AI implementation for Swiss SMEs
Not a tenth pilot, but one use case running in production within weeks: identify suitable tasks, select tools, settle data protection under revFADP, and bring your team to the point where they use the solution themselves.
What is AI implementation?
AI implementation is the introduction of artificial-intelligence tools into existing business processes — with the goal of permanently taking over or accelerating one concrete work step. That is very different from experimenting: what stands at the end is not a prototype but a workflow that continues without a consultant.
In Swiss SMEs the viable use cases are rarely spectacular but reliably useful: extracting data from documents into line-of-business software, drafting quotes and reports, answering recurring enquiries, making internal knowledge searchable, summarising tenders and contracts.
Whether a public model, a European alternative or a locally hosted model is used is decided by the sensitivity of your data and your budget — not by whatever is currently fashionable.
- Use case before tool. Process first, software second — never the other way round.
- Measurable. Before starting, we define what success looks like: time, error rate, throughput.
- Data protection upfront. Data categories, provider location and contracts are settled before rollout.
- One case first. One production use case beats five half-finished ones.
- Anchored in-house. Your team can operate, adjust and explain the solution.
Which companies genuinely benefit from AI today
Not every SME is ready — and that has nothing to do with headcount, everything with the starting point.
Lots of paper and PDFs
Invoices, delivery notes, tenders, reports: wherever people retype information from documents every day, the benefit is immediate and easy to measure.
Recurring communication
When a large share of emails is similar in substance — appointment requests, status enquiries, standard offers — a model reliably produces the draft.
Knowledge in many heads
When experience sits with individuals and looking things up takes long, a searchable knowledge base makes the company less dependent on single people.
From first question to running operation
Four phases, each producing a result that stands on its own.
Potential analysis
Together with your employees we walk through the workflows that cost the most time today. Every candidate is rated by data availability, repetition frequency, cost of errors and implementation effort. The outcome is an honest list — including the cases where AI is the wrong answer.
Selection & data protection
For the prioritised use case, two or three options are compared: capabilities, running costs, processing location, data processing agreement, data portability. In parallel we settle which data categories the model may see and what belongs in the internal policy.
Pilot with real data
The solution is set up and tested with real, deliberately awkward cases — not with the showcase example. Hit rate and time saved are measured against the baseline. Only when the numbers hold does the work continue.
Rollout & enablement
Introduction across the team, short instructions in your own language, clear rules on what must be reviewed before anything leaves the company. After four to eight weeks we adjust whatever behaves differently in practice.
What is different afterwards
No magic — relief in places nobody volunteered for, plus a foundation that makes further use cases considerably faster.
- A work step that used to be manual now runs largely automatically.
- Traceable metrics showing what the deployment actually delivers.
- A data protection concept that withstands scrutiny instead of a grey area.
- An internal policy that gives employees certainty when using AI.
- People who operate the tool — not a consultant who has to run it.
- A prioritised, rated list of further use cases.
What we work with
The choice follows data sensitivity, budget and your existing IT landscape — not partner programmes.
- Cloud models with EU/CH processing
- Locally hosted models (Ollama, vLLM)
- Microsoft 365 Copilot
- Domain-specific assistants
- Retrieval over your own documents (RAG)
- Document recognition & OCR
- Connection to DMS and file shares
- Automatic summaries
- Connection to ERP and CRM
- Interfaces via API or webhook
- Automation platforms
- Roles and permissions concept
What does AI implementation cost?
Fixed prices would be misleading here: automating supplier invoice capture in a ten-person business has little in common with a company-wide knowledge base. What you do get is a written quote with a fixed ceiling before any work starts — and a transparent breakdown of the running costs that remain after go-live.
Worth knowing: licence or usage costs for AI tools are usually the smaller item in an SME. The larger share goes to analysis, integration and enablement — precisely the part that determines whether the solution is still in use six months later.
- Number and complexity of use cases
- Condition and accessibility of your data and documents
- Required interfaces to ERP, CRM or industry software
- Public cloud models versus locally hosted models
- Scope of training and internal policy work
- Whether you take on part of the implementation yourself
Frequently asked questions
What does AI implementation mean for an SME in practice?
AI implementation means introducing artificial-intelligence tools into existing business processes — not trying out a chatbot, but deciding which concrete work step will in future be handled fully or partly by a model. In an SME those are typically tasks such as extracting data from supplier invoices, drafting quotes and reports, answering recurring customer enquiries, or making internal documentation searchable.
Do we need large amounts of data?
No. The widespread claim that "AI requires big data first" dates from the era when companies had to train their own models. Today pre-trained language and vision models are used and simply supplied with your context — price lists, templates, contracts, manuals. For most SME use cases, what already sits on your server is enough. Order matters more than volume: documents that are findable and current are worth more than a large, unmaintained archive.
Is using AI compatible with the revised Swiss data protection act?
Yes, provided it is set up properly. The revised Federal Act on Data Protection (revFADP) requires transparency about which personal data is processed, a lawful basis, appropriate security and clarity about disclosures abroad. In practice: define which data categories a model may see, choose a provider processing in Switzerland or the EU, sign a data processing agreement, extend your record of processing activities, and set out usage in a short internal policy. For particularly sensitive data, locally hosted models are an option — nothing leaves the building.
What does introducing AI cost in an SME?
There is no standard price, but there are clear cost drivers: the number of use cases, the quality and accessibility of your data, whether interfaces to existing systems are required, whether public cloud models or locally hosted models are used, and how much training your team needs. On top come running costs for licences or token consumption, which depend heavily on volume. A single use case with clear benefit costs far less than trying to do everything at once — which is exactly why the work deliberately starts with one. You receive a written quote with a fixed ceiling beforehand.
How long until the first production use?
The potential analysis usually takes two to three weeks. A first production use case is often live after a further four to six weeks — provided the required data is accessible and someone inside the company carries the topic. Delays rarely come from technology; they come from unresolved ownership and unsettled data protection questions. Both are therefore addressed deliberately early.
What if AI is not the right answer for our case?
Then that is what the report says. A considerable share of what SMEs describe as an AI need is in fact a process or data problem: a clean automation or a properly configured line-of-business system solves the task more reliably, more cheaply and with less maintenance than a language model. Since no vendor commissions are involved, that answer costs nothing beyond the analysis.
How do we stop the model producing nonsense?
Through three measures. First, the model is restricted to your own sources rather than left to invent — answers then point to the document they came from. Second, use is limited to tasks where a human reviews the result anyway before it leaves the company: quotes, reports, draft replies. Third, testing happens with real and deliberately awkward cases before rollout, and the hit rate is measured. Only when that figure holds does the solution go live.
Find out whether AI will help your business — before you invest.
In a free 30-minute intro call we look at your workflows. You get an honest assessment of where AI works and where a simpler solution is the better answer.