Every decision a company makes is a financial decision. So finance isn't just the function that reports on the business; it's where the business gets decided.

Which makes it worth asking where your finance team's hours actually go. Reconciling last month. Chasing receipts. Policing a $30 subscription. All of it backward-looking.

That's why I use Brex: corporate cards plus AI-native workflows that absorb the record-keeping. Zero-touch receipt matching, real-time policy enforcement, continuous reconciliation. Controls fire before money leaves the account. Close takes days, not weeks.

Defense handled. Go play offense. Thousands of companies — including Anthropic, Coinbase, and DoorDash — already run on Brex. Head to brex.com/metrics to join them.

The Mostly Metrics world tour continues 🌎

RSVP below, rumor is Walter our editor in chief may make an appearance

Join us in London on October 15th and Boston on October 20th

👋 Hi, it's CJ Gustafson and welcome to Mostly Metrics…

Ya ever work on a project / pptx / book report so long that when you finish you just say PLZ take this thing off my hands because it’s eaten my mind? That’s the point I’m at right now. I feel like Leo in the Revenant, emerging from the arctic after sleeping inside of a horse.

Too much?

Any who!

This is the second post in our series on managing LLM token costs. Both parts are recorded on video, accompanied by the source files and final outputs to download.

Last time we looked at the tokens your product burns, aka the ones you sell through to your end customers to make that dinero. Today we're inside your house…

In this tutorial we’re analyzing your team's Claude and OpenAI and Cursor seats, plus all the tokens getting gobbled at the company level. And then we give you frameworks to budget for them (‘I’ve got the frameworks!’ (read in an Arby’s ‘we’ve got the meats voice’)).

To set some context for what we are about to look at, our friends at Brex went through their card and bill pay data, and the median company now pays for two AI tools in a given month, with the mean a shade over three. This is up from a median of two and an average of three two years ago.

Same trend with model vendors. 61% of companies pay two or more in the same month, up from 20% two years ago, and the median company crossed over from one to two back in January (free tiers / anything not run through a Brex card are invisible to them, so read all of this as a floor).

OK, let’s budget!

First, here's the full walkthrough.

And here's everything behind it.

Open Claude, drop them in, and follow along.

Last tutorial I used the Admin Usage and Cost API, which groups spend by workspace, API key, and model. That's great for your product's AI costs. But it's useless today, because it has no idea who is who within your org and what they're consuming internally.

What you want is the Spend report. Go to Settings, then Analytics, scroll down to "How much is Claude costing?" and hit Export spend report.

Then I marry it to my headcount roster and my seat list.

I grab three months, so I can make a trend out of this (in this case I used April, May, and June). You'll see net spend is what we actually paid after our negotiated rate, and gross is list price (ya boy loves a good discount).

The roster gives me department, role, level, and comp. Department is the one that is critical to this exercise. The seat list gets me what each person's seat costs, because Claude bills a lot of these as a hybrid - a subscription per employee + additional usage layered on top.

Then I hand the whole pile to Claude and have it join them using email identifiers.

In this example, every row matches to a person… no orphaned emails… Guess we didn't fire anybody at this fictional company this month.

Now every dollar of LLM spend has a name and a department attached to it. I'd encourage you to look at it per head rather than in total. Totals will def tell you engineering is your biggest department, which you already knew. Per head will tell you that finance was a lot higher than you expected.

Here's what falls out once you can see it by person.

The ceiling is going to be a critical component of our budget. The move here is to budget everyone else up to the frontier. I call it frontier convergence.

What we're doing now is forecasting by role, rather than by a named person. We are looking at what a staff engineer costs when they're using AI well, what an SDR costs fully ramped, etc.

We have to encourage a mindset shift. AI doesn't get grandfathered in. It competes from zero, against every other dollar. If this dollar could go to a head, a tool, or a token budget, and it has to win on return, what do you actually pick?

Here are three ways to do it, and they build on each other.

First method: $ per head. Everyone starts here… Total spend, divided by headcount, times next year's headcount.

The trouble is AI usage is lumpy. A handful of heavy users spend many times what a normal person does, and they drag the average up with them.

It's fine as a top-down sanity check, or a first budget envelope pass. But it's useless for actually allocating, because it hides all the variance and nuance.

Second method: Budget AI the way companies budget healthcare. A set percentage on top of base comp, because it's becoming a cost of employing a person, same as benefits loading.

Its best use is as an output or reporting number to ground the board. We spend 20% of comp per employee on tokens.

It's clean, and for your board's higher level math they have a driver. But you as the operator responsible for the plan still don't know what's inside that percentage. It's a great communication tool, not a real lever for allocation.

Third method, and this one came from Meredith Finn, the CFO and COO of Front: Go team by team and find what your top-decile users spend per month.

The assumption is your heaviest users are what the company looks like when people are good at this. They've worked out how to get real output from the tools, and the job is getting everyone else to where they already are.

Do it at the team, then the role level. An engineering manager isn't spending like an engineering IC, and a support rep is probably spending more than their support manager.

Next, smooth the outliers. Use the top decile rather than the single max, or trim the highest and lowest in each group, so one power user doesn't set an unrealistic ceiling.

Then ramp your laggards up to that frontier over a few quarters, like someone getting their feet under them in a new job, with a margin of error on top.

Don't make it land all at once. To use a sales quota analogy, you wouldn't assume a rep is fully productive on day one.

What we've done now is shift the unit from a person to a seat. What a great engineer spends… What a great analyst spends… That's a number you can roll up and forecast with some intellectual honesty.

Interestingly enough, in the video two of mine landed within 10% of each other off completely different methods, which helps me hone in on a range. In many ways, it reminds me of working in valuations, where you run three approaches and go looking for where they agree.

Net net, the method I'd depend most on for forecasting purposes is the third, frontier convergence.

And we can convert that into a model with inputs and outputs.

In the model we built you can enter next year's headcount by role, it prices each head at the frontier, ramps adoption over time (because nobody hits frontier immediately), and adds a margin of error. I love a good buffer.

Once again, here are the files, including the model.

1. Join the spend file to your org chart.

2. Find the money you're wasting before rolling it forward.

3. Test the average.

4. Find your power users.

5. Check whether the frontier has a ceiling.

6. Build the persona table.

7. Run three budgets three ways.

8. Turn it into a model.

So go forth and budget with confidence, please.

CJ