should not be spending from the same pot as production.
- For what must stay up, fix the alert. It should go to more than one person, by a route that works out of hours, and say who is allowed to act on it.
- Ask your developers or IT provider one question. "What is the most we could be charged this month if something went wrong, and what would stop it?" If nobody can answer, that is the finding.
How Steelwise can help
Working out what you are metered for, what should have a ceiling, and what happens when one is reached is part of our infrastructure advisory work. Get in touch.