The Meter Was Always Running!

The meter nobody had been reading

The first move cost nothing. Copilot already exposed the cost; it just sat where nobody looked. Together with the project team, we switched on the live credit meter, the exact-credit hover, and the per-session readout that shipped with VS Code 1.126. Now a developer could watch a single chat session tick past fourteen credits while it happened.

We expected mild interest. What we saw was closer to discomfort, the useful kind.

Watching the number move in real time does something no policy memo can. It makes the cost yours, in the moment you spend it, instead of finance's problem three weeks later. Nobody enjoys watching a meter run. That is exactly the point. 

The unpopular part

Then the awkward question: does this task actually need the most expensive model? For months the honest default had been to grab the biggest one and stop thinking, and on a team of dozens that reflex was the single most expensive habit on the project.

The numbers were not subtle. Against the top-tier reasoning model, the everyday model runs about 83 percent cheaper and the balanced one about 60 percent cheaper, and documentation, boilerplate and routine refactoring do not need a frontier model reasoning about architecture. So, we encouraged the team to use automatic model selection and to reach for the top tier deliberately rather than by reflex.  

This is where it got human. Being nudged off the most powerful model feels like a demotion, and some people read it that way. There is a real fear underneath it: cheaper means worse output means my work suffers.  

Telling a skilled engineer that their instinct is expensive is not a popular message. What shifted the conversation was not a policy. It was the meter from the first step, combined with one honest framing:

The frontier model is not the prize; it is the escalation!

Use it when the problem earns it. Most problems do not. 

The fuse in the box

A shared pool invites a tragedy of the commons, so there were two numbers, and they are easy to mix up until you notice they do completely different jobs.

The first is what each user actually gets covered: about 30 dollars a month, paid for by the license (19 dollars from the Business License plus an 11-dollar promotion). Think of it as the data allowance in a phone plan. It is already included.

The second is a hard ceiling of 500 dollars, shown as 50,000 credits. That is not a budget, it is a circuit breaker. It is the point where Copilot stops completely, so that no single user can drain the shared project pool on their own.

The part that matters is what sits between the two. Spend more than your included 30 dollars and Copilot does not cut you off. It keeps working, but those extra credits now come out of the shared pool, which is real money the project pays. You only actually hit the wall at 500. Same phone plan: past your allowance you keep surfing, except now it is billed to the company account, and the 500 is the spending limit the company set so nobody runs up a monster bill alone.

Which is why the small number did the heavy lifting. Almost nobody ever came close to 500. But once people understood that credit number 31 was the team's money and not theirs, they spent differently, with nobody enforcing a thing. The 500 dollar wall was there for the rare disaster. The 30 dollar line changed the daily behaviour.

Did the ceiling ever pinch? Occasionally, yes. Someone deep in a legitimate session would have to stop and ask for a raise. Not comfortable, and we would not pretend otherwise. But a cap that never pinches is just decoration, and the alternative was one runaway session starving everyone else. The friction was the feature.

The guardrail gave people room to work while protecting the shared pool from a single unchecked session.  

A smaller guardrail did also quiet work well. We asked agents to stop once the acceptance criterion was met, because an open-ended agent keeps working, and keeps billing, long past the point it is needed. We also reined in auto-approved agent mode, since an agent running unattended in the wrong direction is the most efficient way to turn budget into nothing.  

An autonomous agent still needs a clear task and a clear point at which to stop. Without both, activity continues while responsibility becomes harder to trace.

What stuck, and what is still open

We cannot publish a clean percentage, and we would distrust it if I could. The change that lasted was quieter. The model dropdown stopped being wallpaper. People started choosing in the moment instead of finding out at month end. And the team question shifted from "are we even allowed to use this" to "how do we get more out of every credit".

“AI Credits highlight a fundamental reality of enterprise AI: technology alone does not create value. Value is created when people understand the trade-offs behind their decisions and consciously choose the right tool for the right task. Building this capability across the workforce will be just as important as selecting the right AI platform.”
Björn Kasten, Partner, MHP

But here is the honest edge, and it is the same one we ended part one on. We became very good at seeing what was being spend. We are still not good at proving what that spending bought. We can say that the credits went further. We cannot yet point to a number that proves the output was worth it. 

Closing the gap between cost transparency and value transparency is the next frontier, and it is where this work goes next.  

The switch to AI Credits felt like a penalty when it landed. It turned out to be the useful kind of pressure, the kind that makes a room full of engineers learn, in a week, what a request costs and when the expensive model is worth it. That instinct is the deliverable. The credits we saved are just the receipt.

As AI becomes part of everyday engineering and industrial work, this skill will matter beyond a single Copilot license. Digital sovereignty starts with knowing how AI is used and keeping the ability to change course.

At MHP, we see platforms as the technological foundation. The teams using them still need to set the rules and remain responsible for the decisions they make. Our role is to help organizations establish that practice while ensuring that responsibility stays where it belongs.

Start by looking at where AI costs become visible today and where model choices still happen by default. That alone will show where the first guardrails are needed.

For the broader AI FinOps argument behind this case, read part one of the series.