The experimentation phase is over in many companies. Staff use AI tools, individual departments have started pilots, and management asks what will remain. Our conversations with companies in Switzerland and Germany show a clear pattern: successful projects start with a problem, not with a tool.
Use cases from your own workflows
The best candidates for AI are workflows in which a lot of text is read, written, sorted or searched: answering enquiries, checking documents, finding knowledge, drafting reports, transferring data from forms. Each case should be rated on four criteria: benefit, feasibility, data readiness and risk. Regularly, part of the ideas turns out to be classic automation without AI, which works faster and more robustly.
Making success measurable
An AI project needs a success measure with a baseline: handling time, accuracy, share of enquiries sorted correctly automatically, or satisfaction. Without a baseline, the benefit cannot be demonstrated later, and the project ends as a matter of opinion. Equally important is a test set of real cases with verified results, against which every new version is measured.
Data and permissions first
The most common mistake with internal assistants is a missing permission concept. An assistant that answers from all documents will sooner or later reveal something not everyone may see. Permissions must therefore be inherited from the source systems. At the same time, it must be clarified which personal data is processed, where models run and whether the provider uses inputs for training. For data that should stay in Switzerland or the EU, suitable offerings exist today.
Classifying under the AI Act and KI-MIG
The EU AI Act differentiates by risk. For most business applications, such as internal assistants or chatbots, transparency duties are what mainly applies: users must be able to recognise that they are interacting with AI. High-risk cases are listed exhaustively in Annex III, including recruitment, creditworthiness assessment and certain applications in education and critical infrastructure; their duties apply from 2 December 2027. In Germany, since 29 July 2026, the Federal Network Agency supervises in principle under the KI-MIG, BaFin in the financial sector. In Switzerland, FINMA expects supervised institutions to have AI governance in line with its Guidance 08/2024.
Human in the loop
For every step with external effect, it should be defined whether the AI may act on its own or a human approves. Starting with suggestions that a human reviews, and automating selected standard cases only after measurably good quality, has proven effective. It lowers risk and builds trust among the staff working with the solution.
Operation is the larger part
A prototype is quickly built. Reliable operation demands more: monitoring quality and cost, keeping content current, handling providers’ model changes and a person who takes responsibility. These tasks should be assigned before launch, not after the first problem.
How to start
Our checklist for starting an AI project summarises the points. If you want to know where AI brings the most in your company, an AI opportunity assessment is the structured entry; for assistants on your own documents, our AI chatbot on your own data offer describes the approach. The overall framework is described by our AI consulting & integration service.