I built an accounts payable system in two days. Claude Code, full OCR, approval workflows, 10,000 invoices running through it. The whole thing worked.
And I would never recommend it to a client. The build was fast, but the maintenance would be endless. Edge cases, liability, frontend upkeep. For that company, the right answer was still to buy Ramp or Concur. The existing tools solve this problem better than anything I could maintain on my own.
A week later, a different company needed data visibility. They were considering a full ERP replacement, but what they actually needed was the ability to see what was happening across the business. I stood up a data warehouse using open-source tools like Postgres, dbt, and Superset. AI configured the connections. Within a week, 1,800 employees had access to row-level reporting across 40 million G/L lines with incremental updates. The cost was almost nothing.
Same person, same tools. One project was a waste of time. The other changed how the company operates. The only difference was the use case.
That distinction, knowing what's worth building versus what isn't, is the skill that matters most right now in finance. And almost nobody has developed it yet.
You're not as behind as LinkedIn makes you feel
I keep hearing the same thing from finance leaders right now. There’s this low-burning anxiety that doesn’t go away. You see the horsepower that’s available. You scroll LinkedIn, and it looks like everyone else has it figured out. But the reality is different. Recently, CJ Gustafson from Mostly Metrics surveyed hundreds of CFOs who are actually using AI, and most teams are still just experimenting in pockets. A couple of determined people trying things on their own. Not pervasive. Not a mandate. If you think you’re behind, there’s almost certainly someone further behind.
I find that reassuring. But it doesn’t answer the harder question, which is: where do you actually aim your energy?
This is something I think about constantly. Finance has never had the equivalent of the 10X engineer. Now the power is there. But people are fatigued by the noise, desensitized to the claims, and stuck in a loop of knowing they need to get better without knowing what to do next. I get it. I feel it too, and I’m building software in this space.
From asking questions to running the strategy
I've been watching this play out, and I've started to see a pattern in how FP&A professionals are evolving with AI. It breaks into four phases.
The Question Asker: You're exploring AI. You ask about accounting rules, reporting logic, and tax questions. Maybe you send in a spreadsheet to analyze. You get decent answers back. The problem is that if you can ask these questions, so can anyone else.
Next step: Go from asking to building. Dream up a tool. Keep pushing until you hit the limit of what AI can produce for you.
The Vibe Coder: Now you're building. Python scripts, automated reports, quick models. It feels like a superpower. But most of what you create is one-off. It works for you, nobody else uses it, and it usually gets abandoned. This is the phase where compounding curiosity kicks in. You try one thing, and it plants a seed. That energy is the engine of this phase, but it's also the trap.
Next step: Get someone else to actually run your tool. Pitch it to your manager. If nobody else touches it, it's still a prototype.
The Builder: This is where the mindset shifts. You stop solving individual problems and start designing systems. Instead of building a report, you build a reporting layer. Instead of pulling data, you connect systems. You become what I'd call a decision architect for the rest of the org. Your work becomes scalable, repeatable, and usable by others.
Next step: Work with someone else to build it. Institute this org-wide. Handle the objections.
The Orchestrator: You're not building everything yourself anymore. You're deciding what should be built, what shouldn't, and how it all fits together. You manage small build teams while spending most of your effort on analysis and strategy.
The skill that matters most now is judgment
The key to reaching Phase 4 is developing a sense for what makes a good use case. The capability to build is now essentially infinite. The discipline is knowing when to stop.
Worth building
Deterministic reports that live in your data warehouse and that everyone across the business can access.
Connecting data sources through APIs, where you used to pay implementers $50k per connection.
Automating variance analysis that flags anomalies across hundreds of line items before your team even opens the close package.
Not worth the risk
Rebuilding an entire A/P system. I learned this one personally.
Vibe-coding your planning process. An LLM doesn't carry your model logic, version history, or assumptions. The output sounds convincing but isn't safe to act on.
Custom internal apps with security implications. Networking is networking, and trusting a vibe-coded solution to handle that is a different kind of risk entirely.
Behind mountains, there are more mountains
My advice for where to start: pick one decision. Not the annual plan. Something micro. Forecasting next quarter's AWS bill. Figuring out how long it takes an enterprise rep to ramp up in a specific geography.
The mountain range is overwhelming. The hill in front of you isn't. One small problem solved well will inform the bigger ones, and that's how you start to change how you work.
The gap between the teams pulling ahead and the ones that aren't isn't linear. It's compounding. Every month you wait, the catch-up gets harder. But the starting point is smaller than you think.
