AI Agency Switzerland.

The range breaks into three cleanly separable blocks: advice before the decision, build up to a working system, and run afterwards. Once you know which block you are buying, you can compare quotes properly — and you notice quickly when a supplier prices one block and silently assumes the second.

• Advice. Process mapping, collecting and scoring use cases, checking the data, defining a pilot with success and stop criteria, the business case, a read on the legal risk. What comes out is documents and a decision, not a tool.
• Build. Data preparation, integration with existing systems, building the application, selecting and connecting the models, a permissions concept, testing against real cases, rollout inside the team. The largest block of effort almost never sits with the model. It sits with everything around it.
• Run. Monitoring quality and cost, keeping the source data current, reacting when a vendor changes its models, support, further development. This is the block buyers forget most often, and the one that decides how long a solution survives.

Around those three blocks sit the familiar shapes: assistants for text work, searchable knowledge bases built from your own documents, pre-sorting of incoming enquiries, analysis of structured data, and agents that carry out several steps by themselves. What that last term means is explained in the glossary entry on AI agents; the practical side is covered in our article on AI agents and business automation.

Advice is billed by time spent, build as a fixed price against a defined scope, run as a monthly line plus usage-based model costs. Those three logics cannot be squeezed into one monthly figure without somebody swallowing a risk — either you or the supplier.

For advice, time-based billing is the honest model, because before the process mapping nobody can seriously say how many use cases will need reviewing. What matters is not the hourly rate but the ceiling agreed in writing up front. For build, a fixed price becomes possible once the scope is described precisely — which is exactly why the advice comes first. At DLM Digital a validation prototype starts at CHF 10,000; websites and web applications sit between CHF 3,000 and CHF 25,000. Comparable orders of magnitude for websites are set out under web design in Switzerland.

The third line is regularly underestimated. Model and infrastructure costs are charged per use and grow with the success of the application. A quote that does not show this line separately is incomplete. Ask for an estimate of the expected monthly running cost, and a second one for the case where usage triples. The gap between the two tells you whether the supplier has understood the operating side of the job.

The most reliable selection process is one identical written brief sent to three suppliers, seven questions in the meeting, and a first assignment with a limited scope instead of a twelve-month contract. Quotes only become comparable when everyone received the same brief. Otherwise you are comparing scopes and mistaking the difference for a price.

The seven questions that separate suppliers fastest:

• Which of our use cases would you advise against, and why? A supplier who names none either was not listening or sells everything.
• Where is our data processed, and what does the contract say about reuse? The answer has to contain a place and a contract type, not the word "secure".
• Which work do you do in house and which do you buy in? Both are legitimate. It only gets hidden when it applies to nearly everything.
• What is the stop criterion for the pilot? Without one there is no failure, only extension.
• What does running it cost in year two? This asks for the line missing from most quotes.
• Who operates the system when you are no longer involved? This tests whether handover and documentation are planned in.
• Can we speak to someone you built something comparable for? A no is not disqualifying. Evasion is.

An eighth point applies specifically to Switzerland: ask whether the supplier raises the revised Federal Act on Data Protection and the question of where data lives without being prompted. Anyone who talks about features first and legal basis only when asked has the order wrong. The same selection logic for digital projects in general is worked through under Choosing a Digital Agency in Switzerland.

The choice hangs on where the value is created: a digital agency for visibility and communication, a software house for a large core system, an AI agency for a function whose worth comes precisely from the AI. The boundaries blur, and most suppliers sit somewhere in between — which is no flaw, as long as they say so.

A digital agency thinks from the market inwards: website, findability, campaigns, conversion. AI is a background tool there, one that speeds production up. A software house thinks from the system outwards: architecture, data model, long-lived operations, integration into an existing landscape. It is the right choice when a core system is being created or replaced, and typically the more expensive one. An AI agency sits between the two: small enough for an undertaking that is not a platform, technical enough for data integration and operations.

For Swiss SMEs there is one pragmatic extra point. Most AI applications end up in a browser — as a form, as a search interface, as an assistant alongside the existing working environment. A supplier who builds web interfaces anyway saves you an integration and a boundary of responsibility. That is precisely where we work: AI Development and web development come from the same team here.

AI projects can largely be run remotely, because the work happens on data, code and documents — with two exceptions where presence pays: the process interviews at the start and the rollout inside the team at the end. We plan both of those on site deliberately and invoice the travel time as a separate line rather than hiding it inside a lump sum.

Our studio is at Gustav-Maurer-Strasse 23 in 8702 Zollikon. In Zollikon, Küsnacht, Zurich and the neighbouring communes an on-site meeting can be arranged at short notice. Central and eastern Switzerland, Bern, Basel, Valais and Ticino we serve remotely on a fixed rhythm: a weekly video session during the analysis phase, a fortnightly one during delivery, shared documents instead of status emails. That structure is not a substitute for proximity. It is the reason distance works, because it forces things to be written down — and in projects with unclear requirements, writing things down is the better mode anyway.

Where an engagement needs work to be observed in a physical place — a workshop, a warehouse, a reception desk, a production line — a travel day is part of the analysis. If you want a large, nationally present organisation looked after, we are the wrong supplier: we are a small team and we turn down engagements we cannot deliver at the quality we promised.

You need no AI agency when your use cases are covered by off-the-shelf subscriptions, nobody has to integrate anything, and nobody has to answer for a running result. That is more often the case than the market suggests — and we would rather say it in the first conversation than in the third month of a project.

• Pure text work in the team. Drafting, summarising meetings, speeding up research: standard subscriptions, an internal usage policy and two half-days of training will do. How to introduce that cleanly is covered in our article on AI content strategy for Swiss SMEs.
• Case volumes that are too small. A handful of cases a month cannot carry build and maintenance. A good template beats any model here.
• Unsettled processes. If there is internal disagreement about how a task should be done, the next investment is a decision, not software.
• No budget for running it. If you can fund the build but not year two, you are building a monument. Start smaller instead.
• The real problem is visibility. If too few enquiries come in, automating how they are handled changes nothing. Then ongoing SEO or visibility inside AI answers is the stronger lever.

We are a small digital agency with our own development team in Zollikon near Zurich, and we cover advice, build and run from the same team — with no published AI reference we could hold up here. Our projects come from hospitality, fashion, delivery services, stationery, kitchen fitting, house clearance and pipe renovation; the technically most demanding one is a web application for tax returns. We put that in writing because this market circulates a great many figures nobody can verify.

What that means for an undertaking in practice: we start with a free sixty-minute conversation in which we give you a read, not a quote. If an engagement makes sense, the analysis phase follows with the effort range fixed beforehand. Only after that does a fixed-price delivery offer appear, because before that nobody can describe a scope honestly. After the analysis you can carry on with us, with another supplier or in house — the documents are written so that this happens without friction.

What we do not do: promise effects, present efficiency percentages from other people's studies as your numbers, or recommend a model before the process is clear. For where AI actually takes hold in a small business, see AI for SMEs. If the route runs through a validating prototype, our guide to MVP development describes the approach.