The AI adoption gap is a support gap, and bespoke software won't close it

Advazia — Wed, 23 Sep 2026 09:18:53 +0000

In 2025, one EU enterprise in five used artificial intelligence. A year earlier it was 13.5 %. That is Eurostat's figure, published on 11 December 2025, for enterprises with ten or more employees.

The number hides a split. The same survey gives 55 % for large enterprises, 30 % for medium ones, 17 % for small.

In Switzerland, the 2025 AXA / Sotomo study of 300 SMEs says the same thing from a different angle: 34 % have integrated AI into their processes, 37 % are trying it, 29 % do not use it. Among the users, 57 % report saving time. And only one SME in three has clear data-protection rules for it; 23 % among the smallest.

We install open-source AI software in small companies, then we run it. This post is what those numbers look like from the install side, and the checklist we ended up with.

What the install side looks like

The tooling is not the problem. Most of it is open source, documented, and a person who knows what they are doing can bring a stack up in an afternoon. What a small company usually lacks is that person.

The big vendors support two audiences well. Very large enterprises get dedicated teams, framework agreements and named architects. Startups get credits and mentoring: AWS Activate, Google for Startups, Microsoft for Startups, NVIDIA Inception. A company with fifteen or eighty people buys a licence, gets a link to the docs, and is on its own.

That is not a complaint. A global vendor cannot send an engineer to every accounting firm in western Switzerland. But that is often exactly what the accounting firm needs. The gap between 55 % and 17 % is not only a software budget gap. We think it is, first of all, a support gap.

We saw it up close. We stood up four popular open-source AI stacks on our own hardware. Three of the four left account creation open on the public internet by default, and none of the documentation said so. A company installing on its own can fall into that hole without knowing it. The write-up is here: Four open-source AI stacks, open by default.

Why we default to proven open source, and what we still build

Faced with that gap, a frequent answer is bespoke development: a services company builds a system specific to the client. Sometimes that is justified. For a small company, it is not always the best fit, for a simple reason: a bespoke system often has one team that understands it, and that team is needed twice. Once to build it, then for as long as it runs, to maintain it, fix it, upgrade it, and keep it working when the models change. And the models change often.

A proven open-source project is in a different position. A whole community works on it, often daily: vulnerabilities are found and fixed by people who are not your contractor, new models are wired in by someone else, the documentation is written by the people who use it. The company depends less on a single vendor, and less on code nobody else can read.

That does not mean nothing gets built. What is specific to each company is the integration: its tools, the rules of its trade, its data to migrate. That is where bespoke work most often belongs. We develop the layer between that software and your systems: the connection to your tools, the rules of your trade, the migration and digitisation of your data. The rest we do not rewrite.

And "proven" is something you verify, not assume: we examine software as it appears, we put it through its paces on our own machines, and we keep only what proves reliable in real service. We prefer open source software whenever it meets the need. We install no software whose business model is your data.

The checklist we now follow

These are the controls we write into every scope before the work begins. None of them is exotic; all of them are the difference between a demo and a system a small company can live with.

Decide where the data lives, first. You decide where your data is processed, and we install the system there: a server on your premises, our hardware in Lausanne, or a Swiss host. For the SME in three without a data rule, this is the rule.
Put a person in the loop where it matters. We define together the moments where a person validates. The file comes back to them with what the decision requires, and the steps concerned are written into the scope before the work begins.
Make answers checkable. A model queried on its own invents. The system answers from your documents, gives its source, and flags the questions it could not answer.
Route every model call through a gateway, and cap it. Every installation includes a gateway, and no model call bypasses it. You see spending use case by use case, and you set the ceiling.
Grant access one tool at a time. The system's access to your tools is defined one tool at a time.
Leave the company owning it. The software, the developments that integrate it into your systems and the documentation are not proprietary: they can be taken over by your own teams, or by another provider's.
One engineer, end to end. The same person carries diagnosis, going live, training and keeping it running. Each stage is decided separately, at a price settled before it begins. None is billed by time spent.

What to take away

One company in five is through the door. The others are not short of tools or of will: 45 % of Swiss SMEs see AI as an opportunity, 13 % remain sceptical. What they most often lack is someone who installs, trains, and stays. And that someone is usually better off relying on what a community already maintains than on what they alone would know how to repair.

If you run infrastructure for small companies and have seen the same gap, or a different one, I would like to hear how you close it.

Sources. Eurostat, 20% of EU enterprises use AI technologies, 11 December 2025 (enterprises with 10+ employees; 20.0 % in 2025, 13.5 % in 2024; 17 % / 30 % / 55 % by size) — ec.europa.eu/eurostat. AXA / Sotomo, 2025 labour-market study, 300 SMEs in French- and German-speaking Switzerland (34 % / 37 % / 29 %; 57 % save time; one SME in three with data-protection rules, 23 % among the smallest; 45 % see an opportunity, 13 % remain sceptical), reported by ICTjournal, 9 October 2025. The startup programmes named are the vendors' public programmes.

Advazia installs open-source artificial-intelligence software in companies, then runs it. Lausanne, Switzerland — advazia.ch/en.

DEV Community: Advazia

The AI adoption gap is a support gap, and bespoke software won't close it

What the install side looks like

Why we default to proven open source, and what we still build

The checklist we now follow

What to take away