I Spent €2,000 and 500 Hours on AI Tools in 2025. 14 Lessons for Finance.

Over 2025 I spent more than €2,000 and over 500 hours on AI tools. I'm a fractional CFO, not an engineer, so this is a practitioner's roundup: what actually earned its place in my finance workflow, what I wasted money on, and what I'd tell another finance leader to try first. Fourteen lessons.

Why a CFO would spend €2,000 and 500 hours on AI tools

Three reasons: I like understanding how things work, I want things done fast and well, and I have very little patience for repetitive tasks. So I went deep on AI tooling, and my credit card noticed.

NotebookLM reviewed a 250-page M&A contract in 2 hours

Reviewing a 250-page M&A purchase agreement used to take me about three days: dense legal text, lawyers at €500+/hour, and plans and timelines drafted by hand. With NotebookLM (free with Google Workspace) it took about two hours. I uploaded the SPA, voice-chatted through the complex questions, and let it draft the project plans and emails, with the M&A lawyer and corporate finance manager on standby rather than in the lead.

Voice replaced most of my typing

About 70 to 80% of how I talk to my computer and phone is now voice, through superwhisper (~€10/month). That is 50+ prompts a day, with roughly 10x better prompt quality and 5x faster getting thoughts out. Emails, Slack, brainstorming: wherever I used to type, I now speak. Typing feels like going back to a flip phone.

n8n automates finance operations without developers

At around €20/month, n8n is the backbone for automating finance operations and running AI agents. It connects the entire finance tech stack, handles custom integrations, and kills the tool-switching friction of the 5-minute tasks I would otherwise repeat 20 times a day. No developers, and no begging IT.

You don't need to become an engineer, but technical literacy pays

As a CFO I don't need to become an engineer, and I shouldn't spend my day building n8n workflows. But understanding a few key concepts got me far more out of every tool: RAG, MCP and APIs, what an LLM actually is, and the difference between deterministic and non-deterministic outcomes. A couple of afternoons with ChatGPT and YouTube got me about 80% of the way there.

The laziest productivity hack: AI meeting notes

AI note-takers (Fireflies.ai at ~€20/month, Gemini free) save about 10 minutes of post-meeting work each. At three meetings a day across 200 working days, that is 600 meetings a year and roughly 100 hours, or 12 days, saved. I stay focused in the meeting instead of scribbling, tasks get extracted automatically and pushed via n8n into the task manager (Asana, Jira, Notion) so nothing slips, and anyone who missed the call catches up fast.

There is no single best AI assistant

The LLMs (GPT-5.2, Gemini 3) are the brains; the assistants (ChatGPT, Claude, Gemini) are the apps you actually use. 2025 brought major app upgrades almost weekly, and I switched apps and engines constantly. The honest conclusion: no one app wins every task.

Claude Desktop became my AI operating system

It connects to Gmail, Drive, Calendar, n8n and Canva, runs strong models (Sonnet 4.5, Opus 4.5), handles financial data, CSV and Excel well, and writes high-quality JavaScript and Python snippets for n8n workflows. The one gap at the time of writing: no document editing yet.

Vibe coding with Lovable: great for simple tools, and a €700 lesson

Vibe coding means describing software in plain language and letting the AI build it (I used Lovable, usage-based, and Google AI Studio, free). I needed a tool to translate 3,000 account descriptions between English and German; 60 minutes in Lovable and it was done. It works great for simple tools and breaks on complex logic, which I learned the expensive way after about €700 in credits.

n8n vs. Lovable: when to use which

I built a Lovable app to ingest DATEV BWA numbers into a SQL database for Power BI analysis. It worked well until the logic got complex: FX conversion, multiple GAAPs, periodicity. So I switched to n8n, where the visual workflow shows every node and a failure tells me exactly where it broke. My recommendation for CFOs: start with n8n for automation, and add Lovable only when you genuinely need a front-end UI.

The real bottleneck: Finance Engineers

Finance automation has one bottleneck: the gap between knowing what to build and actually building it. Finance experts lack the technical skills, and engineers lack the finance knowledge. The people who hold both (call them Finance Engineers) are rare and valuable. The provocation I keep coming back to: why hire the fifth accounts-payable specialist instead of one finance engineer?

Are AI-native ERPs the Swiss Army Knife? Not yet

AI-native ERPs like Rillet and Light are not ready for European scale-ups yet. They will be central to the finance stack eventually, but they will not replace the treasury tool, email, BI, task managers, or the wiki. So I skip the Swiss-Army-knife fairytale and look for WD-40 instead: AI tools that make the existing stack run better. n8n is exactly that. One small workflow, then another.

My daily AI tool essentials

What I actually reach for every day: Claude Desktop as the all-purpose hub, n8n for workflow automation and AI agents, Fireflies.ai for meeting notes, NotebookLM for analyzing documents, and superwhisper for voice across every app.

Was it worth it?

€2,000+ and 500+ hours later, I work faster and better, and I cut the tasks that used to annoy me. So yes, clearly. You don't have 500 hours, and you don't need them. Encouraging your team costs far less: have them pick one small, annoying task and automate just that. Start there.

If you want help picking that first task, an AI Power Hour is built for exactly this: we take one real finance process and automate it together in 60 minutes.