FiscalDoc: replacing CHF 1,400/year in SaaS with a local AI

Yes, a SaaS subscription can be replaced with an owned tool — I did it, and this note documents the case end to end. FiscalDoc, a local application built in three evenings through dialogue with a code assistant, sorts my tax documents using an open language model that runs on my Mac: zero subscriptions, zero data leaving my machine. Here is the mechanics, the real cost, and the honest limits of the demonstration.

Original note, April 2026: a few weeks earlier I had published SaaS versus custom-built in Switzerland, where I argued a simple thesis — for many Swiss companies, owning is once again smarter than renting, because AI has collapsed the cost of code. Several readers asked me for a concrete example. Here it is. And it's sized to fit me.

The problem, a very ordinary one

Two spheres share my administrative daily life: MCVA Consulting SA, my AI-augmented agency based in Haute-Nendaz, and my personal tax affairs. Two scales, two rhythms, two logics — but one shared reality: a constant stream of documents to sort and retrieve.

Between these two spheres, hundreds of documents arrive every year through every possible channel: supplier invoices in PDF, expense receipts photographed at a restaurant, AHV and pension fund statements, cantonal tax assessments, bank certificates, various receipts.

The pain isn't the accumulation. It's the moment, some Saturday in March, when everything needs to be found again for the tax return or the year-end accounts. Who hasn't spent two hours hunting for a receipt in an overflowing "Downloads" folder?

The SaaS options I looked at

The market's obvious answer is a subscription, and I looked seriously at three categories. Tax document management solutions first, like Dext: between CHF 25 and 60 per month per entity, i.e. CHF 600 to 1,400 per year for my two spheres, with storage most often held by a US-based host. Document modules built into ERPs next, like Bexio or Odoo: more complete, but tied to an accounting tool I have no wish to impose on my accountant, and which lock you in a little more every year. Generic document management tools finally, like Notion or Airtable: flexible, but everything has to be built yourself inside the tool — while paying every month for the right to access it.

None of these options ticked my three baseline requirements: absolute data sovereignty (not a single byte should ever leave my machine), genuinely intelligent classification of documents according to my own Swiss tax categories, and a complete absence of subscriptions and marginal cost.

What I built, and how long it took

I built FiscalDoc. A local web application, running on my Mac, that does exactly what I need:

1. I drop a document — PDF or photo — into a Streamlit interface.
2. The document is analyzed by Gemma, an open-source language model installed directly on my machine. No network calls. No API.
3. The model returns structured JSON: sphere (MCVA or personal), tax category (income, social charges, VAT, deductions, mortgage interest…), date, amount, issuer.
4. The file is automatically renamed according to the convention YYYY-MM-DD_TYPE_ISSUER and filed into a clean folder structure.
5. A local SQLite database indexes everything for instant search in a filterable dashboard.

And because I wanted to go a step further, I added a local conversational assistant, also connected to the model. I can ask it, in plain language, very specific questions: "Find me all MCVA supplier invoices over CHF 2,000 for the fourth quarter of 2025," "Add up my third-pillar deductions over the last three years," "Is there a missing receipt for this tax assessment?". The assistant queries the SQLite database, rereads the documents if needed, and answers. Always locally.

The whole thing fits in a folder of a few hundred lines of Python. No data leaves my machine. Ever. I own the tool. I own the code. I own the data. I even own the model that analyzes it.

The real cost

This is where the example gets interesting. The table compares what the subscription option costs against what the owned tool costs, line by line.

What the table shows: where a subscription bills every year for access to a generic function hosted elsewhere, the owned tool concentrates its entire cost in the initial build — three evenings — and then has none left. The only real "cost" is RAM: an open model needs a machine with adequate memory, the fair trade-off for running it at home rather than on a remote server. It's a one-off hardware investment, already amortized by the rest of my professional use. Once that hurdle is cleared, FiscalDoc's annual operating cost is strictly zero. Not "almost zero." Not "a few dollars of API calls." Zero.

What this example demonstrates

For fifteen years, the killer argument for SaaS was: "You can't build this yourself, it's too long, too expensive, too risky, and you'll never get the power of a real AI model." That was true. It no longer is — on any of the four points.

What has changed in eighteen months comes down to two shifts. On one side, code assistants have collapsed the time it takes to turn an idea into a working implementation. I didn't write FiscalDoc: I dialogued it into existence. I described my needs — two spheres, Swiss tax categories, automatic classification, a lightweight interface, a local database, a conversational assistant — and I iterated evening after evening until I got a tool that fits me. I documented this way of building in my hands-on account of AI-assisted development.

On the other side, open models have reached a quality level that makes cloud calls optional for a wide range of professional use cases. What yesterday required an American vendor's API today runs on a properly sized Mac, never connecting to the internet for the core processing.

The combination of the two changes the trade-off. What used to be an agency project costing tens of thousands of francs plus a subscription to an external API is today, for a narrowly scoped need like mine, a three-evening project, zero francs of operating cost, and zero lines of data leaving the user's own boundary.

The lesson, beyond my desk

FiscalDoc is a tiny case. Two spheres, one user, a few hundred documents a year. But it illustrates exactly what I believe, and what I build at MCVA: with local generative AI, SaaS is no longer the default choice — it has become a choice that needs justifying.

For a Swiss company subject to the revised Federal Act on Data Protection^[1], mindful of the CLOUD Act and tired of its vendors' annual price hikes, this shift isn't theoretical. It's available, now, to anyone willing to look at it honestly.

Before the next subscription renewal, four questions deserve to be asked with a cool head. Does this tool do exactly what the company needs, or has the company adapted to its limitations over the years? Do you want to pay rent to a vendor for the next ten years just to keep accessing your own data? Can this data really be entrusted to a third-party cloud, given the revised data protection act and trade secrecy? And is the custom alternative — driven by a code assistant, run on a local model, owned outright — really out of reach, or do you just still believe it is, out of habit?

Fifteen years in the digital leadership of large groups showed me the other end of the scale: software budgets stacked with dozens of redundant subscriptions, worked around with parallel spreadsheets. What I proved on my own Mac at the scale of a one-person practice, the same logic allows, function by function, for SMEs of five to fifty people — that is precisely the territory of my custom-built solutions. And the overarching question, the one that governs this shift and the next, I laid out in What should a Swiss SME do about AI in 2026?

Key takeaways

— A narrowly scoped need, sensitive data, no network effect: that's the zone where an owned tool beats a subscription, and FiscalDoc is the documented proof. — The operating cost of a local application built on an open model is zero; all the cost sits in the build, now collapsed by AI-assisted development. — The shift doesn't apply to everything: large-scale collaboration and heavy functional complexity remain vendor territory — the trade-off is made function by function.

FAQ

Do you need to be a developer to replicate this case? You need someone who can dialogue with a code assistant and judge the result — a technical profile, not necessarily a professional developer. For an SME without that profile in-house, it's typically a short custom project, whose deliverable belongs to you entirely, code and data included.

Is a local model good enough for accounting documents? For structured classification — document type, date, amount, issuer, category — yes, comfortably, and it's verifiable on your own documents before any commitment. For complex multi-document reasoning, open models are improving but I still validate case by case.

And what about Swiss data protection law in all this? It's one of the strongest arguments in this setup: tax documents contain personal data within the meaning of the law, and a local application removes the question of transfer to a subcontractor, applicable law, and extraterritorial access. What the law requires is still your responsibility — but the perimeter to protect shrinks down to your own machine.

Does this case justify canceling all my SaaS subscriptions? No, and I'd advise against reading it that way. It justifies an inventory: what you're paying for, what you're actually using, where your data sits — then a function-by-function trade-off, where some subscriptions survive because they earn their place.

And in your company? The AI Usage Diagnostic: sixty minutes to lay out your actual workflows, identify what deserves custom software, what stays in SaaS, and what needs no AI at all. Book a diagnostic

Sources

[1] Federal Act on Data Protection (FADP), revision of September 25, 2020, entered into force on September 1, 2023. www.fedlex.admin.ch/eli/cc/2022/491/fr [↩]

Jérôme Deshaie is CEO and founder of MCVA Consulting SA, an AI-augmented agency based in Valais. Fifteen years serving major international brands, now working directly with Swiss SMEs. Background.