November 25-27

As enterprises rapidly adopt large language models (LLMs), managing inference costs has become a critical challenge. Many AI applications repeatedly transmit large amounts of static or redundant context, resulting in unnecessary token consumption, increased latency, and higher operational expenses. While organizations invest heavily in AI capabilities, few optimize how prompts are constructed and delivered to the model.

In this session, Vivek Singh demonstrates practical techniques for reducing LLM token usage by up to 50% through intelligent prompt preprocessing, semantic context selection, and context caching in Google Vertex AI. Attendees will learn how to eliminate redundant information, preprocess structured and unstructured data, retrieve only the most relevant context, and leverage Vertex AI’s context caching capabilities to avoid repeatedly sending identical information.

Through a live demonstration, the audience will compare a traditional “send everything” approach with an optimized pipeline, examining token consumption, response latency, and API costs in real time. The session also covers implementation patterns, architectural best practices, and measurable strategies that can be integrated into enterprise AI applications without sacrificing response quality.

Attendees will leave with the practical knowledge needed to build faster, more scalable, and significantly more cost-efficient generative AI solutions, making this session valuable for AI engineers, data engineers, cloud architects, platform engineers, and developers deploying LLM-powered applications in production.

When people bring AI into product innovation, they often end up with generic, safe, “average” outputs because AI tends to regress to the mean. In this workshop, Steffen Kastner teaches participants how to use generative AI specifically for problem discovery and ideation without falling into that trap. Participants will bring their own real problem to work on and see firsthand how AI can help them generate genuinely novel ideas—ones that go beyond merely remixing existing patterns from AI training data. Participants will receive exclusive access to the tools used by teams at Digital Product School Munich. These tools have been refined over ten years of work with several hundred product teams.

Everyone is shipping tools and skills for agents, but almost nobody can tell you whether those tools and skills actually help. “It felt better in the demo” is not a metric, and with agents, that impression is often wrong: something can look great in one transcript and quietly fail much of the time in aggregate. This deep dive shows how to put real numbers on the effectiveness and efficiency of agent-facing tools and skills. The order matters: effectiveness first, efficiency second, because reducing costs is worthless if the agent never reaches the goal. Using real evaluation infrastructure as a worked example, Michael Hablich covers how to run controlled experiments with and without the tools or skills, scored against explicit checks; how to read the results to decide where to invest; an honest case in which adding guidance made things worse; and the efficiency levers and their trade-offs that reduce costs once quality is maintained. Attendees will leave able to design an evaluation for their own agent tools instead of trusting a good-looking demo.

Everyone wants to “do something with AI,” but not every business problem needs an AI solution.

In this practical session, Max Golikov shares a decision-ready framework for evaluating which AI projects are worth funding, which should be simplified, and which should be stopped before they become expensive pilots with unclear value.

The talk introduces the AI Value Matrix, a simple tool for assessing AI opportunities based on business value, technical feasibility, data readiness, implementation complexity, sponsorship, expected ROI, success metrics, and operational risk.

Drawing on real-world work with European and international companies, including cases involving Allkind Group, Japan Tobacco International, C-Leanship, Vienna Insurance Group, and Homesearch, Golikov shows how AI projects are evaluated before implementation. The examples cover areas such as patent intelligence, AI-supported document workflows, maritime reporting, insurance processes, real estate data infrastructure, and AI adoption roadmaps.

Attendees will leave with a practical structure they can use immediately to score AI ideas, define success metrics, identify hidden blockers, and decide whether to greenlight, reshape, or kill a project.

The main message is simple: companies should not start with an AI model. They should start with the business problem, the value they want to create, and the proof they need before committing to implementation.

This talk introduces a practical three-layer framework for AI-native software engineering, shaped by Maulik Sailor’s transformation work at Innovify and his experience building Notchup’s AI Chief of Staff. He outlines how organisations can move beyond surface-level code generation to redesign engineering execution, augment management through AI-driven orchestration of people, products, and processes, and establish governance models that ensure strategic alignment, measurable ROI, and long-term resilience. The session offers a founder-operator perspective on how to adopt AI not merely as a tool, but as part of a structural shift in how software organisations think, operate, and deliver.

“Many enterprises now have an artificial intelligence (AI) governance policy. Far fewer have operationalized it. The gap between a framework on paper and continuous, auditable oversight of live AI systems is where organizations quietly stall—and where the EU AI Act’s 2026 application and enforcement milestones are exposing real weaknesses.

In this session, Kevin Schawinski examines this implementation gap.

Key takeaways:

A clear diagnosis of why AI governance programs stall during implementation, including the specific organizational and technical patterns that distinguish successful programs from unsuccessful ones.
A practical approach to moving from risk classification under the EU AI Act to continuous, board-reportable oversight of AI systems in production, grounded in real enterprise deployments across regulated industries.
Concrete questions that every chief information security officer (CISO), chief information officer (CIO), and audit committee should be asking their AI teams right now, covering shadow AI, agentic systems, vendor accountability, and how to quantify AI risk exposure.”

As organizations move from experimental large language models (LLMs) to autonomous artificial intelligence (AI) agents, the definition of “safety” is evolving faster than the tools designed to ensure it. In this session, Iñaki Velez de Guevara provides a comparative review of the current AI safety landscape, contrasting the robust infrastructure of major public cloud providers with the “frontier risks” identified in recent academic research. He examines real-world implementations of safety frameworks, analyzes what the industry has solved—and what remains unsolved—and discusses how to maintain human-centric control in sensitive sectors.

Most enterprises remain stuck in the demo phase: 88% of companies use artificial intelligence (AI) in at least one function, yet 94% see no measurable value from it (McKinsey, 2025). The gap is rarely the models themselves – it is everything around them: data access, legacy integration, language coverage, and organisational readiness.

Vedran Bajer leads Wonderful’s deployments across 12 markets in one of Europe’s most linguistically diverse and regulatorily complex regions, including deployments for companies in Lithuania, Latvia, and Estonia. In this session, he explains what production looks like at scale: how deployments move from contract signing to production in weeks, why mature systems achieve automation rates above 80% and approximately 64% end-to-end resolution, compared with an industry average of around 3%, and what breaks first when a single-market pilot becomes a multi-market operation.

Attendees will leave with a concrete deployment checklist drawn from real-world enterprise implementations—a practitioner’s view from the ground.

This session will break down what “search” actually means in the agentic era, in which retrieval is no longer a single lookup but a chain of reasoning over context. Ina Koleva will present a practical framework for choosing between three main search approaches—traditional indexes, large language model (LLM)-curated wikis, and fully agentic search—based on latency and accuracy requirements, as well as data format and scale.

Choosing the wrong search architecture can cost enterprises speed, accuracy, or both. Drawing on hands-on experience in applied artificial intelligence (AI) and product management, Ina Koleva, Product Manager at Mistral AI, will explain different approaches to grounding agents in enterprise data.

This session presents a practical architecture for governing decision systems supported by artificial intelligence (AI) in regulated organisations. Using a realistic enterprise case, such as AI-assisted recruitment screening, credit risk support, insurance claims triage, or customer service escalation, Glen Rai will show how an organisation can move from informal AI adoption to a structured governance model.

The session will guide participants through a step-by-step method for mapping AI influence, assigning ownership, categorising risk by tier, designing human oversight, setting review gates, managing vendor dependencies, and building an evidence ledger that can support audits, assurance, and internal accountability. The emphasis is on implementation: what should be documented, who should be accountable, when a system should be reviewed, and how governance can be maintained after deployment.

Participants will leave with a reusable decision-system canvas, risk-tiering criteria, a governance-gate model, an evidence-ledger structure, and a 90-day roadmap for initiating AI governance in their own organisations.

The session will also briefly reference TIRA, an emerging implementation layer developed from this architecture, showing how these governance steps can be translated into a structured workflow for AI inventories, risk reviews, evidence capture, and readiness reporting.

This is a how-to session for organisations that need AI governance to become practical, repeatable, and capable of withstanding audits, scaling effectively, and operating under real-world pressure.