I Pay CHF 2,700 a Year to My Accountant. So I Built an AI to Replace Him.

I'm a self-employed developer in Switzerland. I run my own consulting business, which means I'm responsible for everything — the work, the clients, the invoicing, and the part I hate most: the finances.

Every year, I pay my accountant CHF 2,700 to file my taxes, handle AHV contributions, and submit my VAT return. And you'd think that for that money, you'd feel taken care of. Like someone's got your back.

You don't.

The emails my CPA sends me are dense. Full of references to paragraphs I've never read, deadlines I didn't know existed, and questions I need to go research myself before I can even answer them. Half the time I'm on Google trying to understand what he's asking me. The other half I'm digging through a drawer looking for a document I'm pretty sure I scanned at some point.

And the worst part — the actual bookkeeping, the receipt sorting, the expense categorization — that's all still on me. I do the work. He files the paperwork. For CHF 2,700.

I don't blame him. That's how it works. But it means I spend my evenings doing admin instead of the work I actually get paid for. And there's always this low-grade anxiety that I'm behind on something, that I missed a letter, that a deadline slipped past me.

Every self-employed person I talk to has their version of this.

What I Built

I'm a developer, so when something is broken I build. The question was: what if an AI could handle the bookkeeping — the part I'm actually doing myself — so I only need the accountant for the final submission?

I found OpenClaw, an open-source AI agent framework. It handles messaging, LLM routing, memory management, conversation threading — all the infrastructure I didn't want to build myself. I just needed to build the intelligence on top: the bookkeeping logic, the document processing, the contact tracking, the calendar.

So that's what I did. I built a personal tax assistant that runs on my own server and talks to me over chat. I send it documents as they come in and it handles everything.

Phone bill arrives? Send it.

📎 swisscom_invoice_march_2026.pdf

Booked CHF 65.00 — Swisscom
Telecom, 50% business use → CHF 32.50 deductible
Filed to 2026/invoices/
Contact updated: Swisscom (last seen: 2026-03-10)

Electricity bill, health insurance statement, a letter from the cantonal tax office — whatever it is, I just forward it. The bot reads the document, figures out what it is, books it, files it to cloud storage, updates the contact history, and checks if there's a deadline attached.

But what really changed things for me: I'm in the middle of filing my 2025 taxes right now, and it's looking like I might have to fire my CPA. The bot gave me a clean ledger, organized every document by category and year, and the filing was basically done before I even sat down to do it.

Here's what my S3 bucket looks like — this is all the bot's work:

Five years of financial documents, organized by year

Each year broken down by document type

Five years of financial documents, categorized and filed. I didn't organize any of this manually.

I can't show you the ledger itself — it's my actual finances — but the bot generates a clean, formatted PDF that I cross-checked against my own records. It's verifiably correct. That's the moment I realized this thing actually works.

The Hard Part

Building the skills — the bookkeeping, the document intake, the calendar — that was the straightforward part. The hard part was getting the AI to actually be reliable and predictable.

The thing nobody warns you about with AI agents: they don't remember. Every conversation starts fresh. Your bot processed a stack of documents, built up a picture of your contacts, your tax situation, your filing history — and then the context window fills up and it forgets all of it.

I needed my bot to know that this letter from the SVA is the third one this year, that the AHV reference number is the same one from January, that the Kantonales Steueramt already sent a reminder about this exact deadline two weeks ago. Without that continuity, the bot is useless. It processes each document in isolation like it's never seen your life before.

I built a full framework around it — a contact registry that tracks every institution and company I deal with, correspondence histories per entity so the bot knows this is the third SVA letter this year, a calendar system for deadlines and obligations, and a knowledge base the bot builds during onboarding by researching your jurisdiction's tax rules.

The workspace files are where it all comes together. I think of them as the agent's operating system. There's a file that defines who the bot is and how it should behave. There's one that enforces a mandatory process for every document — because without it, the AI will skip steps to respond faster. There's a heartbeat system that runs periodic checks: upcoming deadlines, overdue items, data integrity. And there's a memory system — daily logs that get distilled into long-term knowledge over time.

None of this was a quick fix. It took a lot of iteration — running the bot on my own finances, catching the gaps, rebuilding the scaffolding until it was consistent.

Privacy

This mattered to me more than most features. My financial documents — tax returns, bank statements, salary slips — I don't want those sitting on some company's cloud or passing through a third-party API without thought.

The bot runs on a Mac Mini at home. Documents get stored in my own S3 bucket, hosted in Germany — not the US. After the bot processes a file and uploads it to storage, it deletes the local copy immediately. No financial documents linger on the filesystem. The bot actively sanitizes its own environment.

Next step: I'm planning to replace the S3 bucket with a NAS, so everything — the server, the storage — lives physically in my apartment. Full end-to-end encryption for the stored documents is also on my list. But even now, it's already more private than any cloud accounting app.

What It Costs to Run

This surprised me. The whole thing runs on almost nothing.

The storage is free — Cloudflare R2 gives you 10GB with zero egress fees. The onboarding research uses Gemini's free web search tier. A VPS is $5/month. The LLM — I use Claude Sonnet for document accuracy — runs maybe $5-15/month depending on volume.

So maybe $15/month total. For something that runs around the clock, never forgets a deadline, and doesn't charge me per hour.

Felix

If you follow OpenClaw, you've seen what FelixCraftAI and others have done — an AI CEO running a real business, $120k+ in revenue. That's what convinced me these agents aren't toys. They do actual work when you build them right.

Mine isn't running a business. It's running my back office. Same principle: give an agent the right skills, connect it to your life, let it run.

I Packaged It

I figured other self-employed people have the same problem. The dread, the receipts, the anxiety about deadlines. So I packaged everything — all the skills, tools, templates, the workspace framework — as a blueprint you can deploy on your own server.

It works in any jurisdiction. Germany, Switzerland, US, UK — the bot researches your local tax rules during onboarding. Works with any LLM. There's a one-command installer, and a prompt file you can paste into Claude Code or Cursor that handles the deployment for you.

Your data stays on your server. You own every line of code.

$79, once.

Built on OpenClaw. If you're self-employed and drowning in receipts, this is the thing I wish existed two years ago.