Replacing a SaaS subscription with a local application driven by an open model: an observed case

Note revised on 25 May 2026. Article originally published in April 2026 — full rewrite in the editorial register.

For fifteen years, the software subscription model — Software as a Service — has imposed itself as the default choice for almost every support function in a company. Accounting, document management, electronic signature, customer-relationship management, expense-claim tracking, invoice filing: every function has found its publisher, every subscription has found its annual budget, and the bill has stabilised on substantial amounts for the smallest Swiss SME.

This stability rested on an implicit hypothesis: developing an internal alternative would necessarily be more costly than the subscription, longer to implement, and of lower quality than what specialist publisher teams produce. This hypothesis is giving way. The conjunction of open generative models that can now execute on standard enterprise hardware and of development assistants that significantly accelerate code production redefines the trade-off. This note documents an observed case on that shift.

The case: fiscal document management for a one-person firm

The studied case is deliberately tiny. An independent consulting firm, operated by one person, which produces several hundred administrative documents each year to be classified for accounting and tax purposes: supplier invoices in PDF format, photographed receipts, various certificates, cantonal tax decisions, bank receipts. The operational need is trivial. Classifying these documents as they come, retrieving them quickly at year-end, retaining the necessary traceability in the event of a tax audit.

The classical market response to this need is a subscription to a fiscal document-management solution, whose rates usually range from a few tens to over one hundred francs per month depending on features. For a one-person firm, the cumulative annual cost is not negligible relative to the effective volume. More structurally: these solutions host the documents at the publisher, who is frequently subject to foreign law — typically US law — on data that may contain financial and personal elements covered by confidentiality obligations. These situations can raise questions of applicable law and extraterritorial access to data.

The observed alternative took a different form. A local application, running on the user's personal computer, driven by an open language model also executed locally, which automatically classifies the documents dropped into its interface, renames them according to a coherent convention, files them in a clean tree, and allows natural-language search across all archived documents. The application was developed over a short period, through iterative dialogue with a coding assistant. The model used is open and executed entirely on the user's machine, with no network call for classification operations.

The mechanics of the deployment

The technical detail of the mechanics illuminates the conditions of replicability of the case. The system articulates around four identifiable components.

An open language model executed locally — the family of open models now reaching useful performance on standard enterprise hardware, provided sufficient memory. This component is the analysis engine that extracts from the dropped document the structured elements: type of document, date, amount, issuer, fiscal category. No data leaves the user's hardware perimeter for this analysis.

A lightweight user interface that allows dropping a document — PDF file or photo taken on a phone — and visualising the proposed classification. This interface is developed in a standard Python framework, which does not require advanced web-development competence.

A local database — typically SQLite — that indexes the classified documents and their metadata for rapid retrieval.

A local conversational agent, also driven by the open model, which queries the base and allows natural-language search: "find me the deduction supporting documents for the 3rd-pillar account over the last three years", "is there a missing document for this tax decision".

The whole represents, in code volume, a few hundred lines — an order of magnitude that would have been unreachable two years ago without a dedicated development team. The current coding assistants make it possible to hold this production volume through iterative dialogue with the developer, who formulates intents and corrects the successive iterations.

What this case demonstrates, and what it does not

The observed case is not a general proof that any software function can be replaced by a local application. It demonstrates a precise zone of shift, which deserves to be identified for what it is.

The shift works when the operational need is circumscribed, when the data processed must remain within a controlled perimeter, when the network effect of a centralised solution is not the main value — a personal document-management tool has no need for multi-user collaboration or synchronisation across dozens of employees. The studied case ticks these three conditions.

The shift does not work — or does not yet work — when the operational need involves tens or hundreds of users in synchronous collaboration, when the workflows rely on integration with dozens of external services, when the functional complexity exceeds what a single developer can hold with an assistant. Enterprise software publishers retain value in these zones.

Between the two, there exists a substantial zone of software tools that Swiss SMEs use by subscription, whose operational need is in reality circumscribed, whose processed data would gain by remaining within a controlled perimeter, and whose functional complexity can now be held by a small team with the available tools. This zone is in observable expansion.

Four questions to ask before the next subscription renewal

For a Swiss SME preparing its next cycle of software arbitrations, four questions emerge from the observed case.

Does the subscription tool used do exactly what the company needs, or has the company adapted to the tool's limitations over the years? Does the used functionality represent ten per cent, twenty per cent or fifty per cent of what is billed?

Can the data processed by this tool reasonably be entrusted to a third-party cloud, in light of the Swiss data-protection framework^[1] and the confidentiality obligations of the profession exercised?

Is the custom alternative — local application, open model executed locally, code owned by the company — effectively out of reach at the scale of the need, or does this judgement rest on hypotheses dating from a time when it indeed was?

Is the exit cost from the subscription — data migration, staff requalification, integrations to be redone — anticipated in the current total cost of ownership, or is it an implicit debt the company accumulates at each renewal?

These questions do not systematically lead to a subscription replacement. They lead to an informed arbitration, rather than a renewal by habit.

The scope of the case, beyond the individual perimeter

The observed case concerns a tiny entity — a one-person firm, a few hundred documents per year, a strictly personal use. Its pedagogical scope, however, exceeds its perimeter.

What it demonstrates applies, at scale, to much larger organisations that have piled up over fifteen years dozens of sector subscriptions, each with its per-user rate, its annual price increases, its functional limitations bypassed by parallel files, its data entrusted to hosts subject to foreign law. The logic that makes the local alternative relevant for an independent consulting firm makes the internal alternative relevant for SMEs and larger structures, on certain precise functions of their software arsenal.

This extension is not a universal promise. It is an arbitration zone that deserves to be qualified for each organisation, in its own context, with its real constraints. The note SaaS versus custom development in Switzerland addresses this question on a larger scale.

What the observed case excludes, by contrast, is the argument that the internal alternative would be as a matter of principle impossible for the Swiss SME. This argument, which had its relevance during the decade when generative models were inaccessible and when the cost of code remained prohibitive for an internal function, no longer holds. The terrain has shifted beneath the hypotheses, and the trade-offs would deserve to be posed afresh.

Sources

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

Jérôme Deshaie is CEO of MCVA Consulting SA, a Swiss firm specialising in strategic consulting on artificial intelligence, based in Valais.