Ensuring cost efficiency with enterprise-scale deployment of AI agents is a tough challenge for every organization. Did you know you already possess all the infrastructure right at your fingertips? Think distributed, managed bare-metal infrastructure offering 24/7 operational AI agents for under 100 EUR/month.

In the age of Cloud, k8s and VMs, an overlook is actual machines your company owns. Full governance and control at an unbeatable price for every business, regardless of their size.

Artificial intelligence promises unprecedented efficiency, yet many organizations still struggle to demonstrate meaningful business outcomes beyond faster content generation and lower costs per task. Too often, success is measured in tokens processed, prompts executed, or hours seemingly saved—metrics that can obscure whether AI is actually improving decisions, customer experiences, or employee effectiveness.

This talk challenges the assumption that efficiency alone is the goal. Instead, it presents a practical framework for designing AI pilots that are measurable, outcome-driven, and intentionally human-centered. Drawing on real-world implementation experience, the session explores how to define meaningful success metrics, identify where human judgment adds the greatest value, and avoid optimizing systems for activity instead of impact.

Many organizations have established AI principles, policies and governance structures. Yet employees often still struggle with practical questions: Which tools can I use? What data may I enter? Who is accountable for the result? And how can I innovate without creating unnecessary risk?

The challenge is no longer simply to define responsible AI; it is to translate governance into everyday decisions and scalable organizational practice.

This talk explores how companies can move from abstract principles to actionable guardrails, practical enablement and shared accountability. Drawing on experience from AI governance, compliance and large-scale employee enablement, Sindy Leffler-Krebs will show why governance should not be designed as a control layer added after innovation, but as an integral part of adoption.

Key takeaways:

• How to translate AI principles into clear and usable employee guidance

• Why AI literacy, leadership and culture are essential governance mechanisms

• How communities and practical learning formats accelerate responsible adoption

• How to balance innovation, accountability and human oversight

Participants will leave with a practical framework for making responsible AI understandable, actionable and scalable.

GenAI prototypes can look impressive long before they are reliable enough for real users. The difficult question is not simply whether the model works, but whether the overall product is actually ready to ship.

Drawing on lessons from building and evaluating production GenAI systems at Meta, I’ll share a practical approach to answering that question. We’ll look at how teams can combine offline test sets, human review, automated and LLM-based evaluation, live dogfooding, prompt comparisons, and production metrics—and why relying on any one of these signals can give a misleading picture of quality.

I’ll also cover what changes after launch: detecting regressions, evaluating multilingual experiences, separating model failures from product failures, and catching issues that only emerge at scale.

Attendees will leave with a framework for designing evaluation alongside the product, defining what “good enough to ship” actually means, and turning evals from an offline scorecard into an engineering system that supports real production decisions.

GenAI keeps getting more capable, yet its value has not reached the people who need it most. They don’t ask for AI. Often they don’t even ask for help. Jan Gerwin of Hybrid Heroes shows how generative AI can work invisibly in the backend, using the example of Finlyfree, an app Hybrid Heroes developed for Finlyfree GmbH that helps people in debt make sense of their mail. A photo of a crumpled letter becomes structured data, explained in plain language and ready to hand over to a debt counselling service. The talk distills three product principles for bringing AI value to people who never asked for AI.

How we rebuilt Softr to make AI-powered software development accessible without compromising security, governance, or control.

Most teams treat the model as the system. In production, the model is the easy part, the scaffolding around it is what determines whether your AI system is dependable, cost-effective, and correct. This session skips the introductions and goes straight into the hard practical problems: how to manage context windows under real-world constraints, when to retrieve versus when to keep things in memory, how compaction and caching decisions trade off latency against fidelity, and how prompt/instruction design shifts once you're optimizing for a system that runs thousands of times a day rather than a single chat interaction. I'll share concrete techniques and failure modes from building harnesses that turn a raw model into something you can actually trust in production; including where naive context management breaks down, how to think about memory as a design surface rather than an afterthought, and what "good" instruction design looks like once you're past the demo stage.

In this session we share how we designed and implemented a pipeline where AI agents collaborate across the entire data onboarding lifecycle, from feed analysis to mapping code generation. Rather than replacing expertise, the solution embeds human knowledge into a scalable framework where specialists guide, validate and continuously improve agent outcome. The result is faster onboarding, improved quality, reduced manual effort and the ability to scale. Attendees will gain insights into building agentic workflows while keeping the human in the lead and turning AI from a productivity tool into a strategic capability that transforms how work gets done.

AI agents often look impressive in demos. The difficulty starts when questions arise pre-release. Is our AI agent ready for release with all requirements met? Are we able to identify unwanted behaviors once live? Will running costs spiral once traffic scales up? Most AI teams struggle to answer these questions like these with certainty.

In this talk, we will examine the three structural pillars required to move an agent from pilot to production. We will cover how teams can build verification that satisfies internal stakeholders and external regulators, alongside methods for catching live failures early. We will also explore how experimentation helps you find the right configurations to keep costs bounded.

Key takeaways:

• What evidence internal teams and regulators require before giving a release green light

• How to identify performance issues in live environments

• How to run experiments that uncover cost-efficient hyperparameter setups

For decades, businesses have invested in software that requires people to adapt to it. Employees navigate applications, learn interfaces, and work around predefined processes to get things done.

But what if advances in AI could fundamentally reverse that relationship?

What if employees could simply express what they want to achieve, and software dynamically assembled the interfaces, workflows and services needed to deliver it?

The hypothesis behind Post-Application Architecture (PAA) is that this shift could fundamentally change the economics of enterprise software — reducing operational friction, lowering costs and enabling businesses to adapt faster.

PAA explores this possibility through a different approach to software design:

Agents decide. Services execute. Policies constrain.

By separating intelligent decision-making, deterministic execution and governance into distinct architectural layers, PAA challenges the assumption that applications must remain the primary building blocks of enterprise software.

This talk introduces PAA as an architectural hypothesis, not a finished solution. Through a practical example, we'll explore how intent-driven software might work, what it could mean for business efficiency and agility, and the challenges that must be addressed before this vision can become reality.

Key takeaways:

• Why application-centric software may be limiting business efficiency and agility.

• How intent-driven systems could fundamentally change the way organisations interact with technology.

• What technical, operational and governance challenges must be overcome to make this approach viable.

AI does not replace humans. It replaces friction. Sovereignty belongs to those who help shape this economy.

AI does not replace humans. It replaces friction. Sovereignty belongs to those who help shape this economy.

In transcatheter edge-to-edge repair (TEER), a leaking mitral valve is repaired through a catheter while the operator watches a live transoesophageal echo image. Every decision, where to place the clip, when to grasp the leaflets, whether to release, is made by eye, and there is not a single number on the screen. When we gave the same frame to four specialists they disagreed by two to four millimetres. That gap is why the learning curve is long, why outcomes vary between operators, and why many hospitals cannot start a programme.

AutoClip is a segmentation model that maps the valve and the clip on every ultrasound frame and turns each decision into a measurement: clip position on the papillary-muscle axis, insertion depth of each leaflet, and grasped leaflet length. The operator still decides; AutoClip gives a number where there was a guess. It is built as a device-agnostic layer that works with any clip and any echo vendor, and its intended use is to improve procedural outcomes: procedure time, clip malposition and re-intervention.

AutoClip was built by a cardiac surgeon with generative-AI coding tools, on a multi-centre dataset of about 2,000 procedures from four hospitals in Europe and China, with an outcome-linked registry so that every readout can be tested against results. It was selected as a Featured Innovation at TCT 2025 while still a prototype. Turicor is now building the real-time version and preparing clinical validation.

Every organisation scaling AI makes the same promise to its board, its customers and its regulators: a human remains accountable. This talk asks what that promise rests on.

I propose a standard: accountability is real when someone with authority can detect a failing system, escalate, decide and intervene before the harm becomes irreversible. Without evidence, accountability becomes risk.

Drawing on two decades in regulated financial services and my work as independent adviser to Intervene, I show how to test that standard on one system, and why the organisations that can evidence it are the ones that can scale.

Every day you re-brief your AI assistant on things you told it yesterday. You carry context from Claude to ChatGPT to the next tool, and you have quietly become the integration layer.

Searching old chats works for facts. It breaks down when the answer is context: why something was decided, what it touches and what else belongs to it. That answer rarely looks like the question. Memory built on connections alone would lose at facts, so Komplyt Zero keeps similarity search and adds connections and time on top.

We compare chat history, vector search and Zero on size, team use, sources and the kind of question each can answer, using cases from our own work. Zero is free, runs inside Claude and ChatGPT, and its engine is going open source under Apache-2.0.

Key takeaways:

• Search finds what looks like your question. Memory also has to find what belongs with it.

• What you capture, and when, decides what you can retrieve. Capture inside the assistant, where the context is.

• You can set up a shared memory for yourself or your team in a minute.

On 6 October, Meta and Sierra announced the Personal Agent Protocol (PAP), with partners including Shopify, Stripe and Walmart. PAP is an open standard for how a customer's personal AI agent finds a business, signs in with the customer's permission and gets things done. It gives companies a front door for agents, instead of bots scraping their websites.

This 10-minute talk explains, without code, what PAP promises and where it fits alongside MCP, APIs and WebMCP. It then covers what PAP changes for brands, retailers and enterprises: who owns the customer relationship, what an agent should be allowed to do, and how to see agent activity. It also names what is still missing.

Key takeaways: (1) Agents will come in through one of three doors: your website, your APIs or your own company agent. Choosing one is a business decision. (2) The value is consent and visibility. The customer decides what the agent can access, the company sets its limits, and you finally see which agent is acting for whom. (3) Plan around the gaps. The spec is not yet published, detailed limits and payments come later, and rival standards exist.

Nobody in Germany doubts that AI is powerful anymore. What still fails is implementation: pilots built on clean demos collapse when they meet real data, real approvals and real silos. Companies are not blank canvases; they run on ERPs, CRMs, spreadsheets and habits built over years.

This talk shares Output-First Discovery, the approach Adriana Carmona Beltran uses with teams such as idealo: start with a business output the team already produces, trace the workflow backwards, map the tools and manual steps, then reproduce it with AI inside the systems that already exist. At idealo, a workflow spanning six tools and three days now runs in under an hour, with the same tools and no rip-and-replace.

Key takeaways: (1) Don't start with tools or wishes; start with work that already happens. (2) The output reveals the workflow, and the workflow reveals the automation opportunity. (3) Keep humans in the lead: people decide what "correct" means before anything goes live.

Organizations invest heavily in training, yet creating engaging learning experiences, delivering courses across languages, and scaling training efficiently remain major challenges. Traditional learning platforms often require significant time and resources to develop, manage, and maintain high-quality learning programs.

Lectful is an AI-powered, white-label learning platform that helps training providers, educational institutions, and organizations build and scale their own digital learning environments. By combining learning management technology with generative AI, Lectful aims to simplify course creation, support multilingual learning experiences, and make digital education more accessible and scalable.

Our approach focuses on applying AI to practical learning workflows rather than adding AI for its own sake. From AI-assisted content creation and translation to more personalized learning experiences, Lectful helps organizations reduce operational complexity while creating new opportunities to deliver training at scale.

In this talk, we will present how Lectful applies generative AI to real-world education and workforce development challenges, share our approach to building scalable learning experiences, and explore how AI-powered learning can help organizations expand access to knowledge across languages, sectors, and markets.

The key takeaway: AI can make digital learning more scalable and accessible when it is designed around real organizational needs and measurable outcomes.

Organizations invest heavily in training, yet creating engaging learning experiences, delivering courses across languages, and scaling training efficiently remain major challenges. Traditional learning platforms often require significant time and resources to develop, manage, and maintain high-quality learning programs.

Lectful is an AI-powered, white-label learning platform that helps training providers, educational institutions, and organizations build and scale their own digital learning environments. By combining learning management technology with generative AI, Lectful aims to simplify course creation, support multilingual learning experiences, and make digital education more accessible and scalable.

Our approach focuses on applying AI to practical learning workflows rather than adding AI for its own sake. From AI-assisted content creation and translation to more personalized learning experiences, Lectful helps organizations reduce operational complexity while creating new opportunities to deliver training at scale.

In this talk, we will present how Lectful applies generative AI to real-world education and workforce development challenges, share our approach to building scalable learning experiences, and explore how AI-powered learning can help organizations expand access to knowledge across languages, sectors, and markets.

The key takeaway: AI can make digital learning more scalable and accessible when it is designed around real organizational needs and measurable outcomes.

The proliferation of agentic AI systems across enterprise and public-sector contexts has outpaced the capacity of general-purpose AI risk frameworks to classify and govern them. To solve this gap, the Responsible AI Institute introduces the TrustX Agent Risk Classification (ARC) Framework, a structured, repeatable instrument that helps enterprises classify internal AI risk and translate governance requirements into measurable controls.

Grounded in internationally recognized standards (EU AI Act, NIST AI RMF, ISO/IEC 42001, etc.), ARC can be applied to seven types of agentic AI systems. It assesses whether the use case is agentic, its autonomy level, and how it scores against twelve risk dimensions derived from existing governance frameworks. The use case is then placed in a risk tier, with generative AI providing personalized controls and policy recommendations.

This framework has supported organizations deploying AI at scale across various industries, ranging from healthcare to financial services. Since its launch, 200+ assessments have been completed on the interactive ARC tool across 12 industries. ARC's methodology has also been published on arXiv, and two working groups within the Responsible AI Institute have been created to help integrate ARC-related initiatives into selected industries.

European fintechs face massive operational overhead under German BaFin and GwG compliance rules. Manual onboarding, complex Ultimate Beneficial Owner (UBO) tracing, and risk scoring take up to 45 minutes per case, creating severe growth bottlenecks. In this pitch, we explore the architecture and real-world deployment of a production-grade, hybrid agentic AI compliance auditor. Developed at Cedony Consultancy Services, this solution bridges the gap between the flexible reasoning of Large Language Models and strict, deterministic validation layers. By combining autonomous document parsing with rule-based compliance guardrails and human-in-the-loop oversight, the system operates in full alignment with the EU AI Act's high-risk framework.

Key Takeaways:

i. Zero-Hallucination Design: How to blend probabilistic LLMs with deterministic guardrails to ensure 100% legal citation accuracy in regulated sectors.

ii. Agentic Workflows in Action: Practical strategies for automating unstructured multi-page document ingestion and UBO graph parsing.

iii. Measurable Enterprise Impact: Real-world metrics demonstrating an 82% reduction in review cycle times, a 65% drop in false positives, and fully auditable decision trails.

Most companies don't struggle with GenAI technology – they struggle to find the right use cases. The valuable ones sit in the daily work of hundreds of employees: in handovers, workarounds and routines that management never sees. Workshops reach a handful of people, consultancies a few more. Nobody talks to everyone.

Buddy, the AI assistant inside the MeinGPT platform, now can. It holds confidential voice conversations with employees about their work, asks follow-up questions and captures where AI can genuinely help. Because no manager reads along, people talk openly. Buddy works in rounds: it prepares follow-up interviews, suggests whom to talk to next and builds a knowledge graph in the background that connects what no single person could see. Smaller use cases it builds directly on the platform.

Just as important is what Buddy doesn't do. Trust, priorities and rollout stay with people: a consultant briefs Buddy beforehand, leadership decides afterwards, and teams carry the change.

In this talk, we share what we learned from our first customer sprints and an internal test, where participants rated the conversation 4.3 out of 5 for being on equal footing.

Key takeaways:

• Why use-case discovery, not technology, is the real bottleneck of GenAI adoption

• How confidential AI interviews surface use cases that workshops miss

• Where to draw the line between AI and humans in an adoption process

• How a company can start listening to itself – continuously, not once a year

Most companies don't struggle with GenAI technology – they struggle to find the right use cases. The valuable ones sit in the daily work of hundreds of employees: in handovers, workarounds and routines that management never sees. Workshops reach a handful of people, consultancies a few more. Nobody talks to everyone.

Buddy, the AI assistant inside the MeinGPT platform, now can. It holds confidential voice conversations with employees about their work, asks follow-up questions and captures where AI can genuinely help. Because no manager reads along, people talk openly. Buddy works in rounds: it prepares follow-up interviews, suggests whom to talk to next and builds a knowledge graph in the background that connects what no single person could see. Smaller use cases it builds directly on the platform.

Just as important is what Buddy doesn't do. Trust, priorities and rollout stay with people: a consultant briefs Buddy beforehand, leadership decides afterwards, and teams carry the change.

In this talk, we share what we learned from our first customer sprints and an internal test, where participants rated the conversation 4.3 out of 5 for being on equal footing.

Key takeaways:

• Why use-case discovery, not technology, is the real bottleneck of GenAI adoption

• How confidential AI interviews surface use cases that workshops miss

• Where to draw the line between AI and humans in an adoption process

• How a company can start listening to itself – continuously, not once a year

Most companies don't struggle with GenAI technology – they struggle to find the right use cases. The valuable ones sit in the daily work of hundreds of employees: in handovers, workarounds and routines that management never sees. Workshops reach a handful of people, consultancies a few more. Nobody talks to everyone.

Buddy, the AI assistant inside the MeinGPT platform, now can. It holds confidential voice conversations with employees about their work, asks follow-up questions and captures where AI can genuinely help. Because no manager reads along, people talk openly. Buddy works in rounds: it prepares follow-up interviews, suggests whom to talk to next and builds a knowledge graph in the background that connects what no single person could see. Smaller use cases it builds directly on the platform.

Just as important is what Buddy doesn't do. Trust, priorities and rollout stay with people: a consultant briefs Buddy beforehand, leadership decides afterwards, and teams carry the change.

In this talk, we share what we learned from our first customer sprints and an internal test, where participants rated the conversation 4.3 out of 5 for being on equal footing.

Key takeaways:

• Why use-case discovery, not technology, is the real bottleneck of GenAI adoption

• How confidential AI interviews surface use cases that workshops miss

• Where to draw the line between AI and humans in an adoption process

• How a company can start listening to itself – continuously, not once a year

What if every matter made a law firm smarter? Every negotiation strengthened its institutional memory. Every expert correction improved how the next task was done. Decades of experience became knowledge the entire firm could put to work.

That is the future we are building toward at Eigenwelt Labs: law firms that own and continually improve their AI, with an open-source foundation developed alongside the legal community.

Our approach connects three capabilities. A legal workspace where lawyers and agents work together. Computer and browser workflows that let agents act across the firm’s existing software. Knowledge graphs that connect matters, documents, people, and decisions into a living institutional memory.

Together, these create the foundation for a larger shift: legal expertise becoming an asset that compounds through everyday work. The community improves the shared tools. Each firm builds its own private intelligence. Lawyers shape the system through their judgment and corrections.

This talk presents that vision, demonstrates the foundations we are building today, and explores how an open ecosystem could give firms of every size the ability to develop AI around their own expertise.

Large companies sometimes break the law. Since 2023, the EU allows class action lawsuits to fix that at scale with. But most people never claim, which means the companies keep the money and face little consequences.

Advocated.eu makes joining a collective claim as simple as it gets. Users see open EU class actions in plain language, check eligibility with a few clicks, and use instruction pre-sets to join lawsuits, and follow the status of their claim in one place. The platform is free and funded by voluntary contributions from users who choose to give.

Key takeaways

• The rights exist, but the legal language and effort stops people from using them.

• Class action lawsuits incentivise companies to better behaviour only when payouts are actually claimed at scale

• Joining a claim is free, risk-free and now also simple.

Whether AI pays off rarely depends on the model. It depends on whether a company understands how its work actually runs. That knowledge is scattered: employees know the workarounds, handovers and waiting times; documents describe how things should work; systems record what happened. Leaders see the result, but not the causes, so it stays unclear where AI or automation actually pays off.

In this talk, Szattalov Valerio, co-founder of insivo, shows how GenAI can listen to an organization at scale, bring these perspectives together and turn operational knowledge into targeted improvements.

Key takeaways:

– Why AI success depends on context, not on the model

– How GenAI makes it possible to listen to an entire organization

– How to find where everyday work loses time and money, and where AI actually pays off

– Why AI should surface evidence while people decide

Buying or selling a home is the biggest deal most people ever make, and in Germany it is still run on PDFs, phone calls and portal fees. Buyers cannot see what a home really costs. Agents spend their day answering the same enquiries by hand.

Athaus is a property marketplace with the agent's software built in. Buyers describe the home they want in plain language. Behind every listing, AI drafts each reply, reads the documents, values the property and books the viewing, and a licensed agent approves before anything goes out. In our first two months with listings, 32,000 property seekers used athaus.ai.

What happens when machines, buildings and assets can explain their condition, request maintenance and coordinate with the agents managing orders, energy and procurement?

This talk explores a possible next chapter of generative AI: industrial and enterprise agents connecting physical assets with business decisions.

Through an illustrative operational scenario, Dr. Omid Givehchi shows how agents representing production equipment, building systems and enterprise functions could balance delivery commitments, asset health and energy costs. He examines what makes this coordination useful: shared operational knowledge, captured engineering experience, explicit decision rights and independent checks before consequential actions.

The session offers a practical vision of how organizations could progress from isolated assistants to coordinated operations across factories, commercial buildings and infrastructure—while retaining human authority and control over their data.

Eric Anderegg, Co-founder of GenAI Zürich, GenAI London, GenAI Berlin and GenAI Shenzhen, opens the GenAI Berlin 2026 conference at Humboldt Carré, steps away from Berlin's historic Unter den Linden boulevard. His opening remarks set the stage for a day of insights, innovation and conversations on applied GenAI.

Nexoria is a digital workspace designed to support GDPR-compliant use of generative AI in everyday business. It combines company knowledge, cloud storage, semantic search and the AI assistant Nex with integrated tools for documents, spreadsheets and presentations. Email and meeting tools are available as additional modules.

Data protection is central to the platform’s architecture. Files, the knowledge index and search infrastructure are hosted in Germany, while Mistral AI processes relevant content within the EU. Workspace permissions control access to company information. A data processing agreement and documented data flows support organisations in assessing and meeting their GDPR obligations.

In this session, co-founders Caan Grueneberg and Christian Klever demonstrate how Nexoria connects information from documents, emails and meetings, provides answers with supporting references and assists with business workflows. An invoice example illustrates how information retrieval, document filing, human approval and transfer to an ERP system can work together.

The presentation also explains how companies can introduce Nexoria alongside Microsoft 365. Business and IT decision-makers will gain a concrete understanding of how to combine generative AI with existing processes while maintaining oversight of access, data processing and automated actions.

Nexoria is a digital workspace designed to support GDPR-compliant use of generative AI in everyday business. It combines company knowledge, cloud storage, semantic search and the AI assistant Nex with integrated tools for documents, spreadsheets and presentations. Email and meeting tools are available as additional modules.

Data protection is central to the platform’s architecture. Files, the knowledge index and search infrastructure are hosted in Germany, while Mistral AI processes relevant content within the EU. Workspace permissions control access to company information. A data processing agreement and documented data flows support organisations in assessing and meeting their GDPR obligations.

In this session, co-founders Caan Grueneberg and Christian Klever demonstrate how Nexoria connects information from documents, emails and meetings, provides answers with supporting references and assists with business workflows. An invoice example illustrates how information retrieval, document filing, human approval and transfer to an ERP system can work together.

The presentation also explains how companies can introduce Nexoria alongside Microsoft 365. Business and IT decision-makers will gain a concrete understanding of how to combine generative AI with existing processes while maintaining oversight of access, data processing and automated actions.