🚀 How to Build Your Own AI Spend Dashboard | The AI CFO

Yesterday was a big day for airCFO - we officially launched a custom internal AI platform aimed at helping our team operate more effectively. I wrote more about this platform here; I'll be documenting the rollout over the next few months as we get initial feedback & iterate with our team.

Our goal is to create a 'self-driving' infrastructure that gives our team superpowers & facilitates faster, deeper collaboration amongst team members & with clients. We're still just getting started on our AI enablement journey, but this is a big milestone.

Getting a grip on your Claude usage (without an Enterprise plan)

As a builder, I love Claude¹. I use it every day, and many of airCFO's core workflows rely on it.

But as a CFO, it drives me a little crazy. We know our total bill is skyrocketing, but Claude doesn't give me a straight answer on where that spend is going.

The root cause of the issue is that Claude spend happens across three separate systems with their own billing surfaces. Your committed seat cost lives in one export; variable overage charges are buried in another. Then there's an entirely separate API bill, in a completely different console and accessed via API key.

Anthropic has the ability to pull these different spend buckets into one view, but they've decided to gate comprehensive spend analytics as an Enterprise feature. Thousands of companies on Team plans (including airCFO) are seeing their spend skyrocket, but it's difficult to understand the full picture.

In my last article, I described the 'AI chimera' of managing AI spend. The takeaway from that piece in one sentence:

Step one in taming said chimera is seeing it - hopefully, if you're reading this article you're smart enough not to approach a chimera blindfolded. We need to measure AI spend across products, models, and people.

I built a dashboard that does just that, and I'm going to walk you through exactly how I did it.

Dashboard step-by-step

Creating this dashboard is actually not that hard once you know where to pull the data from. Here's what you'll need to create yours:

1. Pull the Claude workspace data. Head to your admin console and export the spend report. It's one CSV that notes down every user, split by product (Chat, Claude Code, Cowork, the Office agents) and by model. This will be the backbone of your entire dashboard.

2. Pull the API platform data. Your API platform data lives in an entirely different console with its own login, and the numbers are attributed by workspace and API key, not by person. So the first time you do this, make sure you map each key to whoever owns it. After that, it's a five-minute lookup every month instead of a mystery.

3. Build a roster file to compile monthly seat costs. Create a simple spreadsheet with each team member's name, email, department, and seat price. For some reason, Claude's 'Spend Report' export only includes overage costs, so you need to input the seat costs by hand. This is also the join key that helps you slice & dice the combined dataset by department.

4. Assemble it using Claude. Fully loaded cost per person is their subscription fee + whatever they spent on overage + API keys attributable to their work. Cut total spend numbers by product, department, or person.

A few things to watch for:

• These reports show raw token usage (not tokens billed on overages).
• Exports only show totals for the entire time period, so if you want to analyze spend month-over-month, you need to pull each month's data in separate exports.
• Claude's export window is only 90 days, so set a recurring reminder or you'll lose the history for good.

That's the manual version. I've built a Skill² at the end of this article that, after a one-time configuration, runs this entire assembly & analysis process (including data exports) end-to-end. But honestly, it's very easy to drop these files into Claude and get the exact report that you need. Plus, it forces you to actually understand what you're looking at.

The Numbers

When you put the three buckets together, you get the full picture.

Now you can see your actual trend line, no guessing. And once you have that, you start noticing the areas that actually need attention. Is a model burning cash on the wrong tasks? Is a department costing more per person than you'd expect? None of it shows up until it's all in one place.

Here's what that looked like for airCFO:

• Total AI cost went from $3,094 in May to $3,562 in June. July is already on pace to land around $4,241 once the month closes out. My initial reaction upon seeing this: looks like a bargain to me!
• However, there are some opportunities to save. Cost per request varies by roughly 30x depending on which model handles the task. Shifting half our routing to the cheaper option would save about $475 a month.
• There were several individuals who are spending well over $100 a month in overages, so I should have a conversation with them to understand what they're doing and potentially upgrade them to a Premium seat (or make their AI usage more efficient).
• The dashboard also flagged seats still active for three people who'd left the company months ago - turning off those seats was a quick win!

Why This Matters

Every dollar in that dashboard was already spent. But now that I can finally see it, I can do something about it going forward.

I already made the fairness case up top. None of this should require an Enterprise contract. But there's a business case for Anthropic here too. When we can't see our spend clearly, CFOs get nervous and cut back indiscriminately. Give users more visibility, and we'll actually spend confidently instead of cutting out of caution.

¹ Also, if I haven't told you yet, I was their Fractional CFO (I like to remind myself of this a few times a week when I'm feeling down).

² 🛠️ The Skill: If you've made it this far, you have the manual steps to build your dashboard down. I've since automated most of this into a Skill that pulls the exports, runs the analysis, and hands me the same breakdown. You can download the skill and read the user guide here.

• How a 20-Person Startup Won Math Olympiad Gold: Watching Harmonic (an airCFO client) tie the frontier labs at the Math Olympiad reframes the whole "can we trust AI with the numbers" debate. The difference between Math Olympiad Gold & hallucinated results is trust, and trust is something you can engineer.
• The Self-Driving Company: An early report-out from a company on the frontlines of rebuilding its operating model to put AI at the center. I like the 'self-driving' analogy: humans have their ability to take the wheel at any point, but AI can handle the freeways on its own.
• 50% of the S&P is AI. So much for market broadening. I'm as bullish on AI as anyone, but this statistic still makes me nervous. Any speedbump could cause a pretty massive ripple effect (and there will almost certainly be speedbumps) 🫣

That's all for this edition of The AI CFO - we'll be back in your inbox soon with more musings on the future of AI-powered finance & operations. Please reply to this email directly with any feedback/suggestions/just to say hi!

Cheers,

👋 Alex Wittenberg, CEO @ airCFO