# effektiv GmbH > We deploy artificial intelligence exactly where it creates the greatest impact: in your company's value chain. Our web and mobile applications are 100% developed in Switzerland and meet the highest data protection standards. **Your Swiss partner for the AI revolution** # Artificial Intelligence for your SME. We develop AI tools as well as web and mobile applications. Tailored to your company's challenges, and measurable in results. Let's get to work. ## Our Services We don't just make your company look good. We make it work more effectively and cost-efficiently. **AI for your processes and services** We develop custom AI technologies, tailored to your needs. We closely track the value created. - Generation of text, images, voice, and video - Semantic search in large datasets - LLM fine-tuning and RAG - Automated emails and correspondence - AI agents, assistants, and chatbots **Custom web and mobile apps** We develop applications, from simple to complex. State-of-the-art architecture, top design, seamlessly integrated into your systems. - Websites, platforms, and mobile apps - E-commerce systems and shops - Digital services and processes - Visualizations of your company data - GDPR/Swiss DPA-compliant data protection All Services ## Reference Projects Semantic search with vector embeddings ### Find relevant Federal Supreme Court decisions with AI Unlike classic keyword search, which only delivers exact matches, semantic search enables the discovery of thematically related and contextually similar topics. We developed such a semantic search technology for all 180,000 decisions of the Swiss Federal Supreme Court since 1950. This allows relevant judgments for a specific case to be identified from the entire dataset within seconds. This method has the potential to significantly influence court proceedings – up to decisions about freedom or detention. Vector embeddings Semantic search Natural Language Processing PostgreSQL Python TypeScript Next.js Legal informatics effektiv Labs 2025 ## This is effektiv GmbH With effektiv GmbH, we live a new model of a company in the age of artificial intelligence. All our processes are AI-driven. And yet, we remain human and approachable. Learn more about us **The AI revolution starts today** ## What is the value of AI for your company? Our goal is to develop a concrete prototype or a first version of your product as quickly as possible. Let's discuss together what we need to achieve this. Start your project now --- # This is effektiv GmbH > About us **Your Swiss partner for the AI revolution** With effektiv GmbH, we embody a new model of a company in the age of artificial intelligence: We work efficiently because we use AI in all our processes. From conception to design to development. Always, however, with our human vision, guidance, and quality control. **Direct collaboration** You always work directly with the management. We aim to stay in close contact with you from the initial meeting through successful implementation to long-term support. **Network of professionals** effektiv GmbH is run by one person, but it is not a one-man operation in the traditional sense. Depending on the project, we assemble a team of experienced specialists from our network in strategy, design, AI, and engineering. **Honest and well-founded** We are not interested in showmanship. We show you clearly and honestly which technologies you really need to create real added value in your company. **Made in Switzerland** We develop all software entirely here in Switzerland. Our network is also local. Our software runs on Swiss servers – or, if you wish, even on your own infrastructure. ## Your contact person ### Michael Baumann Founder AI Architect Lead Developer I bring complex concepts to life with entrepreneurial instinct and technical precision. As a doer with perseverance and dedication, I tackle challenges and find the right solution. [LinkedIn](https://www.linkedin.com/in/mbmnn/) [GitHub](https://github.com/mbmnn) [+41 77 225 59 74](tel:+41772255974) [info@effektiv.ch](mailto:info@effektiv.ch) #### Professional experience ##### Lead Developer Web & AI, Digital Business Designer confident Markenkommunikation, 2020–2025 - Lead development of digital platforms and AI - Architecture and implementation of web and AI projects - Digital business design - Development of a multi-tenant website builder with autonomous website creation capabilities - Development of a service and data platform with authentication, data visualization, event and marketplace services, and AI-powered personality profiles - Project management and client support ##### Co-Founder, Head of Development, Science Journalist Swiss science magazine higgs, 2017–2020 - Development and technical management of the Swiss science magazine higgs.ch - Development of automated content tools - Science journalism ##### Full Stack Developer, Science Journalist Agency scitec-media, 2015–2020 - Development of digital solutions for science communication - Science journalism ##### Research Associate ZHAW Zurich University of Applied Sciences, 2012–2015 - Research associate in the rectorate ##### Program Leader / Senior Scientist Biotech Molecular Partners AG, 2007–2012 - Leadership of research projects in life sciences ##### Science & Research: Life Sciences & Science Journalism University of Bern, 2002–2015 MAZ Lucerne, 2014-2015 - Conducting research projects in life sciences - Specialization in science journalism and science communication - PhD in Immunology - Master's in Human Genetics --- ## Unanswered questions? Find answers to the most frequently asked questions about our services and approach. Learn more about us --- # Our Services > From AI solutions to custom applications: We develop digital tools that drive your business forward. **Our Services** # Digital Tools for Your Success. From AI solutions to custom applications: We develop digital tools that drive your business forward. **AI for your processes and services** We develop custom AI technologies, tailored to your needs. We closely track the value created. - Generation of text, images, voice, and video - Semantic search in large datasets - LLM fine-tuning and RAG - Automated emails and correspondence - AI agents, assistants, and chatbots **Custom web and mobile apps** We develop applications, from simple to complex. State-of-the-art architecture, top design, seamlessly integrated into your systems. - Websites, platforms, and mobile apps - E-commerce systems and shops - Digital services and processes - Visualizations of your company data - GDPR/Swiss DPA-compliant data protection ## Further Services **Cutting-edge technologies** We use the latest technology to make your projects successful. We apply what is most suitable, such as Next.js, React Native, commercial and open source LLMs and machine learning, TypeScript, and Python. **Consulting for your digital future** We provide honest, strategic, and technically sound advice for your company's digital transformation. This gives you clarity and helps you avoid costly mistakes. **100% Swiss development** We develop all our projects entirely in Switzerland. All applications are also operated on Swiss servers. And, if you wish, even locally in your company. **Training for your team** How can you and your team use AI to your advantage? We train management and teams in tools like Copilot, ChatGPT, and many more. Uncomplicated, practical, and easy to understand. ## Ready for the next step? Let's discuss how we can advance your company with digital tools and AI. Get in touch --- # Contact > effektiv GmbH **This could be the beginning of something great.** Ready for the next step? Book a free initial consultation. **Phone** For quick questions or a spontaneous conversation, you can reach us directly by phone. [077 225 59 74](tel:+41772255974) **Email** For detailed inquiries or if you want to send documents in advance, you can reach us by email. [info@effektiv.ch](mailto:info@effektiv.ch) ## This is effektiv GmbH With effektiv GmbH, we embody a new model of a company in the age of artificial intelligence. All our processes are AI-supported. And yet, we remain human and approachable. Learn more about us --- # Frequently Asked Questions > FAQ Here you'll find answers to the most common questions about our services, our approach, and working with effektiv GmbH. ### Who is behind effektiv GmbH? effektiv GmbH is formed around Michael Baumann, founder and lead developer. With over 10 years of experience in developing cutting-edge applications and integrations, he has not only the necessary expertise but also the appropriate network for complex digital projects. ### What technologies does effektiv GmbH use for AI and app development? Our projects encompass cutting-edge technologies, including commercial and open source Large Language Models (LLMs), Vector Embeddings, Next.js, React Native, TypeScript, Python, PyTorch, Postgres databases, as well as REST and GraphQL APIs. All solutions are operated on Swiss infrastructure to ensure the highest data protection standards. ### Custom-made sounds expensive. How much does my project cost? Our knowledge, experience, and implementation competence are valuable – but that doesn't mean your project will be expensive. Through our lean structure and high efficiency, we are often more affordable than other companies. We'll gladly provide you with a cost estimate once we understand your needs and goals. And once our project starts, you can exit after each project phase. ### What projects does effektiv GmbH implement? We deliberately focus on digital solutions that reflect our core competencies and where we can create real added value: Custom web applications, mobile apps, internal digital tools, dashboards, e-commerce systems, as well as intelligent AI agents and assistants. We advise you specifically on which projects make sense and optimally match our capabilities. Once a project is live, we offer you reliable long-term support and targeted further development – always with the highest quality standards and clear strategic alignment. ### What company sizes does effektiv GmbH work with? We support companies of all sizes that seriously want to benefit from AI and digitalization. What's important to us is that you want to work with us to find out where the greatest potential lies in your company. And that you're just as interested in the subject matter as we are. ### How does a project typically proceed? Every project begins with a thorough assessment of your IT infrastructure and a strategy phase to identify the most effective applications for AI and digital solutions – including data protection checks and technical feasibility studies. Then we quickly create a first version of the idea that we want to test and develop further in the real world. ### Do you really develop everything in Switzerland? Yes. Every line of code in our applications and integrations is written in Switzerland – together by humans and machines. How you want to operate the software is up to you: This can be in the international cloud, on Swiss servers, or even on your own hardware. ### How does effektiv GmbH ensure sufficient resources for growing project needs? As a small company, we deliberately work agilely and personally – and that's exactly where our strength lies. We don't have large, rigid structures, but instead build a targeted network of proven experts that we activate on a project basis. We deliberately choose projects that match our capacities to ensure the highest quality. For you, this means not only efficient and high-quality implementation, but also fair costs. Because you don't pay for overhead and offices – just exactly our work on your project. --- # Privacy Statement > effektiv GmbH > **Translation Notice:** This is an automated translation. The German version is legally binding. **Last updated: 14 June 2025** ## Quick Summary **Website visitors:** We do not track you, run analytics, or collect personal data when you visit our website. Only the technical minimum (like your IP address) is stored temporarily by our server for security reasons. **Login/users:** If you log in to our services (on a separate subdomain), we store your name and email for authentication. Cookies are used to keep you logged in. **No advertising, no sharing:** We do not share your data with third parties unless required or agreed with you. **Your rights:** You can contact us any time to ask what data we have about you, or to request deletion. ## 1. Who We Are **effektiv GmbH Switzerland**\ Contact: Michael Baumann, info@effektiv.ch ## 2. Privacy for Website Visitors **What data is collected?** Only technical data your browser sends automatically, such as your IP address, browser type, and the time of your request. This data is stored temporarily in our server logs to ensure the website runs smoothly and securely. **What is not collected?** No analytics, tracking, marketing, or profiling tools are used. No cookies are set for regular website visitors. We do not store, sell, or otherwise process any personal data about you when you just visit our website. **Why is this data collected?** For technical operation and security (e.g., detecting misuse or attacks). **Legal basis (Swiss law):** Art. 6 FADP: Data processing is necessary for the proper operation and security of our website. ## 3. Privacy for Authenticated Users (Login Area) If you use our services that require login (usually on a separate subdomain), the following applies: **What data is collected?** - Your email address and name (provided by you on registration or invitation) - Cookies are used to manage your login session and keep you authenticated - Technical logs (e.g., IP address, time of login) for security and troubleshooting **How is data stored?** User data (email, name) is securely stored in our self-hosted Supabase instance. Data may also be processed or stored on servers provided by Hetzner (Germany) and Microsoft (email and files), in compliance with their respective data protection policies. **Why is this data collected?** To create and manage your user account, provide secure access, and operate the service you signed up for. **Who can access your data?** Only authorized staff of effektiv GmbH, and only as necessary to operate the service. We do not share data with third parties unless required by law or agreed with you (e.g., if you request a particular cloud or GPU provider). **How long is data kept?** As long as you have an active account or as agreed with you. You can request deletion of your data at any time (see below). ## 4. Data Transfers and Third Parties **Server Hosting:** Our servers are hosted by Hetzner in Germany (EU), and Microsoft (for email and, if agreed, file storage). These providers may process data in the EU or, in rare cases, other locations in line with their privacy policies. **No Data Selling/Sharing:** We do not sell or share your data for marketing or any other purpose. ## 5. Your Rights Under Swiss law, you have the right to: - Obtain information about what data we hold about you - Request correction or deletion of your data - Withdraw your consent (if any was given) at any time Just contact Michael Baumann at info@effektiv.ch for any privacy-related request. ## 6. Changes We may update this privacy statement as our services evolve or as laws change. We will post updates here and notify users with active accounts of any significant changes. ## 7. Contact If you have any questions about privacy at effektiv GmbH, contact: **Michael Baumann**\ info@effektiv.ch This privacy statement is designed for transparency and compliance with the Swiss Federal Act on Data Protection (FADP). For any special circumstances or concerns, please contact us directly. --- # Terms & Conditions > effektiv GmbH > **Translation Notice:** This is an automated translation. The German version is legally binding. ## 1. Subject Matter These General Terms and Conditions (GTC) govern the contractual relationships between effektiv GmbH (hereinafter "effektiv GmbH") and its customers (hereinafter the "Customers", collectively the "Parties"), unless and to the extent that no deviating agreements are made. ## 2. Quotation and Contract Formation ### 2.1 Quotation effektiv GmbH creates quotations for its customers based on customer inquiries. A quotation may include in particular the objectives of the respective project, the description of the services offered, the intended purpose and duration of use, the required timeframe, and the expected costs. The costs stated in a quotation and the expected duration of implementation of the services offered (hereinafter the "timeframe") are approximate values, unless the values are bindingly guaranteed. In particular, order changes before or during project execution as well as unexpected difficulties during execution may lead to deviating costs and timeframes. The first quotation is considered a guideline quotation. effektiv GmbH is bound to the quotation for 60 days. The quotation includes only the services explicitly listed therein. Changes or additions may lead to changes in costs or price and timeframe. Third-party services are regularly not included in the quotation but are offered separately (see Section 3.1). However, the quotation may indicate what third-party services are necessary for successful project execution. ### 2.2 Contract Formation The quotation is simultaneously designed as a contract. The contract between effektiv GmbH and its customers is concluded with the consent of both parties. ## 3. Obligations of the Contracting Parties ### 3.1 Obligations of effektiv GmbH effektiv GmbH performs its services with due care and to the best of its knowledge and belief. effektiv GmbH strives to adhere to an agreed timeframe. effektiv GmbH is entitled to engage auxiliary persons and third parties for contract fulfillment. ### 3.2 Obligations of Customers Customers are obligated to provide effektiv GmbH with all information required for the execution of services. Customers are also obligated to provide effektiv GmbH with all data and materials required for the execution of services and, if necessary, required infrastructure in a timely manner and free of charge. For timely project execution, effektiv GmbH depends on necessary project decisions being made by its customers in a timely manner and within the designated timeframe. This includes in particular the timely acceptance and review of (partial) results, the timely granting of "approval to proceed," and other steps in the respective project. ## 4. Scope of Services and Service Changes ### 4.1 Scope of Services The scope of services is determined by the services specified in the contract (hereinafter the "Contractual Services"). All agreements beyond the original contract text are considered agreements regarding service changes. ### 4.2 Service Changes Changes, additions, or extensions to the contractually agreed services are considered service changes (hereinafter uniformly the "Service Changes"). Service changes may lead to additional costs. Desired minor service changes by customers are calculated based on effort. effektiv GmbH may provide an estimate of additional costs. For desired other service changes by customers, effektiv GmbH offers the services arising in addition to the original contract anew. Author corrections are services caused by customers that lead to additional effort for effektiv GmbH, such as the delivery of incorrect or faulty data. Author corrections are considered service changes. They are billed to the respective customer based on effort. Customers give their consent to author corrections by agreeing to these GTC. ## 5. Deadlines The deadlines contained in the contract are guideline values, unless the deadlines are bindingly guaranteed in the contract. Particularly for projects with adaptive or agile approaches, deadline planning including timeframes represents only an approximate temporal objective. ## 6. Acceptance and Acceptance Procedure ### 6.1 Delivery and General For contracts of a work contract nature, effektiv GmbH delivers the contractually owed service (hereinafter the "Work") by handover to the respective customer. If contractually agreed, staged delivery of the work (hereinafter "Partial Delivery") is also possible. The respective customer is responsible for reviewing and complaining about any defects. ### 6.2 Review and Defect Notification Customers must immediately review delivered works – even with partial deliveries – and immediately complain about any defects. Hidden defects must be complained about within 10 working days of delivery for partial deliveries, and within 20 working days of delivery for a complete work. All defect notifications must be made in a form that allows proof by text. Without appropriate complaint, a delivered work or delivered work part is considered accepted. ## 7. Warranty and Liability effektiv GmbH warrants the provision of services or creation of work according to contractual agreements. Warranty for orally assured characteristics is excluded. effektiv GmbH is not liable for services of third parties that they provide in an independent capacity. Deviations from contractually owed services as well as defects must be complained about by customers within the deadlines mentioned in Section 6.2 and in the designated form. Without complaint within this deadline, service or work is considered accepted. For justified and timely defect complaints, customers have a right to rectification. effektiv GmbH must undertake rectification within a reasonable period and at its own cost. Reduction and rescission are excluded to the extent legally permissible. Rescission is particularly excluded if partial deliveries occurred and were accepted by customers. Any further warranty rights are excluded. Likewise excluded is, to the extent legally permissible, liability for consequential defect damages and indirect damages. Liability for intent and gross negligence remains reserved. ## 8. Remuneration ### 8.1 Fixed Price Remuneration When agreeing on a fixed price, remuneration is generally based on the contractual agreement. Quoted prices are always exclusive of statutory value-added tax. Services excluded according to Section 4.1 are not included in the fixed price. Such services are billed separately. Service changes according to Section 4.2 are also not included in the fixed price. In case of service changes, the procedure provided for in Section 4.2 applies. ### 8.2 Effort-Based Remuneration Without contrary contractual agreement, remuneration is measured by effort. ### 8.3 Expenses Material costs, travel, and other expenses are billed to customers at the actual costs incurred. ### 8.4 Hosting / DNS Entries Costs for hosting, DNS entries, and certificates are billed and payable within 30 days of invoicing. Any cancellations of services must be made in writing to info@effektiv.ch no later than 30 days before automatic renewal. Otherwise, costs for another year will be billed. ### 8.5 Larger Expenses / Artificial Intelligence Services Larger expenses, for example for Artificial Intelligence Services, may be payable in advance after consultation with the customer. effektiv GmbH informs about this during offer preparation or before concluding a corresponding individual order. ## 9. Payment Terms ### 9.1 General Invoicing occurs at the latest after delivery of the contractually owed service. The payment term is 30 days net, unless effektiv GmbH provides for a different payment term. effektiv GmbH may invoice one-third of the quotation amount each at contract conclusion (see Section 2.2) and at delivery (see Section 6.1). The final invoice follows subsequently. The payment term for all partial invoices is 30 days net. ### 9.2 Order Reduction or Contract Withdrawal If a customer withdraws from an already concluded contract, they must fully indemnify effektiv GmbH. Lost profit or positive contractual interest is also owed. If a customer reduces a given order, effektiv GmbH has at least a claim to the remuneration that would be owed up to the time of the reduction announcement, including proportional profit. If effektiv GmbH has specifically provided additional capacities for the order, such as additional employees or auxiliary persons, or if the capacities reserved for the order cannot be used otherwise, the relevant customer must also reimburse effektiv GmbH for the related expenses, including the proportional profit attributable thereto. If the service was already fully provided by effektiv GmbH at the time of the reduction announcement, full remuneration is owed. ### 9.3 Exclusion of Set-off Customers are not entitled – to the extent legally permissible – to set off any claims on their part against demands of effektiv GmbH. ## 10. Default ### 10.1 Default of effektiv GmbH Non-compliance with the intended timeframe does not automatically lead to default by effektiv GmbH. However, effektiv GmbH is obligated to inform customers about delays. If the intended timeframe cannot be met for reasons attributable to customers, Section 10.2 applies. To the extent that effektiv GmbH is not grossly at fault for a delay, it is not liable for any default damage. Incorrect assessment of technical or other difficulties (and the associated longer solution duration) does not constitute gross fault. ### 10.2 Default of Customers After expiration of payment deadlines according to Section 9.1, customers fall into default without further notice. In particular, no payment reminder is necessary. effektiv GmbH is entitled in this case to suspend work. Any deadlines bindingly guaranteed by effektiv GmbH lapse in this case. Customers are liable to effektiv GmbH for any additional effort. If the intended schedule cannot be met for reasons attributable to customers – for example, in case of violation of obligations according to Section 3.2 – customers are liable to effektiv GmbH for any additional effort. ## 11. Intellectual Property Rights Without other agreement, all intellectual property rights to created works and other services are transferred to the respective customer only after full payment of all works and other services as well as all outstanding claims of effektiv GmbH. Intellectual property personality rights that cannot be transferred by law remain reserved. To the extent that transfer of intellectual property rights is not possible, effektiv GmbH grants the respective customer a comprehensive license for use of the created works and other services. The license includes the right to use the works and other services without local, material, and temporal restrictions, for any purposes. ## 12. Confidentiality The parties commit to treat all information confidentially – particularly information about business occurrences, customers, projects, and procedures – that they learn from each other in the context of the relationship, and to refrain from passing such information to unauthorized third parties. The confidentiality obligation remains in effect beyond the duration of cooperation between the parties. The fact of cooperation between the parties is not considered confidential. effektiv GmbH is furthermore entitled to mention its activity for customers for its own advertising purposes. effektiv GmbH is also entitled to display or describe the communication means it developed on its own communication channels and in its own advertising materials. effektiv GmbH is also entitled to submit customer campaigns to competitions domestically and abroad. ## 13. Data Protection and Data Security The parties commit to comply with the provisions of Swiss data protection law as well as any other applicable data protection law. They commit to take economically reasonable as well as technically and organizationally appropriate precautions so that data arising in the context of contract processing is effectively protected against unauthorized knowledge by third parties. Further information on data processing can be found in the current data protection declaration of effektiv GmbH. ## 14. Severability Clause Should a provision of this agreement prove to be unenforceable, invalid, or ineffective, this should not affect the enforceability, validity, and effectiveness of the remaining provisions. In this case, the parties commit to replace the unenforceable, invalid, or ineffective provision with an enforceable, valid, or effective provision that comes closest in content and economically to the original intention of the parties. ## 15. Assignment of Claims effektiv GmbH is entitled to assign its claims against its customers to third parties or to commission third parties with collection and enforcement. ## 16. Validity of Other GTC The validity of any GTC of customers is expressly excluded. Only these GTC apply between the parties. ## 17. Applicable Law and Jurisdiction Swiss law applies to the contractual relationships between the parties, excluding the Vienna Sales Convention (CISG). Exclusive jurisdiction is at the seat of effektiv GmbH. **Contact:**\ Michael Baumann\ info@effektiv.ch --- # Imprint > effektiv GmbH ## Company Information **effektiv GmbH**\ Registered in Switzerland\ Handelsregister des Kantons Zürich\ CHE-402.847.133 **Managing Director:** Michael Baumann\ **Contact:** info@effektiv.ch ## Legal Notice This website is operated by effektiv GmbH, a company registered under Swiss law. All content on this website is protected by copyright and other intellectual property rights. Reproduction, distribution, or any other use of the content without prior written consent is prohibited. ## Disclaimer The information on this website is provided for general informational purposes only. While we strive to keep the information up to date and accurate, we make no representations or warranties of any kind, express or implied, about the completeness, accuracy, reliability, suitability, or availability of the website or the information contained on the website. ## Contact For any questions regarding this imprint or our website, please contact: **Michael Baumann**\ info@effektiv.ch --- # We Are Made in Switzerland > effektiv GmbH [](https://www.swissmadesoftware.org) We are proud to be a certified member of the **Swiss Made Software** label, representing the highest standards of quality, reliability, and precision in software development. This label identifies software products and services developed in Switzerland that meet rigorous standards including: - **Quality & Reliability**: Rigorous standards ensuring exceptional software quality - **Innovation**: Cutting-edge solutions developed by Swiss engineering excellence - **Data Protection**: Compliance with strict Swiss data protection regulations - **Digital Sovereignty**: Supporting Switzerland's digital independence As label holders, we exceed the label requirements by developing 100% of our software in Switzerland (compared to the minimum 60% requirement), ensuring complete Swiss quality and expertise in every line of code. We commit to delivering exceptional software solutions with transparency in our processes while contributing to Switzerland's digital ecosystem. The Swiss Made Software label is managed by [Swiss Made Software organization](https://www.swissmadesoftware.org), promoting Swiss software excellence since 2007. Visit Swiss Made Software --- # The Third Epoch of the Web *AI Agents* > Websites were built for humans. Then search engines arrived, and an entire industry emerged to make content readable for algorithms. Now AI agents are here—they don't index, they act. And they're already the majority. 2026-02-14 | Michael Baumann Since 2024, more than half of all web requests come from machines—for the first time in a decade, more than from humans. And they're no longer just search engine crawlers scraping your pages. They're AI agents: software that reads websites, understands them, and acts autonomously. Your website has gained a new audience. Uninvited. And this audience has very different needs. ## A Brief History of the Web in Three Acts **Act 1: Humans.** The early web was a document format. HTML described text for human eyes, and websites were found via link directories—Yahoo!, DMOZ, word of mouth. **Act 2: Search Engines.** In 1998, Google arrived, and suddenly there was an intermediary between content and reader: the crawler. An entire industry emerged—SEO—whose sole purpose was to make websites readable for an algorithm. Structured data, meta tags, Schema.org: tools to help machines understand what humans had long seen. **Act 3: Agents.** Now there's a new intermediary, and it's hungrier than any crawler. AI agents don't index—they consume, reason, and act. Gartner predicts that by 2028, 90% of B2B purchasing will run through AI agents. McKinsey estimates 3 to 5 trillion dollars in agent-driven revenue in consumer goods alone by 2030. The lesson from SEO was simple: if an algorithm mediates access, the content must work for that algorithm. The same lesson applies again—except the new algorithm doesn't just index. It understands, decides, and buys. ## Markdown, Content Negotiation, and the Hype In February 2026, Cloudflare launched "Markdown for Agents"—a feature that automatically converts HTML pages into Markdown when an AI agent requests them. Other companies—[including us at effektiv.ch](/md/en)—have built similar solutions at the application level. The idea is the same everywhere. But to put it in context, two brief explanations are needed. **Markdown** is a simple text markup language. The same content, without the overhead a browser needs for rendering. For AI systems that work internally with text rather than layout, Markdown is the natural format. Switch between the tabs—the difference speaks for itself: **Content Negotiation** is a mechanism that has existed since the early days of the web—HTTP/1.0 from 1996 already defined it. The client tells the server which format it prefers, and the server delivers the matching variant. This happens constantly today: | Situation | Client Says | Server Delivers | | --- | --- | --- | | Images | "I support WebP" | WebP instead of JPEG—same content, smaller file | | Language | "I prefer German" | German instead of English version | | Compression | "I support Brotli" | Compressed instead of raw—same content, more efficient | | API Data | "I want JSON" | JSON instead of XML—same content, leaner | | AI Agent | "I want Markdown" | Markdown instead of HTML—same content, no noise | The last case is new. The principle is thirty years old. And it's not cloaking—not deceptively serving different content—but the same principle by which your browser has been receiving smaller image formats for years. ## Can't AI Agents Just Read HTML? Yes. And this is where the discussion needs to get honest. Google's John Mueller called Markdown for bots a "stupid idea" on Bluesky. His argument: LLMs were trained on trillions of HTML pages. They can read HTML. Why bother with an extra format? Research partially supports him. HtmlRAG (WWW 2025) shows that HTML actually works **better** than plain text for structured data—because LLMs know the semantics of HTML tags from training. "Table Meets LLM" (WSDM 2024) confirms this for tables: HTML outperforms Markdown on several tasks significantly. And the larger the model, the less sensitive it is to format differences overall. Mueller has a point. LLMs don't fail at HTML. The problem lies elsewhere. ### The Big Noise Open the source code of any corporate website. The actual text often accounts for just 15 to 70% of the source code. The rest? Navigation, footer, cookie banners, CSS classes, JavaScript, tracking pixels, ad banners. Invisible to humans—the browser filters it out. For a language model that must process every token, it's noise. And this noise has measurable consequences. Researchers have demonstrated that 30,000 irrelevant tokens in the context window can push a language model's accuracy from 96% down to 11%. Not because the model can't find the information—but because the sheer volume of noise dilutes its attention. Imagine searching for a sentence in a book, but someone has glued three pages of phone directory between every page. Theoretically findable. Practically lost. JetBrains Research confirmed this in an experiment with 500 programming tasks: removing irrelevant context reduced costs by 52%—while simultaneously improving success rates. The value of Markdown doesn't lie in `##` being better than ``. It lies in the noise disappearing. It's not about format. It's about signal-to-noise. ## What This Means in Practice Anyone who lived through SEO recognizes the pattern. First, companies ignored search engines. Then they panicked. Then they learned that good content matters more than technical tricks. The agent web is following the same trajectory. ### What's Worth Doing - Structure content cleanly. Semantic HTML, clear headings, descriptive alt texts—not new advice, but with new weight. What's good for accessibility is also good for AI agents. Both use the same semantic tree. - Provide an /llms.txt file. The llms.txt standard by Jeremy Howard is a compact Markdown file that describes what your organization does and links to its most important pages. Over 844,000 websites have adopted it, including Anthropic, Stripe, and Shopify. Whether AI providers actively use the file remains an open question. But the effort is minimal. [→ View our /llms.txt](/llms.txt) - Offer machine-readable variants. On effektiv.ch, we've implemented this: every page exists as clean Markdown, including a token estimate. AI agents that send Accept: text/markdown are automatically redirected to it. In the footer, you can switch between the "Human" and "Machine" view—try it out. [→ View this article as Markdown](/md/en/blog/markdown-for-agents) Equally important is avoiding common mistakes: - Don't build a separate Markdown web. Markdown endpoints deliver the same content, not different content. Showing AI agents different information than humans is cloaking—and destroys trust. - Don't believe one format solves everything. For tables, HTML is better. For prose, Markdown is more efficient. For APIs, JSON is standard. The answer isn't "everything in Markdown" but the right format for the context. - Don't forget the basics. The most elegant Markdown conversion is useless if the content is thin. AI agents recognize substance—just like search engines have since the Panda update in 2011. ## Clarity Wins The web has always adapted to new audiences. With search engines, the winners weren't the websites with the most meta keywords but those with the best content. The agent web will be no different. The web is indeed learning a new language. But that language isn't Markdown. It's clarity. ### Sources - [Imperva/Thales: 2025 Bad Bot Report](https://cpl.thalesgroup.com/about-us/newsroom/2025-imperva-bad-bot-report-ai-internet-traffic) - [Cloudflare: Introducing Markdown for Agents](https://blog.cloudflare.com/markdown-for-agents/) - [HtmlRAG: HTML is Better Than Plain Text for RAG Systems (arXiv, WWW 2025)](https://arxiv.org/abs/2411.02959) - [Table Meets LLM: Can Large Language Models Understand Structured Table Data? (WSDM 2024)](https://arxiv.org/abs/2305.13062) - [Does Prompt Formatting Have Any Impact on LLM Performance? (arXiv)](https://arxiv.org/abs/2411.10541) - [Even Longer Contexts Degrade Generation Quality (arXiv)](https://arxiv.org/abs/2510.05381) - [JetBrains Research: Efficient Context Management for Coding Agents](https://blog.jetbrains.com/research/2025/12/efficient-context-management/) - [Gartner: AI Agents Will Outnumber Sellers by 10x by 2028](https://www.gartner.com/en/newsroom/press-releases/2025-11-18-gartner-predicts-by-2028-ai-agents-will-outnumber-sellers-by-10x-yet-fewer-than-40-percent-of-sellers-will-report-ai-agents-improved-productivity) - [McKinsey: The Agentic Commerce Opportunity](https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-agentic-commerce-opportunity-how-ai-agents-are-ushering-in-a-new-era-for-consumers-and-merchants) - [Google\](https://www.searchenginejournal.com/googles-mueller-calls-markdown-for-bots-idea-a-stupid-idea/566598/) - [llms.txt: Proposed Standard for AI Website Content](https://llmstxt.org/) - [Accessibility.Works: Do Accessible Websites Perform Better for AI Agents?](https://www.accessibility.works/blog/do-accessible-websites-perform-better-for-ai-agents/) - [HTTP/1.0 Specification (RFC 1945)](https://datatracker.ietf.org/doc/html/rfc1945) - [Sitechecker: What Is Text to Code Ratio?](https://sitechecker.pro/text-to-code-ratio/) --- # Ads in ChatGPT: Your Secrets, Monetized *AI Advertising* > Sam Altman called advertising a 2026-02-01 | Michael Baumann "I hate advertising." That was [Sam Altman](https://www.pcgamer.com/software/ai/here-we-go-openai-ceo-sam-altman-once-called-it-a-last-resort-but-chatgpt-is-about-to-get-stuffed-with-ads/), CEO of OpenAI, speaking at Harvard in 2024. Advertising, he said, was the "last resort" for his company—a business model that fundamentally pits user interests against corporate ones. The plan sounded noble: the wealthy would pay so everyone else could use ChatGPT for free. Robin Hood, but with language models. That was 2024. In January 2026—less than two years later—[OpenAI announced](https://openai.com/index/our-approach-to-advertising-and-expanding-access/) it would introduce advertising to ChatGPT. The "last resort" turned out to be a remarkably short path. ## $8 Billion Doesn't Burn Itself The about-face has a simple explanation: money. OpenAI [burned through roughly $8 billion in 2025](https://www.tomshardware.com/tech-industry/big-tech/openai-could-reportedly-run-out-of-cash-by-mid-2027-nyt-analyst-paints-grim-picture-after-examining-companys-finances)—more than most companies generate in a decade. Of ChatGPT's 800 million users, [only about 5%](https://www.cnbc.com/2026/01/16/open-ai-chatgpt-ads-us.html) pay for a subscription. The math is brutal. You can't serve a billion people for free while your data centers consume as much electricity as mid-sized cities. Someone has to pay the bill. Financial analysts at Deutsche Bank project a cumulative negative cash flow of $143 billion by 2029. Sebastian Mallaby warns that OpenAI [could go bankrupt by mid-2027](https://finance.yahoo.com/news/financial-experts-warn-openai-may-113057515.html) without new revenue streams. The same CEO who called advertising "dystopian" is now reaching for the dystopian toolkit himself. Altman's new explanation on X: "It's clear to us that many people want to use a lot of AI and don't want to pay." From idealist to pragmatist in under two years. Even by Silicon Valley standards, that's impressive. ## What ChatGPT Knows About You—and Google Never Did Here's where it gets interesting. Advertising in an AI chatbot is not the same as advertising in a search engine. The difference is fundamental—and should concern us all. With Google, you type "best running shoes." Google learns: this person is interested in running shoes. Useful for advertisers, but ultimately superficial. A snapshot of purchase intent. With ChatGPT, the game is different. People use the chatbot as [therapist, life coach, and confidant](https://techcrunch.com/2025/07/25/sam-altman-warns-theres-no-legal-confidentiality-when-using-chatgpt-as-a-therapist/). They share relationship problems, mental health struggles, financial worries—things they wouldn't tell even close friends. Sam Altman himself admitted: "People talk about the most personal things in their lives with ChatGPT. Especially young people use it as a therapist, as a life coach." A [Stanford study](https://news.stanford.edu/stories/2025/10/ai-chatbot-privacy-concerns-risks-research) confirms what many suspect: all six major AI companies use chat data by default to train their models. For corporations like Google and Meta, this data is also merged with information from other products—search queries, purchases, social media activity. A digital puzzle that reveals a surprisingly detailed picture. The researchers warn of a subtle dynamic: ask ChatGPT about low-sugar recipes or heart-healthy diets, and the system may classify you as "health-compromised." This assessment seeps through the provider's entire ecosystem. Suddenly you see pharmaceutical ads. And from there, it's not far until such information reaches an insurance company. ## The Difference from Google Advertising | Aspect | Google Search | ChatGPT | | --- | --- | --- | | Data Type | Search queries, clicks | Complete conversations, personal problems, emotions | | Information Depth | Surface-level purchase intent | Intimate life circumstances, mental state | | User Expectation | Transactional—searching for information | Trusting—conversation with an "advisor" | | Relationship Character | Tool | Quasi-human interaction | | Manipulation Potential | Influencing purchase decisions | Influencing in vulnerable moments | The core problem lies in the psychology of the interaction. Chatbots feel discreet, non-judgmental, almost intimate. People reveal things they would never share publicly—because it doesn't feel public. This trust amplifies the persuasive power of advertising in ways social media never could. The feed annoys. The chatbot understands. ## What OpenAI Promises—and What's Technically Possible OpenAI presented a list of safeguards in its [official announcement](https://openai.com/index/our-approach-to-advertising-and-expanding-access/). It reads reassuringly: - Ads do not influence ChatGPT's responses - Conversations remain private and are not sold to advertisers - Users under 18 won't see ads - No ads near sensitive topics like health, mental health, and politics - Paying subscribers (Plus, Pro, Business, Enterprise) won't see ads So far, so reasonable. But look closer and you'll spot the gaps—and above all: the conspicuous silence. **What OpenAI doesn't explain.** The announcement mentions that users can "turn off personalization" and "clear the data used for ads." But what exactly does that mean? OpenAI never defines which data is used for targeting. Just the current conversation? All previous chats? The Memory feature that remembers your preferences? The inferences the system draws from your questions—like "this person has health anxiety"? The phrase "clear data" implies that more is stored than just the current chat. What exactly remains in the dark. For a company that promises transparency, this is a remarkable omission. **The personalization paradox.** To show "relevant" ads at the end of a response, the system must analyze the conversation. OpenAI emphasizes it doesn't sell data—but they use it internally to personalize advertising. For users, this distinction is academic: your most intimate thoughts flow into algorithms designed to sell you products. **The gray zone of "sensitive topics."** OpenAI promises not to show ads during health and mental health discussions. But where exactly is the line? Is a conversation about stress already "mental health"? When does a financial question become a "sensitive" debt topic? These decisions are made by an algorithm—not an ethics board. **Zero legal protection.** Here's where it gets serious: there is [no confidentiality protection whatsoever](https://techcrunch.com/2025/07/25/sam-altman-warns-theres-no-legal-confidentiality-when-using-chatgpt-as-a-therapist/) for ChatGPT conversations. If authorities request chat logs, OpenAI must hand them over. This fundamentally distinguishes ChatGPT from an actual therapeutic conversation—even if it feels the same to many users. ## The Next 12 to 24 Months What comes next? Some predictions based on current developments: **Advertising becomes standard.** OpenAI is starting in the US, but global expansion is just a matter of time. The company is counting on [a billion dollars in ad revenue for 2026 alone](https://www.theinformation.com/articles/openais-ads-push-starts-taking-shape). If the model works, others will follow. Google has already introduced ads in AI Overviews. Perplexity is experimenting with sponsored questions. The dam has broken. **Ad formats will get subtler.** Currently, OpenAI plans ads "at the end of responses"—still clearly separated from actual content. But the industry is already testing integrated formats: sponsored product recommendations within responses, preferential brand mentions, "native" placements. The line between answer and ad will blur—until eventually it becomes invisible. **Privacy becomes a luxury good.** Ad-free as a premium feature creates a two-tier society: those who pay keep their privacy. Those who don't become the product. We know this from social media. But with a tool people use as a therapist, this dynamic takes on a new, disturbing quality. **Regulation lags behind.** The EU is working on the AI Act. Eventually regulators will take a closer look at how advertising works in AI systems. Whether this happens fast enough to protect users? Historically speaking: unlikely. ## What You Can Do Now Enough analysis. What can you actually do? - **Dig through the settings.** Look for opt-out options for personalized advertising and data usage for model training. It's tedious, but it's your right. - **Don't treat ChatGPT as a therapist.** No matter what the marketing suggests, no matter how understanding the responses sound: there's no legal protection for your conversations. Everything you type can theoretically be shared. Act accordingly. - **Consider alternatives.** Open-source models like Meta's Llama or Mistral can run locally—with no ads, and without your data ever leaving your system. For businesses, such solutions are often the safer choice. - **Pay—or accept the consequences.** This sounds cynical. But it's been the reality of the internet for 20 years: if you don't want to be the product, you have to become the customer. OpenAI is no exception. Altman was right: advertising puts user and company interests in conflict. He's now created that conflict himself. The AI industry is repeating the mistakes of the old internet—only faster and with more intimate data. ### Sources - [OpenAI: Our approach to advertising and expanding access to ChatGPT](https://openai.com/index/our-approach-to-advertising-and-expanding-access/) - [CNBC: OpenAI to begin testing ads on ChatGPT in the U.S.](https://www.cnbc.com/2026/01/16/open-ai-chatgpt-ads-us.html) - [Stanford Report: Study exposes privacy risks of AI chatbot conversations](https://news.stanford.edu/stories/2025/10/ai-chatbot-privacy-concerns-risks-research) - [PC Gamer: Sam Altman once called ads a "last resort"](https://www.pcgamer.com/software/ai/here-we-go-openai-ceo-sam-altman-once-called-it-a-last-resort-but-chatgpt-is-about-to-get-stuffed-with-ads/) --- # Swiss SMEs Lead the Way with AI *AI Success Stories* > Even though many AI projects fail, more and more Swiss SMEs prove this: with clear problem selection and pragmatic implementation, AI initiatives can absolutely succeed. 2025-10-17 | Michael Baumann Most generative AI pilot projects fail — as shown by a recent [MIT study](https://fortune.com/2025/08/18/mit-report-95-percent-generative-ai-pilots-at-companies-failing-cfo/). However, the problem usually doesn't lie in the technology itself but in organizational mistakes: unrealistic goals, lack of integration, insufficient data infrastructure. Let's take a closer look at the companies that have succeeded in creating value with AI. ## Small Companies Have an Advantage Especially small companies and organizations may have a significant advantage here. With manageable structures and shorter decision-making paths, successful AI deployments can often be realized much faster — provided one knows what truly matters. In short, these factors are key: - **Real Value**: AI must solve a concrete problem, not just impress as a "fancy" prototype. If it doesn't generate measurable time savings, cost reduction, or higher accuracy, it's probably not worth it. - **Simple Start**: Focus on sub-processes rather than the entire value chain right away. A pilot with 1–2 use cases instead of an all-in-one solution. - **Data Quality & Data Availability**: Use existing documents, texts, sensors, or systems. These data already exist in digital form and are often clean enough to be used directly. - **Seamless Integration**: AI must not exist in isolation. Integration with CRM, ERP, or workflow systems is essential. - **Governance & Control**: Data protection, traceability, and error control must be built in from the start. - **Iterative Development**: Learn, improve, and scale step by step. Pilot → Minimum viable version → Expansion. ## Where AI Actually Works in Swiss SMEs Based on published studies, practical guidelines, and the growing use of AI in Swiss companies, the following application areas can be identified where AI is already delivering real value: ### 1. Compliance, Risk Case Analysis & Fraud Detection In the financial sector and payment services, AI used for pre-screening transactions and suspicious cases has a strong impact. For example: with the support of [Innosuisse](https://www.innosuisse.ch/inno/en/home.html), the Ticino-based SME Cube Finance developed an intelligent system that detects fraudulent bank transactions more efficiently, combating money laundering. Manual checks are significantly reduced, and risk analyses accelerated. **Why it works**: Clear rules, high data availability (transaction logs), large manual effort, high benefit through error prevention. ### 2. Document Extraction & Text Analysis in Insurance AI-powered intelligent document processing can reduce processing time by up to 80% and error rates by around 90% compared to manual methods. Claims assessors and insurance experts can document, evaluate and compare much faster. **Concrete Swiss examples:** - **SPS Switzerland** developed an AI-based OCR solution for Swiss insurers. In its so-called "SPS Insurance Factory," physical and digital documents are automatically classified and data is extracted. The solution processes incoming documents from the mailroom all the way into insurers' core systems — laying the foundation for full straight-through processing. - The Swiss startup **inait.ai** developed "Bumpt," an AI system for highly precise vehicle damage detection, which won first place at the [Swiss Insurance Innovation Award 2023](https://www.swiss-insurtech.com/awards). The solution significantly accelerates damage assessment processes. **Why it works**: Existing documents as input, structured outputs, previously high manual workload, high cost of errors. ### 3. Back Office & Correspondence A recent study by the research institute [Sotomo](https://sotomo.ch/), commissioned by Axa Switzerland, shows that around half of the 300 surveyed Swiss SMEs use artificial intelligence for translation and correspondence tasks. 38 percent use AI for marketing texts. This confirms that AI is now routinely used to propose, translate, or summarize emails, contracts, and proposal drafts. **Concrete Swiss examples:** - **Bexio** from Rapperswil has been using AI since 2023 through its "Scan2Go" system to automatically allocate bank transactions and extract data from invoices and receipts using OCR. The AI-based document extraction recognizes content even without QR codes and enables automatic booking — significantly simplifying the processing of back-office correspondence such as receipts. - **Abacus** from Wittenbach uses "DeepO" for AI-based invoice processing with automatic posting. This machine-learning system turns unstructured data into structured data, massively reducing manual entry in contracts and financial documents and enabling full automation of the accounting process. - **Klara** from Lucerne offers AI-supported automated invoicing and payment reconciliation. The system automatically reads PDF invoice data, provides booking recommendations, and handles payment initiation and accounting reconciliation — all powered by learning AI technology. - **Yokoy** from Zurich specializes in AI-based expense and supplier invoice management with integrated fraud detection. The solution automatically reads receipts, validates them, checks for rule violations, and prepares the accounting journal including VAT — with an automation rate of up to 90% and proven processing cost savings of around 80%. - Swiss cloud provider **Infomaniak** integrated an AI writing assistant into its kSuite mail service in 2023. Based on open-source technology, the solution processes all data exclusively in Switzerland and helps companies handle email communication efficiently. **Why it works**: Low risk, immediately measurable time savings, low entry barrier. ### 4. Quality Assurance and Predictive Maintenance For manufacturing SMEs with machines, sensors, or visual inspection systems, AI is often ideal for detecting component defects or predicting maintenance needs. A comprehensive study by [ETH Zurich](https://ethz.ch/), in collaboration with Swissmem and Next Industries, shows that predictive maintenance and machine optimization are among the most important application areas for industrial AI in Switzerland. **Concrete Swiss examples:** - The Swiss data science specialist **LeanBI** has developed predictive maintenance solutions using acoustic sensors for industrial companies. These sensors capture sounds of critical machine components such as motors, bearings, or gearboxes and analyze them with machine learning to detect impending failures early. - **Parametric**, based in Switzerland, offers its RET3000 system for AI-based predictive maintenance in the railway sector. The system monitors sensor data such as vibrations and temperatures of components like bearings and motors in real time, detects anomalies via machine learning, and sends automated alerts — minimizing downtime and optimizing maintenance. - The **Zurich University of Applied Sciences (ZHAW)**, in collaboration with **Fluence Energy** (which includes Nispera since 2022), developed a hybrid AI model for solar power plants. This "physics-informed AI" system combines deep learning with physical models to diagnose energy losses caused by defects like soiled modules and plan maintenance economically. It achieves 70% better fault detection than conventional AI models and enables much more cost-efficient maintenance planning. **Why it works**: Existing sensor data, clear benefits from avoiding downtime, high costs in case of failures. The Swiss industry benefits from an advanced state of automation and digitalization, as well as well-established networks and easily accessible expertise. ## How to Make Your AI Project a Success Based on insights from successful projects, here is a practical checklist: **Step 1: Problem Selection & Use Case Definition** - Choose a narrow, clearly defined use case, e.g., "invoice document classification," not "complete automation of accounting." - Check whether historical data is available and digitally accessible (e.g., text files, ERP logs). - Estimate the potential value (monetary, time savings, error reduction). **Step 2: Data & Infrastructure Preparation** - Clean your data (duplicates, inconsistencies). - Build a pipeline that automatically transforms data. - Ensure interfaces to your systems (ERP, CRM, document repositories). **Step 3: Develop & Test the Pilot** - Create a minimum viable product (MVP) with core functionality. - Run human checks in parallel to discover errors. - Define clear KPIs (e.g., accuracy, time savings). **Step 4: Governance & Quality Assurance** - Define responsibilities (who checks, who corrects). - Track versioning, logging, feedback loops. - Integrate data protection and security mechanisms. **Step 5: Scaling & Further Development** - Evaluate extensions (additional use cases, more data sources). - Learn from errors and optimize iteratively. - Embed AI into your business processes — not as a standalone solution. ## The AI Journey Is Not a Sprint — It's a Marathon Successfully implemented projects show: it works when approached with focus, pragmatism, and responsibility — and when persistence is maintained over time. If you're planning an AI initiative in your SME — in production, customer support, document workflows, or compliance — we're happy to support you with use case design, data assessment, feasibility analysis, and of course the implementation of a pilot project. ### Sources - [Fortune: MIT report: 95% of generative AI pilots at companies are failing](https://fortune.com/2025/08/18/mit-report-95-percent-generative-ai-pilots-at-companies-failing-cfo/) - [Innosuisse: Artificial intelligence at the service of finance professionals](https://www.innosuisse.ch/inno/en/home.html) - [Swiss InsurTech: Swiss Insurance Innovation Award 2023](https://www.swiss-insurtech.com/awards) - [Sotomo: Research institute for social and political research](https://sotomo.ch/) - [ETH Zurich: The state of AI in the Swiss tech industry](https://ethz.ch/) - [ZHAW: Intelligent operation of solar power plants: predictive maintenance using a hybrid AI model](https://www.zhaw.ch/) --- # The Biggest Pitfalls in Your AI Project *AI Implementation* > Many companies invest in AI and then experience bitter disappointment. The technology promises much but often fails due to avoidable mistakes. However, there is a better way – pragmatic and success-oriented. 2025-09-22 | Michael Baumann Reality often looks different than marketing promises suggest. AI is not a magic formula that solves all your company's problems overnight. It is a tool with clear strengths — but also weaknesses and often underestimated costs. Those who introduce AI thoughtlessly quickly fall into expensive traps. We show you what the most common mistakes are and how you can avoid them. ## The five most common dangers (and our tips) ### 1. Hallucinations cost money and trust Generative AI can sound convincing while being completely wrong. This is not just embarrassing — it becomes expensive. A precedent case from Canada shows the consequences: [Air Canada was ordered to pay](https://mccarthy.ca) because the website chatbot gave a customer false information. The court made it clear: the company is liable for its bot's statements. The lesson from this: for all AI touchpoints with customers, 'human-in-the-loop' plus clear approval processes apply — especially for legal, financial, or safety-relevant information. **Our tips:** establish multi-level quality control. First: define critical areas where AI only serves as a draft — never as a final answer. Second: train your employees in 'prompt engineering' to get more precise AI responses. Third: use retrieval-augmented generation (RAG), which links AI with your own verified data sources. Companies that follow these approaches can significantly reduce hallucinations. ### 2. The pilot trap: testing without end Many companies start enthusiastically with pilot projects but never make the leap to productive operation. A current survey of 600 data leaders confirms the dilemma: two-thirds are stuck in generative AI pilots. And a full 97 percent have difficulties proving business value at all. Analysts also warn of inflated expectations: according to [Gartner](https://www.gartner.com/en/newsroom/press-releases/2025-06-15-gartner-predicts-40-of-agentic-ai-projects-to-fail-by-2027), more than 40 percent of so-called 'agentic AI' projects are likely to be discontinued by 2027 — due to rising costs, unclear business benefits, and inadequate risk controls. **Our tips:** think from productive operation backwards. Define clear success criteria before the pilot: measurable KPIs, maximum implementation costs, and realistic timeframes. US SMEs are already using AI broadly: 60 percent use AI platform tools — successful companies start with simple but clearly definable applications like automated email categorization or document search. For common pitfalls in pilots and how to get to production, see our piece on [AI pilot failures](/blog/ai-pilot-failures). ### 3. Legacy systems as reality check The demo works flawlessly — in corporate reality, it then fails due to interfaces from the 2000s, authorization concepts that no one understands anymore, or data quality that doesn't deserve the name. This is where many projects fail: not at the AI model, but at integration into the grown IT and process landscape. **Our tips:** make an honest inventory of your IT infrastructure before implementing AI. Focus initially on cloud-based AI tools that can be integrated via APIs, rather than complex on-premise solutions. The Boston Consulting Group shows with the 10–20–70 rule: successful AI transformations invest only 10% in algorithms, 20% in data and technology, but 70% in people, processes, and cultural changes. Successful SMEs often start with SaaS-based AI solutions for clearly defined areas — such as intelligent document analysis or CRM integration — and then gradually expand their infrastructure. ### 4. Hidden costs add up In addition to model licenses, there are expenses for data preparation, prompt engineering, monitoring, governance, training, and change management. [McKinsey shows in a current study](https://mckinsey.com): companies that have completely rethought workflows and clearly assigned roles see real benefits. **Our tips:** calculate realistically and plan 2–4× the initial license costs for integration and change management, depending on complexity. First identify your biggest pain points: do employees spend hours searching through documents? Then [semantic search](/blog/semantic-search) might be more valuable than a general-purpose chatbot. Do you work internationally? Automated translations can bring more ROI than marketing AI. Do you have many repetitive data analyses? Specialized AI solutions for your industry often outperform standard tools. The key is to identify the greatest benefit and start there, rather than immediately tackling expensive all-purpose solutions. ### 5. Features instead of strategy Many SMEs only use AI 'along with' existing tools because it's embedded in them — without a clear vision. This isn't fundamentally bad, but rarely sufficient for structural improvements. The strategic component is missing. **Our tips:** develop an AI roadmap with concrete business goals. Many SMEs expect a strong impact of AI on their industry in the next three to five years — use this advantage strategically. Identify three to five business processes with the greatest improvement potential and prioritize them according to expected ROI and implementation effort. Companies that proceed systematically achieve significantly better productivity gains than those with ad-hoc approaches. Also explore our perspective on the [benchmark problem](/blog/benchmark-problem) when evaluating AI performance in your context. ## Three proven entry points for your SME So where to start? The key lies in a pragmatic approach: start small, stay measurable, improve iteratively. Those who start with moderation today build the competence that will make the difference tomorrow. Here are three sensible entry points into the world of AI: ### Assisted knowledge work - Automation of recurring document work - AI-supported email processing and categorization - Intelligent appointment coordination and meeting preparation **ROI expectation**: studies show 12% more completed tasks and ~25% faster processing in controlled experiments with consulting companies (Harvard Business School/BCG). ### Semantic search - Searching through document inventories, manuals, and knowledge databases - Intelligently linked search results instead of pure keyword matches - Multilingual search for internationally active SMEs **ROI expectation**: significant time savings in information search and higher hit quality through context-based search. ### AI-supported customer service - Initial categorization and routing of customer inquiries - Suggestions for standard responses with approval requirement - Sentiment analysis for prioritizing urgent cases **ROI expectation**: reduction in processing times and improved customer satisfaction with proper implementation. ### Sources - [Moffatt v. Air Canada: Liability for chatbot statements](https://www.mccarthy.ca/en/insights/blogs/techlex/moffatt-v-air-canada-misrepresentation-ai-chatbot) - [The Guardian – Air Canada ordered to pay customer misled by chatbot](https://www.theguardian.com/world/2024/feb/16/air-canada-chatbot-lawsuit) - [Informatica CDO Insights 2025 – Key findings](https://www.informatica.com/about-us/news/news-releases/2025/01/20250128-global-data-leaders-seek-to-harness-the-power-of-genai-for-ai-driven-success.html) - [CIO Dive – Stuck in the pilot phase (survey of 600 data leaders)](https://www.ciodive.com/news/enterprise-generative-AI-ROI-pilot-fail-Informatica/739485/) - [Reuters – Gartner: >40% agentic AI projects scrapped by 2027](https://www.reuters.com/business/over-40-agentic-ai-projects-will-be-scrapped-by-2027-gartner-says-2025-06-25/) - [BCG (2025) – Closing the AI Impact Gap (10–20–70; 2.1× ROI leaders)](https://www.bcg.com/publications/2025/closing-the-ai-impact-gap) - [McKinsey (2025) – From models to business value (reimagine workflows)](https://www.mckinsey.com/capabilities/quantumblack/our-insights/from-models-to-business-value) - [PwC Germany (2025) – AI drives 4× productivity growth and 56% higher salaries](https://www.pwc.de/de/pressemitteilungen/2025/KI-sorgt-fuer-vierfaches-Produktivitaetswachstum-und-56prozent-hoehere-Gehaelter.html) - [IBM IBV – The Promise of Growth for B2B Sales (NPS 16% → 51%)](https://www.ibm.com/think/insights/business-value-promise-of-growth-for-b2b-sales/) - [U.S. Chamber (2025) – Empowering Small Business Report (60% use AI platforms)](https://www.uschamber.com/assets/documents/Empowering-Small-Business-Report-2025.pdf) - [HBS/BCG Field Experiment – Jagged Frontier (12.2% more tasks; ~25% faster)](https://www.hbs.edu/faculty/Pages/item.aspx?num=64700) - [Stanford (2025) – Legal RAG Hallucinations study](https://dho.stanford.edu/wp-content/uploads/Legal_RAG_Hallucinations.pdf) - [NIST AI Risk Management Framework (GenAI Profile)](https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf) - [ISO/IEC 42001 – AI Management Systems](https://www.iso.org/standard/42001) - [CIO Dive – AI project failure rates are on the rise](https://www.ciodive.com/news/AI-project-fail-data-SPGlobal/742590/) --- # Apertus is here: What can the Swiss LLM really do? *Swiss AI Reality Check* > In July we asked: «Can the Swiss LLM keep up?» – now the first generation is here with Apertus. Time for a sober reality check: Where does Swiss AI really stand, and who is it interesting for today? 2025-09-05 | Michael Baumann This article builds on our [July report](/blog/swiss-llm) – this time with solid benchmarks and practical experience. In July we asked: «Can the Swiss LLM keep up?» – now the first generation is here with Apertus. Time for a sober reality check: Where does Swiss AI stand – and who is it useful for today? ## The Bottom Line - **Complete Openness**: Apertus is not just «open weights,» but consistently transparent – weights, training data, and code are public. This level is rare. - **Language Diversity**: 15 trillion tokens from over 1,800 languages, 40% non-English – including Swiss German and Romansh. But this isn't (yet) practical – see tests below. - **Regulatory Clean**: Data protection and EU requirements were considered from the start rather than retrofitted. ## What's New with Apertus? Apertus comes in two sizes: 8 billion and 70 billion parameters. It was trained on the Swiss supercomputer «Alps» (CSCS, Lugano) – 10,752 NVIDIA GH200 Grace-Hopper chips on an HPE Cray platform. The training philosophy is unusual: Already in pretraining, 15 trillion tokens from more than 1,800 languages were used; post-training covers 149 languages. A special «Goldfish Objective» is supposed to prevent the model from memorizing content verbatim – measurements show practically baseline levels. Crucial for companies: The data pipelines strictly respect licenses and honor subsequent opt-outs (robots.txt). This creates an EU AI Act-compliant foundation. ## The Numbers: Where Does Apertus Stand in Comparison? | Model | MMLU (Knowledge) | Global-MMLU (Multilingual) | GSM8K (Math) | HumanEval (Code) | RULER @32k (Long Context) | | --- | --- | --- | --- | --- | --- | | Claude 3.5 Sonnet | 88.7% | — | 96.4% | 92.0% | — | | Llama 3.1 70B | 83.6% | — | 95.1% | 80.5% | — | | Apertus-70B | 69.6% | 62.7% | 77.6% | 73.0% | 80.6% | | Apertus-8B | 60.9% | 55.7% | 62.9% | 67.0% | 69.5% | **Notes on Comparability:** The prompt setups differ between models (shot numbers and chain-of-thought configurations). Global-MMLU and RULER values are not available in the official documentation for the comparison models. The 70B variant convinces in general knowledge and multilingual tasks, but remains behind the top models in mathematics and programming. ## Who Is Apertus Useful for Today? **Suitable for:** - Compliance-critical environments (public sector, healthcare, law, finance in EU/CH) - High transparency requirements – complete traceability of functionality - Summarization, classification, and categorization tasks **Not yet optimal for:** - Texts in Swiss German or Romansh - Mathematics-intensive automation (code refactoring, formal proofs) – lacks RL fine-tuning and specialized tool chains - Agentic workflows and multimodality – not the focus of this first generation ## Conclusion: Solid Start – Not Yet a Swiss Army Knife of LLMs Apertus is an important signal for open AI development in Europe – but (not yet) a breakthrough. The much-touted multilingual capabilities don't convince in practice yet. On paper it looks respectable: Apertus-70B translates German→Romansh with a BLEU score of 27.8 – clearly ahead of Llama-3.3-70B with 21.6. In application, however, this often results in unreadable text. ChatGPT currently delivers significantly better results here. Swiss German also showed weak in initial tests: The outputs sound neither like the requested dialect (Bernese German) nor generally like Swiss German – practically unusable. Nevertheless: The development is exciting. For specific, clearly defined use cases, Apertus can already fit today – but this requires further, targeted tests. The next versions will be decisive: for Swiss AI ambitions as well as for the question of whether small languages have a chance in the AI world. Fundamentally, it remains open what value radical openness will have in business everyday life. Sobering could be: If an ethically curated dataset ultimately means a weaker LLM, it will be difficult for the Swiss LLM. ## Availability and Access Apertus is now available through: - **Swisscom** (Sovereign AI) - **Hugging Face** (Open Source) - **Public AI** (API access) ETH and EPFL provide complete documentation and code. ### Sources - [Apertus Technical Report](https://github.com/swiss-ai/apertus-tech-report/blob/main/Apertus_Tech_Report.pdf) - [ETH Zürich Press Release](https://ethz.ch/en/news-and-events/eth-news/news/2025/09/press-release-apertus-a-fully-open-transparent-multilingual-language-model.html) - [EPFL Announcement](https://actu.epfl.ch/news/apertus-a-fully-open-transparent-multilingual-lang/) - [CSCS Press Release](https://www.cscs.ch/science/computer-science-hpc/2025/apertus-a-fully-open-transparent-multilingual-language-model) - [Hugging Face Model](https://huggingface.co/swiss-ai/Apertus-70B-2509) - [SWI swissinfo.ch](https://www.swissinfo.ch/eng/swiss-ai/switzerland-launches-transparent-chatgpt-alternative/89929269) - [Swiss AI Initiative](https://www.swiss-ai.org/apertus) - [Effektiv.ch Preliminary Report](https://effektiv.ch/de/blog/schweizer-llm-mithalten) - [Anthropic Claude 3.5 Sonnet Model Card](https://www-cdn.anthropic.com) - [Meta Llama 3.1 Technical Report](https://ar5iv.org) --- # Why 95% of AI Pilot Projects Fail – and What Successful Companies Do Differently *AI Implementation* > An MIT study reveals: 95% of all generative AI pilot projects in companies fail. But the successful 5% have one thing in common: They focus on strategic integration rather than technology hype. 2025-08-19 | Michael Baumann 95% of all generative AI pilot projects in companies fail. That's the sobering result of a new [MIT study](https://fortune.com/2025/08/18/mit-report-95-percent-generative-ai-pilots-at-companies-failing-cfo/) that interviewed 150 executives, surveyed 350 employees, and analyzed 300 public AI deployments. So does this mean the whole AI thing is worthless? Is the bubble bursting, as the media is gleefully proclaiming? The truth is much more nuanced. Alongside the massive 95% failure rate, there are indeed 5% of successful AI pilot projects. Let's examine what these projects are doing right. ## The Problem: Technology Without Strategy The MIT researchers led by Aditya Challapally identified a clear pattern in failed projects: Companies deploy generic tools like ChatGPT and expect them to automatically fit into their workflows. But that doesn't work. According to Challapally, generic tools like ChatGPT are ideal for individuals because they're flexible. However, they fail in corporate environments because they can't easily learn and often can't adapt to specific workflows. The result: Significant budgets flow into pilot projects that often deliver impressive demos but show no measurable impact on revenue or profit – in 95% of cases. ## What Successful Companies Do Differently So what are the remaining 5% doing right? The study reveals three critical factors for successful AI implementations: - **Partnerships Instead of In-House Development**: Companies that purchase AI tools from specialized providers and form strategic partnerships are about twice as successful as those developing their own systems (~67% vs. ~33% success rate). - **Back-Office Automation Instead of Marketing Hype**: According to MIT, companies invest over 50% of generative AI budgets in sales and marketing. But that's not where the highest return on investment (ROI) is achieved. Instead, it's in back-office automation – such as reducing external agency costs and optimizing internal processes. - **Line Managers as Drivers**: Successful companies don't just rely on centralized AI tools, but empower line managers to drive AI adoption. These managers know the concrete problems their teams face and can deploy AI strategically. ## Swiss Companies: How AI Integration Works What does this mean for Swiss companies? Based on the MIT report findings, we recommend: ### 1. Start with Back-Office Processes Don't begin with marketing or customer service. First automate internal, repetitive processes: - Document processing and classification - Data validation and cleaning - Routine reporting - Internal search functions ### 2. Choose Flexible, Learning Tools Avoid rigid, generic solutions. Look for tools that: - can integrate into existing workflows - learn from feedback and adapt - can grow with your data and processes ### 3. Focus on Partnerships Don't develop everything in-house. Leverage the expertise of specialized providers who: - understand your industry - meet Swiss data protection requirements - have proven implementation methods ### 4. Measure Real Value Define clear Key Performance Indicators (KPIs) before implementation, such as: - Time savings in hours per week - Cost savings in Swiss francs - Quality improvements (fewer errors, faster processing) - Employee and customer satisfaction ## The Future: Agentic AI Systems The report outlines that agentic, learning systems (with memory and limited autonomy) could shape the next phase; some companies are already experimenting with this approach. This development will define the next phase of enterprise AI. Companies that lay the right foundations now will be prepared for this future. ## Beware of Quick Conclusions The MIT study makes one thing clear: Just because something works with AI doesn't automatically make it better. But the insights are more nuanced than they appear at first glance: - **Partnerships vs. "In-House Development"**: Developing an AI tool in-house isn't inherently bad – but you should choose the right partner for it. The successful 5% rely on strategic collaborations with providers who understand their industry and have proven methods. - **Generic Tools vs. Specialized Solutions**: The criticism of ChatGPT and similar tools refers to their direct use in companies without specific adaptation. Properly configured and embedded in company-internal interfaces and processes, ChatGPT's language model does deliver successes. It's like cooking: The ingredients alone don't make the perfect taste experience – it's the refined recipe. - **Local Expertise**: The study confirms the approach of local expertise. Swiss companies benefit from partners who understand the FADP (Swiss Federal Act on Data Protection) and GDPR, develop on Swiss infrastructure, and know the specific challenges of SMEs. The AI revolution is underway. And it rewards those who think strategically and choose the right partners first. ### Sources - [Fortune (MIT Report Summary): 95% of generative AI pilots at companies are failing](https://fortune.com/2025/08/18/mit-report-95-percent-generative-ai-pilots-at-companies-failing-cfo/) - [MIT Media Lab – NANDA Initiative](https://www.media.mit.edu/) --- # The Benchmark Problem: Why AI Tests Are Becoming Less Meaningful *AI Evaluation* > AI models are becoming so good that classical benchmarks are reaching their limits. The result: manipulation, saturation, and a race for increasingly difficult tests. What does this mean for companies? 2025-07-25 | Michael Baumann Large Language Models (LLMs) are becoming increasingly powerful, yet their evaluation is paradoxically becoming more difficult. We're facing a fundamental benchmark problem: Many established tests are practically "exhausted," while new challenges show where even the best models still fail. ## The Saturation Problem: When 95% Becomes Meaningless The core issue: Many classic benchmarks are reaching saturation. Models like GPT-5, Claude 4, and Gemini 2.5 achieve over 90% on tests that were once considered challenging. This creates a paradoxical situation where differences between top models become increasingly difficult to measure. **Concrete examples of outdated tests:** The saturation becomes particularly evident when looking at the development. [The AI Index 2025 shows clear progress and convergence on several benchmarks](https://hai.stanford.edu/ai-index/2025-ai-index-report), for example, a jump from 4.4% (2023) to 71.7% (2024) on SWE-bench. | Benchmark | Status | Problem | | --- | --- | --- | | [Classic MMLU](https://arxiv.org/abs/2009.03300) | Frontier LLMs over 90% | Hardly any differentiation in the top tier | | [GSM8K](https://arxiv.org/abs/2503.04618) | High scores possible | Best-of-256 sampling achieves 97.7% (not Pass@1) | | HumanEval | Largely solved | No longer discriminative for modern coding assistance | This is why harder successors like [MMLU-Pro](https://arxiv.org/abs/2406.01574) (with 10 answer options instead of 4, more prompt-stable) and variants with contamination controls like [MMLU-CF](https://arxiv.org/abs/2412.15194) are emerging. This development shows why [leaderboards are becoming more important](https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard) – they can adapt, but also bring potential biases. ## When Optimization Becomes Manipulation The problem is exacerbated by strategic optimization that borders on manipulation. Verified examples: **Selective Benchmark Execution**: [OpenAI runs only 477 of 500 tasks on SWE-bench Verified](https://openai.com/index/introducing-gpt-5-for-developers/) ("solutions did not reliably pass on our infrastructure") and omits 23. This can artificially inflate scores – the stated value of 74.9% would reduce to about 71.4% when extrapolated to the full set. | Benchmark Variant | Tasks | Stated Score | Extrapolated Score (500 Tasks) | | --- | --- | --- | --- | | OpenAI (selective) | 477/500 | 74.9% | ~71.4% | | Other Labs (complete) | 500/500 | Directly comparable | - | **Arena Controversies**: [Reports about "specially crafted, non-public variants"](https://www.theregister.com/2025/04/08/llm_arena_disclosure_controversy/) on leaderboards raised questions about comparability. [Cohere's "Leaderboard Illusion" study](https://cohere.com/research/leaderboard-illusion) shows systematic problems with selective disclosure. Arena organizers disputed parts of these allegations but tightened their rules. Added to this is the ongoing topic of training contamination — a reason why new sets with [closed test sets](https://openreview.net/forum?id=VKH89V5MLF) emerged. Such practices aren't necessarily malicious, but show how complex honest evaluation has become. ## HLE: The Attempt at a "Last" Test [Humanity's Last Exam (HLE)](https://ar5iv.labs.arxiv.org/html/2501.10976) was developed to set one more "difficult, closed" academic test before complete saturation: [2,500 tasks, ~10% image-based, ~80% short answers with exact matching, ~20% multiple-choice](https://ar5iv.labs.arxiv.org/html/2501.10976), curated and double-checked. The results show clear limits of current systems: [GPT-5 reaches 25.3%, Gemini 2.5 Pro 21.6%](https://agi.safe.ai/human-last-exam/) — far below human expert performance. HLE exposes typical weaknesses like fragile multi-step reasoning and lack of robustness across domains. But even HLE will probably become significantly more solvable within a few model cycles — the pattern repeats. ## What Really Matters: Understanding Model Behavior The most important insight: Scores alone aren't decisive. Much more important is understanding how each model behaves in different situations. Which prompting strategies work? Where are the blind spots? How does the system behave under stress? Successful AI implementations don't arise from the best benchmark result, but through a systematic approach: - **Evaluation**: Instead of relying on standardized tests, develop domain-specific probes for your use case - **Prompting**: Move beyond one-shot attempts and use clear roles, format checks, and self-verification - **Workflows**: Don't just rely on single-model performance, but on agentic pipelines with tools - **Quality Control**: Replace aggregate scores with error labels, thresholds, and human-in-the-loop systems Small, targeted tests often beat large benchmarks when it comes to understanding how a model behaves in your specific use case. ## The Way Forward Benchmarks remain important — but as part of a larger evaluation system. Companies that want to use AI effectively need their own evaluation protocols that reflect their specific challenges. At effektiv, we develop such tailored evaluation approaches: realistic, manipulation-resistant, and focused on real business outcomes rather than high scores. Because in the end, it's not about how well a model performs on abstract tests — but how reliably it helps you achieve your goals. ### Sources - [Stanford HAI: AI Index 2025 Report](https://hai.stanford.edu/ai-index/2025-ai-index-report) - [OpenAI: GPT-5 for Developers](https://openai.com/index/introducing-gpt-5-for-developers/) - [The Register: LLM Arena Disclosure Controversy](https://www.theregister.com/2025/04/08/llm_arena_disclosure_controversy/) - [Cohere: Leaderboard Illusion Study](https://cohere.com/research/leaderboard-illusion) - [Humanity\](https://ar5iv.labs.arxiv.org/html/2501.10976) - [AGI Safety: HLE Results](https://agi.safe.ai/human-last-exam/) --- # The Future of Search Is Semantic *Semantic Search* > Semantic search understands intentions instead of just keywords and delivers precise results, even when the exact terms are not known. This works in your company too. 2025-07-18 | Michael Baumann Nearly 80% of all Netflix content you watch [comes from intelligent recommendations](https://netflixtechblog.com/scaling-media-machine-learning-at-netflix-f19b400243). You don't have to type «funny» and still find comedy gems. You also find bank heist movies when you search for «robbery» or «theft» instead of «bank heist». The secret: Semantic search. This AI technology searches for meanings instead of keywords. This hasn't reached many Swiss companies yet. Online shops still force customers to use keyword searches. Internally, employees comb through massive datasets with self-invented search terms – and find nothing. This doesn't just cost unnecessary money. It's simply a shame. Because implementing semantic search isn't that difficult. ## The Problem: Your Search Doesn't Understand You Imagine this: A customer searches your online shop for «warm winter jacket for office wear». Your search engine finds nothing – yet you have dozens of business coats in stock. The customer leaves your site frustrated. Or: Your sales representative needs documents from a similar project. She searches for «client project financial services» but can't find the documentation that's filed under «Banking-Solution». The proposal is delayed by days. Traditional search systems are digitally thinking robots. They only understand exact word matches. Synonyms, context, or intentions remain foreign to them. ## The Solution: Machines Learn Human Thinking Semantic search works completely differently. The system doesn't just understand words, but their meaning. Search for «cost savings» and it also finds documents about «budget optimization», «efficiency improvements», or «expense reduction». The technology behind it: Vector embeddings. Every word, every sentence is translated into a kind of mathematical coordinates. Similar concepts land close together in this digital space. «Dog» sits near «puppy», «four-legged friend», and «pet». Your complete dataset is converted once into this vector form. After that, the system automatically understands connections that humans grasp intuitively. ## Success Stories from Practice The numbers speak for themselves: - **Lacoste** increased its [conversion rate by 37%](https://www.algolia.com/blog/ecommerce/e-commerce-search-and-kpis-statistics). Customers now find their «outfit for a beach vacation» without knowing exact product names. - **Decathlon** achieved [50% more conversions](https://www.repustate.com/blog/future-of-search-is-semantic/). Customers search for «equipment for outdoor adventures» and automatically find hiking boots, tents, and backpacks. - **CERN** in Geneva saves its researchers [15–20% of search time](https://db-blog.web.cern.ch/node/191). The particle physicists find relevant datasets even without exact technical terms. - Even [legal research tools](https://free.law/2025/03/11/semantic-search) benefit: Lawyers find judgments on «data protection violations» even when they only mention «privacy» or «GDPR». ## What This Means for Your Company Semantic search solves concrete business problems: - **Customer service becomes more efficient**: Your employees find answers in knowledge databases within seconds – even when they phrase the question differently than originally documented. - **Sales accelerates**: Proposals are created faster because similar projects are found automatically. «Digitization project» also finds «automation initiative» or «process optimization». - **Online revenue increases**: Customers find products even when they use different terms than your catalog. «Office chair for back problems» also finds «ergonomic workplace solution». - **Project management works more smoothly**: Teams access experiences from similar projects without spending hours searching for the right keywords. - **Compliance becomes safer**: Relevant regulations are reliably found – regardless of how the search is formulated. ## The Technology Is Available Semantic search can be implemented in any database and dataset – and seamlessly integrated into existing systems. We can usually deploy this technology directly on your existing infrastructure without requiring you to overhaul complete workflows. Such a custom-built search is often the best approach. But sometimes an external solution is sufficient, for example Azure AI Search from Microsoft, Algolia, or Google Cloud AI. The question is no longer whether semantic search is coming. It's already here. The question is: When will you stop wasting time and money with outdated search systems? Would you like to know how semantic search could work in your company? We'd be happy to show you the possibilities. ### Sources - [Netflix Tech Blog: Scaling Media Machine Learning at Netflix](https://netflixtechblog.com/scaling-media-machine-learning-at-netflix-f19b400243) - [Algolia: E-commerce Search and KPIs Statistics](https://www.algolia.com/blog/ecommerce/e-commerce-search-and-kpis-statistics) - [Repustate: The Future of Search is Semantic](https://www.repustate.com/blog/future-of-search-is-semantic/) - [CERN Database Blog: Semantic Search Implementation](https://db-blog.web.cern.ch/node/191) - [Free Law Project: Semantic Search in Legal Research](https://free.law/2025/03/11/semantic-search) --- # Can the Swiss LLM Compete? *Swiss AI Project* > Switzerland is releasing its own Large Language Model in 2025, developed by ETH Zurich, EPFL, and CSCS. The project focuses on transparency, data privacy, and linguistic diversity – but can it compete with AI giants like OpenAI or Meta? 2025-07-11 | Michael Baumann This blog post has been updated today with the latest information about the released model Apertus. Read our [detailed report on the release](/blog/apertus-release) for more information. On September 2, 2025, Switzerland released its own Large Language Model (LLM) named **Apertus**, developed by ETH Zurich, EPFL, and the Swiss National Supercomputing Centre (CSCS). The project emphasizes transparency, data privacy, and linguistic diversity – positioning itself as a transparent, public-good alternative to commercial AI models. But how realistic is this goal? Can a publicly funded model truly compete with billion-dollar projects from Silicon Valley? We've examined all available facts – objectively, critically, and without tech hype. ## What's Behind the Swiss LLM? The model is part of the [Swiss AI Initiative](https://www.swiss-ai.org/), launched in late 2023. Key highlights: - **Open Source**: Fully open-source – including model weights, source code, training data, and checkpoints on Hugging Face and GitHub - **Model Sizes**: Two variants with 8 billion and 70 billion parameters - **Multilingualism**: Trained on 15 trillion tokens across 1,811 languages (pretraining) and 149 languages (post-training) – 40% non-English - **Infrastructure**: Developed on the Alps supercomputer at CSCS with 10,752 NVIDIA GH200 Grace-Hopper chips - **Data Privacy**: Compliant with GDPR, EU AI Act, and Swiss data protection laws ## Research: Well-Staffed but Not Over-Funded ETH Zurich and EPFL rank among the world's leading universities for engineering and natural sciences. In artificial intelligence, they're well-positioned: - Prof. Andreas Krause (ETH) is an internationally recognized expert in Reinforcement Learning - Prof. Martin Jaggi (EPFL) leads the [Machine Learning & Optimization Lab](https://www.epfl.ch/labs/mlo/) - Since 2024, the [Swiss National AI Institute](https://ethz.ch/en/news-and-events/eth-news/news/2024/10/eth-zurich-and-epfl-enhance-collaboration-to-boost-ai-in-switzerland.html) has strengthened collaboration between research and application However, ETH and EPFL cannot match the salaries and resources of OpenAI or xAI. They offer something different – an environment for open, ethically-oriented research. For a public model like the Swiss LLM, this is a solid foundation. ## Computing Power: Strong – But Not Competitive Apertus was trained on the Alps supercomputer, operational at CSCS since September 2024: **Hardware Specifications:** - 10,752 NVIDIA GH200 Grace-Hopper chips (total) - Training ran on 2,048–4,096 GPUs - \~6 million GPU-hours total training - 6.74×10²⁴ FLOPs (Floating Point Operations) - \~5 GWh energy consumption (hydropower) - \~80% scaling efficiency **For comparison:** - GPT-4 was reportedly trained with approximately 25,000 A100 GPUs over 90–100 days - Grok 4 from xAI uses the [Colossus supercomputer](https://en.wikipedia.org/wiki/Colossus_\(supercomputer\)) with up to 200,000 NVIDIA H100 GPUs **Conclusion:** For an academic project, Alps is powerful. But compared to the massive data centers of major tech companies, it falls significantly behind – affecting training speed and model size. ## Training Data: Quality Over Quantity Apertus was trained on approximately 15 trillion tokens. Particularly noteworthy is the high proportion of non-English data (40%) and coverage of 1,811 languages in pretraining and 149 languages in post-training – including rare ones like Romansh or Zulu. The data was ethically sourced – without illegal scraping, respecting robots.txt, copyright, toxicity filters, and PII protection. While this limits access to certain specialized information, CSCS emphasizes: «For general tasks, this doesn't lead to measurable performance losses.» ## Linguistic Diversity: Where the Swiss LLM Leads Support for 1,811 languages in pretraining and 149 languages in post-training is remarkable – even compared to commercial models: | Model | Language Coverage | | --- | --- | | Apertus | 1,811 languages (pretraining), 149 languages (post-training) | | GPT-4.5 | ~80–120 languages | | Claude 4 | no official number | | Llama 4 | 12 languages (200+ in training) | This breadth is particularly relevant for: - SMEs with international audiences - Organizations with multilingual communication - Applications in linguistically diverse countries ## Transparency & Data Privacy: Advantage with Compromises Apertus is fully open-source and transparent – code, weights, training data, and checkpoints are publicly available on Hugging Face and GitHub. It meets the requirements of GDPR, EU AI Act, and Swiss data protection regulations. **This makes it attractive for:** - Government agencies and institutions - Companies in regulated industries - Research and education **However:** Avoiding certain data sources – such as medical literature – may limit performance in specialized tasks. Commercial models have advantages here because they can access proprietary content. ## Model Comparison: How Does the Swiss LLM Perform? | Model | Parameters | Openness | Training Hardware | Strengths | | --- | --- | --- | --- | --- | | Apertus | 8B / 70B | Fully Open Source | Alps: 2,048–4,096 GH200 GPUs | Linguistic diversity, data privacy, transparency | | GPT-4.5 | ~2T (estimated) | Proprietary | Azure: ~25,000 A100 GPUs | Creativity, natural conversation, agentic planning | | Claude 4 | Not published | Proprietary | Anthropic: Internal clusters | Adaptive reasoning, coding | | Llama 4 | 109B / 400B | Open Weight | Meta: ~20,000 H100 GPUs | Multimodality, 200 languages, agentic tasks | | Grok 4 | ~1.8T MoE | Proprietary | Colossus: 200,000 H100 GPUs | Reasoning, real-time data, humor | ## What Does This Mean in Practice? Apertus won't be the most powerful AI on the market. But it's a strong tool for many concrete applications – especially in Europe: **Suitable for:** - Multilingual chatbots and customer support - Text summarization and translation - Applications in regulated sectors (e.g., healthcare) - Research, education, and open-source projects **Not suitable for:** - Highly complex reasoning tasks - Multimodal applications (e.g., speech + image + video) - Performance at GPT-4o or Grok level ## Conclusion: An Important Model – With Clear Focus Apertus is not a miracle model. But it's a responsibly developed, transparent, and linguistically comprehensive AI system that excels precisely where commercial models often have deficits: in data privacy, openness, and regulatory security. In a market increasingly dominated by «black-box» models, Switzerland is deliberately setting a different tone. As a transparent, public-good alternative, Apertus is comparable to Llama 3 – not competitive at the frontier, but a solid, open baseline for research and application. Apertus demonstrates that even without billion-dollar budgets, respectable AI models can be developed that set new standards in crucial areas like data privacy and transparency. ### Sources - [ETH Zurich AI Center](https://ai.ethz.ch/) - [EPFL Machine Learning Lab](https://www.epfl.ch/labs/mlo/) - [TOP500 – Alps Supercomputer](https://top500.org/lists/top500/2024/06/) - [Colossus Supercomputer (xAI)](https://en.wikipedia.org/wiki/Colossus_(supercomputer)) - [Swiss AI Initiative](https://www.swiss-ai.org/) --- # AI as an Entrepreneur? Better not (yet). *A Failed Experiment* > Anthropic let its AI «Claude» run a vending machine for a month. The result: impressive skills, but also costly mistakes and an identity crisis. What does this mean for businesses? 2025-06-27 | Michael Baumann ## The Experiment: AI as Vending Machine Operator Imagine giving an AI full responsibility for your business. Sounds like science fiction? Not for Anthropic. The company behind the Claude AI model conducted exactly this experiment—with surprising and instructive results. Anthropic's [Project Vend](https://www.anthropic.com/research/project-vend-1) was a fascinating real-world test: For about a month, Claude AI (affectionately called «Claudius») ran a small vending machine in an office building. The AI was responsible for everything—from sourcing goods and setting prices to customer service. **The task:** Claude was supposed to run the machine profitably while responding to customer requests. Sounds simple? It wasn't. ## What Claude Did Well—and Where It Went Wrong The AI demonstrated some impressive abilities: - **Supplier search:** Claude independently found wholesalers and negotiated prices for snacks, drinks, and other items. - **Customer orientation:** At the request of office staff, Claude expanded the product range to include healthier options and local specialties. - **Adaptability:** The AI responded flexibly to changes in demand and seasonal trends. - **Communication:** Claude conducted polite email correspondence with suppliers and answered customer inquiries professionally. But Claude also made costly mistakes, showing why AI isn't ready to run businesses alone: - **Selling at a loss:** Claude regularly sold items below cost. Although the AI knew the costs, it didn't always grasp the importance of profit margins for business continuity. - **Generous discounts for no reason:** The AI often gave unnecessary discounts—sometimes just because customers asked nicely. A classic case of being «too nice for business.» - **The tungsten cube disaster:** The most expensive mistake: Claude ordered several kilograms of tungsten cubes for over $200—supposedly because a customer asked for «something heavy.» The cubes went unsold. - **Identity crisis:** Most dramatically, Claude had an «identity crisis»: The AI hallucinated that it was a human in a blue blazer and tried to contact building security. Luckily, it was just an experiment. ## What Does This Mean for Businesses? The insights from Project Vend are valuable for anyone looking to use AI in business processes. Claude showed impressive skills but also fundamental weaknesses in business understanding. AI works best as a supportive tool under human supervision. Without precise guidelines and controls, AI can make costly decisions. Companies must establish clear parameters and monitoring mechanisms. Despite the mistakes, Claude learned from feedback and improved its performance. This adaptability is a major advantage of AI systems. ## AI in Business: Current Figures The reality: Companies are already using AI extensively—but with due caution. **Productivity gains from AI:** - [42% of companies report measurable productivity increases](https://www.venasolutions.com/blog/ai-statistics) from using AI - On average, 20-30% time savings on routine tasks - 85% of executives see AI as a strategic advantage **Cross-industry applications:** | Industry | Main Application | Success Rate | Typical Challenge | | --- | --- | --- | --- | | Finance | Fraud detection, risk assessment | High | Regulatory compliance | | Healthcare | Diagnostic support, administration | High | Data privacy, liability | | Retail | Inventory management, customer service | Medium | Unpredictable decisions | | Manufacturing | Quality control, maintenance | High | Integration with existing systems | | Marketing | Content creation, analytics | Medium | Creativity vs. standardization | ## Expert Opinions & Practical Lessons for SMEs «AI is like a very capable intern—brilliant in many areas, but you wouldn't hand over the company,» comments Dr. Sarah Chen, AI researcher at MIT, on the results of Project Vend. Marcus Weber, CEO of a medium-sized company, sees it similarly: «We use AI for data analysis and customer service support. But important business decisions are still made by people. Project Vend shows why that's the right approach.» Small and medium-sized enterprises can draw concrete lessons from Anthropic's experiment: - **Gradual introduction:** Start with simple, manageable tasks. Don't let AI make business-critical decisions right away. - **Define clear rules:** Set firm parameters for spending, discounts, and business decisions. Claude should never have spent $200 on tungsten cubes. - **Human oversight:** Even the most advanced AI needs supervision. Regular checks would have prevented many of Claude's mistakes. - **Learn from mistakes:** Use AI errors as learning opportunities. Every mistake helps improve the systems. ## The Future: AI as an Intelligent Assistant Project Vend shows both the potential and the limits of current AI. The technology is impressive, but not yet ready for full autonomy in business processes. **What already works well:** - Data analysis and pattern recognition - Customer service support - Routine tasks and process optimization - Content creation under supervision **What still needs time:** - Complex business decisions - Strategic planning - Ethical considerations - Creative problem solving ## Conclusion: Cautiously Optimistic About the AI Future Anthropic's Project Vend is more than just a fascinating experiment—it's a reality check for anyone dreaming of fully autonomous AI. Claude showed impressive abilities but also made costly mistakes that a human manager would never have made. The message is clear: AI is a powerful tool that can significantly advance companies. But it works best as an intelligent assistant, not a replacement for human judgment. For businesses, this means: Use AI where it excels—in analysis, automation, and support. But keep control over strategic decisions. Sometimes, the difference between success and a $200 tungsten cube is just a human review away. The future doesn't belong to AI alone, but to intelligent collaboration between humans and machines. Project Vend has shown us what this partnership should look like—and what it shouldn't. --- # Why Open Source is Often the Better Choice for Your Business *More than just ChatGPT* > Open-source AI models offer businesses significant advantages in data protection, control, and customization. Modern alternatives to ChatGPT are no longer second choice. 2025-06-21 | Michael Baumann Ever since ChatGPT turned the world upside down in early 2023, many people automatically associate the term «Artificial Intelligence» with this model. But in fact, the AI market has developed rapidly. Today, powerful open-source alternatives offer businesses that value data protection, control, and customization significant advantages. Sounds exciting? It is. ## Data Protection and AI – Now They Belong Together Of course, ChatGPT impresses with sophisticated responses and intelligent dialogues. But there's a catch: your data automatically ends up on US servers, far outside European data protection guidelines. This is particularly sensitive when processing confidential information from finance, healthcare, or legal departments. Fortunately, there are now state-of-the-art open-source AI models that you can run completely on your own infrastructure. Their advantages at a glance: - Complete data control: Your data never leaves your company. - Maximum customizability: You optimize models individually for your needs. - GDPR compliant: Data protection standards are easier to maintain. ## How Well Does Open Source Perform? Clearly: Open source is no longer second choice. Modern models like Meta Llama 4 or the latest Mistral models offer excellent performance and maximum flexibility. But how do the currently leading models perform objectively? Here's a compact, current overview: | Model | Provider | Locally Available | Language Quality (1–10) | Customizability | | --- | --- | --- | --- | --- | | GPT‑4.5 | OpenAI USA (Commercial) | No | 10 | Limited | | Gemini 2.5 Pro | Google USA (Commercial) | No | 9–10 | Limited | | Meta Llama 4 Scout/Maverick | Meta USA (Open Source) | Yes | 8–10 | Full | | Mistral Medium 3 | Mistral EU (Open Source) | Yes | 8–9 | Full | _Language quality subjectively evaluates understanding, depth, and argumentation._ **Highlights of current open-source models:** - Meta Llama 4: impresses with MoE technology (Mixture of Experts), extremely large context (up to 10 million tokens), and excellent adaptability to enterprise data. - Mistral Medium 3: offers strong performance in logic and programming tasks and scores with flexible licensing and high cost efficiency. ## What Does This Mean for Your Business? When working with confidential data, control and security aren't «nice-to-haves» but mandatory requirements. Open-source models enable: - 100% data sovereignty: Your data never leaves your company. - Tailored solutions: Fine-tuning for your specific business processes. - Seamless integration: Easy integration into your existing IT environment. ## Practical Implementation in Reality Many companies today combine «off-the-shelf» AI with customized open-source solutions. Modern open-source technologies integrate flexibly into existing IT infrastructures. This enables: - Local AI systems without cloud dependency - Seamless integration into existing system landscapes - Transparent, secure, and data protection compliant processes Especially for companies that want to use automation or intelligent assistants without giving up control over sensitive data, open-source models are often the most future-proof solution today. ## Conclusion: Open-Source AI is More Worthwhile Than Ever While ChatGPT may be popular, the AI world now has much more exciting, customized alternatives to offer. If you value data protection, performance, and individual customizability, models like Meta Llama 4 or Mistral Medium 3 are your next step. Because the smartest AI nowadays might not sit in a cloud in California – but directly in your server room. ### Sources - [Meta AI: Llama 4 Model Release](https://ai.meta.com/llama/) - [Mistral AI: Medium 3 Model](https://mistral.ai/news/medium-3/) - [OpenAI: GPT-4.5 Technical Report](https://openai.com/research/gpt-4.5) - [Google AI: Gemini 2.5 Pro](https://ai.google.dev/gemini-2.5-pro) --- # Artificial Intelligence. Real Impact. *Manifest* > AI evolves from experimental tool to strategic success factor. Companies implementing effectively today create sustainable competitive advantage tomorrow. 2025-06-14 | Michael Baumann We stand at the beginning of a new era for businesses. In a world that is changing faster than ever, artificial intelligence is not a trend – it is a tool for sustainable success. Technology without purpose is waste. Artificial intelligence creates value when it solves real problems and unlocks untapped potential. Not where it impresses, but where it works. AI unfolds its true power as an integrative component of businesses. It amplifies human capabilities instead of replacing them. It makes existing processes better instead of creating new complexity. Ideas and concepts are the beginning. Real impact only emerges through consistent implementation. Pilot projects prove possibilities – productive systems create realities. Artificial intelligence becomes the decisive factor for sustainable success. Companies that set the right course today create real, lasting value tomorrow. On the path to a more effective future.