I was sitting in a meeting recently where we were walking through AI credits and how they get consumed across the Microsoft stack. At some point we asked the obvious question: what does a credit actually cost in real money? The answer came back at roughly a penny per credit.
Everyone nodded and moved on. I didn’t. Something about the whole structure felt weirdly familiar, and it bugged me for the rest of the meeting until it finally clicked.
Text messages.
If you’re old enough to remember flipping open a phone and pecking out a message on a T9 keypad, you probably also remember the bill. Every text sent cost you. Every text received cost you. You could buy a bucket of messages for the month and share it across the family plan, and if your teenager blew through it by the 15th, everyone paid for it. Fifteen-plus years later, here we are again, except now the thing being metered is AI.
They literally called them messages
Here’s the part that made me laugh out loud when I went digging afterward. My comparison isn’t just a vibe. Until fairly recently, Microsoft’s unit of consumption for Copilot Studio agents was literally called a “message.”
On September 1, 2025, Microsoft renamed that currency from messages to Copilot Credits. The pack sizes didn’t change. The pay-as-you-go rate didn’t change. A 25,000-message pack simply became a 25,000-credit pack. Some admin pages still carry “messages” in their URLs, like a fossil left behind from the old naming.
So when my brain said “this is text messaging,” it wasn’t making a stretch. It was reading the original label under the new sticker.
The family plan, reborn
Once you see it, the parallels are hard to unsee.
Per-unit metering. Pay-as-you-go Copilot Credits run $0.01 each, billed monthly through an Azure subscription. That’s the old ten-cents-a-text model with a smaller decimal.
The bucket. If you’d rather prepay, a capacity pack runs $200 a month for 25,000 credits, which works out to about $0.008 per credit. That’s your monthly text bundle, priced to make the meter look expensive by comparison.
The shared plan. Those credits are pooled across the entire tenant. Every agent, every department, one shared bucket. In other words, the family plan, with IT playing the role of the parent staring at the bill.
Use it or lose it. Unused credits don’t roll over to the next month. Carriers eventually turned rollover into a selling point. We’re not there yet.
The overage math. Prepaying is about 20% cheaper per credit, but only if you actually use the pack. Underuse it and your effective rate climbs above the pay-as-you-go price. Anyone who ever bought the 1,000-text plan “just in case” and sent 300 knows exactly how that feels.
Where the analogy breaks (and why it matters)
Here’s where it gets more complicated than the old cell bill. A text was a text. Whether you sent “k” or a novel, it cost the same.
Credits don’t work that way. A simple scripted agent response might cost 1 credit. A generative answer costs 2. Heavier prompts that pull in more context and do more reasoning can land somewhere in the range of 70 to 200 credits. Imagine if your text messages were priced based on how hard the recipient had to think about them.
Then there’s the second meter. Internal agents used by people who already hold a Microsoft 365 Copilot license run at no extra credit cost. Everyone else draws from packs or pay-as-you-go. That’s the line access fee plus usage charges of the old phone bill, and it means two departments running the exact same agent can cost wildly different amounts depending on who is licensed.
So how do you slice the pie?
If you run IT for an organization of any real size, you already know the departments asking for AI don’t look anything alike. An HR team wants a policy FAQ bot that mostly gives short, scripted answers. A research department wants to summarize literature and analyze data, burning credits by the dozen on every interaction. Finance wants report automation. Clinical or operations teams want something in between. Headcount tells you almost nothing about who will consume what.
And because the pool is shared, the pain isn’t. Whoever adopts fastest drains the bucket, and the slower, more careful teams find it empty at the end of the month. Here are the allocation models I keep coming back to, each with its cell plan equivalent.
- Even split (everyone gets the same minutes). Simple and fair-looking. It’s also wrong almost immediately, because need is nowhere near even.
- Headcount-weighted (per-line pricing). Easy to defend to finance, but disconnected from actual usage. A 10-person research group can outspend a 200-person department.
- Use-case based (picking a plan by how you actually text). Departments submit their intended use cases, you estimate credits per interaction and expected volume, and you allocate from there. More work upfront, far more accurate.
- Showback, then chargeback (reading a few bills before choosing a plan). Meter a pilot period and show each department its burn without billing it. Once real patterns emerge, move to actual chargeback. This is the most honest option, and it builds trust instead of arguments.
- Central pool with caps and a reserve (the family plan with per-line limits). Keep a shared bucket for experimentation, set hard caps per department or environment, and hold back a reserve for spikes so nobody’s agent goes dark mid-month.
In practice, the answer is usually a blend. Start with showback to learn, then layer in caps and a reserve once you know who your heavy texters are.
Don’t hold your breath for unlimited
We all know how the texting story ended. Once a text cost carriers next to nothing to deliver, the meter stopped making sense and unlimited plans became the norm. Nobody counts texts anymore.
AI isn’t there. Every one of those credits represents real compute, and the heavier the task, the more it costs to run. The meter isn’t going away anytime soon, which means AI budgeting is a governance and leadership problem right now, not a licensing footnote to sort out later.
The good news is that we’ve been here before. We learned to read the bill, set limits, and have the awkward conversation with whoever ran up the overage. The more things change, the more they stay the same.
Now if you’ll excuse me, I need to go find out who used up all the family’s credits.