I did a full financial analysis with a coding agent. Cap table, bank transactions, PostHog, Stripe. All in a clean file system.
I told Claude Code what I wanted. It navigated the data like an engineer.
How it worked
The agent used ls, grep, find, cat. It traced relationships between files. It validated assumptions. It generated charts with Python and matplotlib.
This isn't special AI magic. It's a coding agent with access to a file system. But the time-to-value compared to the traditional process was massive.
The communication tax
When you work with a human analyst — or a CFO, or a consultant — you pay a communication tax. Context, intent, rationale, edge cases — all of it needs to be explained before anything happens.
Then you wait. They do the work. They come back with questions. You clarify. They revise. You review. You ask for changes.
Every step involves waiting and explaining. The work itself might take two hours, but the communication around it takes two weeks.
Why a coding agent changes this
With a coding agent, you iterate directly on the source of truth. Change an assumption, rerun the analysis, get an answer in seconds.
Experiments are cheap and reversible. You're not worried about wasting someone's time or asking a dumb question. You just try things.
If you want to see what happens when you exclude a certain category of expenses, you change the filter and run it again. If you want a different visualization, you describe it and the agent generates it.
The feedback loop drops from days to seconds.
What this means for hiring
Tobi Lütke, the CEO of Shopify, said: "Before asking for more headcount, demonstrate why you can't get it done using AI."
This framing is right. Hiring someone is expensive — and not just salary. You're paying for communication overhead, onboarding time, and the ongoing cost of coordination.
If you can get the same result with an AI agent iterating on files you already have, the case for hiring weakens considerably.
When you still need a human
A CFO isn't just someone who runs financial reports. A good one understands the business well enough to push back on your assumptions and spot things you're not looking at.
AI can run the reports. It can't tell you which assumptions are wrong. That requires someone who has seen enough businesses to notice patterns you're too close to see.
I'd hire someone for that kind of input — the "you're thinking about this wrong" conversation. I wouldn't hire someone to build spreadsheets I can generate in five minutes with Claude Code.
That line keeps moving as the tools improve.
