Your AI Runs Out Before the Month Does

Somewhere out there, a small company has a support bot on its website, and one afternoon a customer typed a question and got this back:

"This agent is currently unavailable. It has reached its usage limit."

That is not a guess. It is the exact sentence Microsoft publishes in its own documentation as what your customers see when your AI agent runs past what you paid for. Written down, in a table, next to the second option: "There is a billing issue."

Nobody gets a warning in the moment. The agent finishes the conversation it is in, then stops taking new ones. An email goes to whoever is listed as the tenant admin. That is it.

I keep meeting owners who budget for AI the way they budget for phone lines. Pick a plan, pay it, forget it. That model is being replaced, and most people have not noticed, because the change did not arrive as an announcement. It arrived as a rate table.

The unit of AI is no longer a seat

For twenty years, business software was priced per person. Ten employees, ten licenses, one predictable number at the bottom of the invoice.

AI does not price that way, because AI does not cost the vendor that way. Somebody who asks one question a week and somebody who wires up an agent to process every order overnight are not the same customer, even at the same size desk.

So the unit moved. Microsoft sells Copilot Studio in packs of 25,000 Copilot Credits at $200 a pack per month, or on a pay-as-you-go meter with no commitment at all. Credits are what you buy. Seats are not in the sentence.

That is not a Microsoft quirk. Once you see it, you start noticing the same words on every pricing page you open: credits, tokens, tasks, actions, capacity, runs.

Not every answer costs the same

Microsoft publishes a rate card. A canned, pre-written answer costs 1 credit. An AI-generated answer costs 2. An action the agent actually takes costs 5. Grounding an answer in your company's own Microsoft 365 data costs 10. And the premium reasoning tools, the deep-thinking ones, bill at 100 credits per ten responses.

Read that ordering again. The cheapest thing your AI can do is repeat a script. The most expensive thing it can do is think carefully using your own information. Which is, of course, the entire reason you bought it.

Microsoft's own worked example is a website support agent handling four scripted answers and two generated ones per visit, at 900 customers a day. That comes to 7,200 credits a day.

Scale that down to something recognizable. Say you are smaller, 60 conversations a day, same mix. Eight credits a conversation, 480 a day, roughly 14,400 a month. One $200 pack covers it with room to spare, and you would tell your bookkeeper AI costs you two hundred bucks.

Then, six weeks later, somebody turns on the setting that lets the agent ground its answers in your company's own files, because the generic replies were not good enough. Now each of those two generated responses costs 12 credits instead of 2. Same 60 conversations, same customers, same business, and the monthly number is about three and a half times what it was.

Nothing broke. Nobody made a mistake. Somebody just made the tool better.

The shutoff is already written down

Microsoft's published enforcement policy kicks in when a tenant hits 125% of its prepaid capacity. At that point custom agents are disabled. A conversation already in progress finishes normally. Every attempt after that is rejected, and the person on the other end gets one of those two sentences up top.

There is a grace band built in, which is decent of them. But the shape matters more than the number. The failure is silent on your side and loud on the customer's side. You find out from an email in an admin inbox. They find out from the thing on your website that stopped working.

If your AI answers your phone, greets people on your site, or triages your inbox, that is not a billing problem. That is your front door.

The flat fee has a meter inside it too

You might be reading this thinking none of it applies, because you do not build agents. You just pay a flat monthly price for software you already had. Look closer at that software.

On August 1, Intuit changed monthly pricing for QuickBooks Online Essentials, Plus, and Advanced, and said plainly that the increase supports its new AI capabilities. Fine, prices move. But read the footnotes on those features. Invoicing on Autopilot is in beta and, in Intuit's words, "subject to monthly usage capacity limits." The plain-language business questions feature comes with "limited inquiries available each month, depending on your QuickBooks plan." The AI chat gets "limited capacity each month."

So you are paying a higher flat fee, and the AI inside that flat fee is still rationed. The meter did not go away. It moved behind the curtain, where you cannot watch it.

Four things worth doing this month

None of this is a reason to back off AI. It is a reason to run it like something that costs money per use, because it does now.

Find out which of your tools are metered. Open the pricing page for every AI-touching tool you pay for and look for the words credits, tokens, tasks, actions, or capacity. Ten minutes per tool. Most owners are surprised by at least one.

Ask what happens when you hit the ceiling. There are only three answers: it queues, it degrades, or it goes dark. You need to know which one, especially for anything a customer can see. Vendors will tell you if you ask.

Set your own cap before the vendor sets one for you. Microsoft, for example, lets you cap monthly credit use per agent in the admin center. A ceiling you chose is a budget. A ceiling you did not choose is an outage.

Make sure the overage email lands on a human. Not a shared inbox, not the address you set up in 2019. Somebody who checks mail on a Saturday.

Price it per job, not per month

The real shift here is not that AI got more expensive. In most cases it got cheaper per unit of work. The shift is that the cost now moves with your volume, which means your AI bill behaves like a cost of goods sold rather than a subscription.

That is actually good news, once you treat it that way. Run the same arithmetic on that 25,000-credit pack and one eight-credit customer conversation lands at about six cents. Six cents is a number you can hold up against what that conversation is worth to you. A flat $200 a month tells you nothing at all.

The owners who get burned this year will not be the ones who spent too much. They will be the ones who never knew there was a meter, until a customer told them.

If you want a straight answer on what your AI stack actually costs you per job, and where it is going to break, that is a conversation we have with clients most weeks. Come find me at nexeraintelligence.com.