Accelerate Tomorrow AI Summit 2026 (day 1)

Till Schmid: Opening the Doors to Tomorrow

Founder, ATS - ~10:00

The founder's opening, and the point of the whole event: people in completely different fields are quietly working on the same problems, and they almost never meet. We all look for AI answers inside our own industry and our own network, so the questions stay stuck where everyone shares the same blind spots. The conference is built as the missing room: put the hospital person, the rail person and the bank person at the same table, and let one field's solved problem answer another field's open one.

• Built by practitioners for practitioners, period (his framing). And he asked people to share their failures, not just the wins.
• First edition, so nothing will be perfect, and his point was that perfect is the enemy of starting something new.

Johann Strauss: Quo Vadis AI?

CTO, AI Solutions, Dell Technologies (responsible for AI across the EU) - ~10:20

The Dell keynote. The core claim: AI isn't new, but it's now moving four times faster, and AI by itself is not a strategy (it's a new form of digital transformation, one that fails quietly instead of loudly). Dell sells the boxes all of this runs on; keep that in mind throughout.

• 368,000 companies changed their domain from .com to .ai in three years. That, he noted, is not a strategy.
• Efficiency with AI is becoming a commodity, so company size stops mattering: a one-person company, his point was, can now run on agents.
• His image: agents as "Minions", with you staying the master. Without oversight AI doesn't crash loudly like a driverless F1 car; it goes quietly and logically, fulfilling every KPI you gave it, in the wrong direction.
• Six "core capabilities" rather than products: RAG (retrieval-augmented generation: the model looks things up in your documents before answering), content, data, coding, analytics and IT agents.
• Sovereignty as a design decision, not a location: something you choose, wherever you run.

Christian Gondek: Building up an Agentic Organisation

Head of Digitalization & AI, thyssenkrupp - ~10:45

They're building an "agentic organisation" across four decarbonization businesses.

• His framing was that the one-huge-super-agent pitch has faded; what they build now is many narrow agents, each constrained, each with a human in the loop.
• AI won't fix a broken process, and bolting it onto bad data with no governance from day one just fails faster.
• Their rollout: don't wire every source system into the middle (expensive, slow, and it doesn't transform anything). Do it one use case at a time (engineering, then construction, then procurement) so the departments start learning from each other.
• A partial data-mesh: each unit owns its data, and consolidation happens at the segment level via data contracts. The data is connected, but there's no central database.
• His numbers, his to back up: first 5 use cases in 6 months, segment IT cost cut by almost half.

His closing line, paraphrased: prioritization is less about focus than about deciding what you won't do.

Till Behnke: Beyond the Black Box - Bringing Legal Reasoning to AI

CEO & Founder, Rulemapping Group - ~11:50

His argument: a probabilistic LLM is the wrong tool when a legal decision has to be deterministic, auditable and hallucination-free.

• His cautionary example: the German government putting 100 million euros into an agentic-AI setup that hands the law and the cases to an agent and asks it for legal decisions. Dumping thousands of pages on an LLM gives you thousands of probabilistic answers, which means more work for lawyers, compliance managers and courts, not less. You want something you can hand to a court to audit.
• Decompose one decision into thousands of micro-decisions, let AI touch only the ~5% that genuinely needs it, and have a human check only that 5%. The rest is plain data and legal logic.
• "Law as Code" as an open, standardized format (the editor is at rulemapping.org, funded by SPRIND, the German agency for disruptive innovation). Open source and EU-funded.
• The demo: a 1,975-page wind-turbine permit checked against paragraph 9 of the BImSchV (the German emissions-control regulation such permits run under), a process that normally takes three to four years, run as a three-minute job.
• The legislation process, he claims, drops from 18 months to about 4 weeks, and you get a digitally executable version of the rules as a by-product.

Eduard Singer: Cyber Security in the Era of AI

Founding Member / Head of Working Group "Cybersecurity & AI", KI-Bundesverband (the German AI industry association) - ~12:10

Cyber security with AI on both sides of the table. A couple of ideas stuck.

• The attacker's AI has never seen the inside of your house, so your own internal, context-rich data is the home advantage.
• The "Agentic SOC" (a SOC is the security operations centre, the team that watches for and responds to attacks) is three chained agents (ingestion, enrichment, validation) with a human approving at the end, taking incidents from hours to minutes. Speed is the whole pitch.
• Four autonomy levels matched to risk: assist, recommend, execute-but-overridable, execute-independent.
• An AI doesn't get an employee login, it gets a just-in-time key that vanishes after the task: if it's compromised, the attacker hits a dead end.
• His number: attackers sit inside a company for 256 days on average before they're detected.

Aatraye Almast-Gentzcke: Cautious by Design - Germany's AI Paradox

Director Strategy & Digital, BioNTech - ~14:00

[Name per the official agenda.] She flagged the views as her own, not her employer's. Germany has the ingredients (DFKI, the national AI research centre, in 1988; Industrie 4.0, the smart-manufacturing programme, in 2011; around 935 active AI startups): early, but not fast, and then it does very little with them. As of 2025, 64% of medium and 77% of small firms still don't use GenAI.

• The "four comforts" that cost us, in her German: Perfektion über Iteration (her image: polishing silver while others ship gold), Vorsicht über Experiment, Prozess über Ergebnis (a 47-page concept paper for a 6-week pilot; living in the scaffolding instead of the building), Effizienz über Innovation.
• The practical asks were good engineering practice with a German accent: a 90-day ship-or-kill review for executives, a six-week pilot rule for builders, and measure stickiness, not deployment.
• The vision: the Germany that built BASF, SAP and BioNTech, not the one still waiting for the Termin (the appointment).

Her sharpest point: caution does not eliminate risk, it only moves it into your future.

Fabio Baerwald: Building the AI workplace at REWE Group

CPO, REWE digital - ~14:20

A "we actually shipped this" talk: getting AI to ~20,000 office staff inside a cooperative of ~380,000 employees and ~10,000 stores whose owners think in generations.

• Two years ago they built their own GPT: a solution looking for a problem, with (as he put it) no adults in the room. The lesson that stuck: the main focus is adoption, not technology. Don't spend three months picking the "best" tool; pick one and get people using it daily.
• His model: an adoption curve (a long tail of 30-40k people doing basic prompting, a thin expert tip) sitting behind a plain digital-literacy curve (colleagues still emailing Excel files around). IT's job shifts from switching on tools to moving people along that curve.
• Strategy in three columns: custom frontline tools (a store-knowledge RAG (a chatbot that looks answers up in their own documents) cut search time; a butcher's job is half documentation), back-office as a fast follower (their use-case wishlist amounted to 20 years of IT work, so they decentralised; roughly half is doable with no-code/low-code), and the workplace itself.
• Numbers, honestly caveated: Copilot bought ~2 hours/week for most people and ~4 for the ~25% who leaned in (the mediator was adoption, not the licence). An IT-support agent halved ticket volume and surfaced more tickets (people had been suffering in silence). On that frontline RAG: easy to get from 80/20 to 95/5, much harder to be accurate every time at scale.
• The most-requested feature by far: people want their own chatbot, for their own project or team. They won't hand-build thousands of them, so it has to be self-serve.

His closing bias: don't build your own GPT (big tech will always be faster), be a fast follower, and over-invest in execution and adoption instead of strategy decks. The honest wrinkle he admitted: the "2 hours saved" number didn't convince the board; they wanted it tied to concrete process change.

Thomas Henschen, Leon Orr & Mik Quinlan: What if AI can't afford to be wrong?

Thomas Henschen (EPAM), Leon Orr (Head of Solutions, First Derivative, an EPAM company), Mik Quinlan (AI Architect, EPAM) - ~15:00

This one was openly a vendor demo (EPAM, on Microsoft Azure, with a First Derivative tool for KYC, the know-your-customer checks banks must run on clients).

• In a client meeting he retold, an architect was proud of hitting 90 to 95% accuracy; the COO across the table only wanted to know what happens to the other 5%.
• AI has to be predictable, repeatable and auditable. Veracity, groundedness, explainability: they all just mean "it's probably right", which is fine sometimes, but not when you need absolute certainty.
• Keep the probabilistic model away from the decision. "Intent Engineering" (the rules and policies are the what; the intent is the why), plus a deterministic control stack (eliminate capabilities outright, push logic into deterministic code and locked-down MCP servers (MCP: the protocol that gives a model controlled access to tools and data), hard human-authorization gates, policy-as-code).
• In the KYC demo, the model sifts the messy data, but once a counterparty turns up in a sanctioned country, policy applies deterministically. You never let the model decide whether someone is a "good actor".

Their concrete example of why: you can't tell an agent to never approve claims above 100,000 pounds and trust that it won't. It may comply, it may not, and for this class of work "may" is not good enough.

Dr. Jochen Kokemüller: The Agentic Leap - Scaling Value in the Era of Autonomous Intelligence

Head of AI Governance, Bosch - ~15:40

Bosch's head of AI governance , and an advisor to the EU Commission on the AI Act (the EU law that regulates AI systems by risk class). His opening: so far, all of us have been doing "innovation theater", and the promised benefits keep not showing up in the numbers. The rest was an architecture for fixing that.

• An agent is not a chatbot: it's a capability that runs for hours or even days and creates value over that whole stretch.
• On UX, his bet is that the prompt box dies: give a goal plus a guardrail and work backwards from a KPI. And don't "pave the cow paths" by automating a process built around human limits an AI doesn't have.
• His thesis: you can't govern agents at human speed, so governance itself may need agentic support. Rules and a regex first, then LLMs as checkers, then a policing agent over the swarm.
• Sovereignty, in his telling, is not really optional given US Cloud Act risk (a US law that can compel US providers to hand over data wherever it lives; here that means both reading your data and being able to switch your agents off). And compliance, in his telling, is not a hurdle but a speed booster.

Deepak Alse: The New Change Equation - Humans, Agents, and the Leaders Who Bridge Them

Chief Data, AI & Analytics Officer, ProSiebenSat.1 - ~17:05

The closing talk (lots of acronyms, which the speaker cheerfully admitted he likes).

• Agency is not autonomy: agents mostly can't pick their own goals, a human still does.
• The soccer-camp metaphor for the whole moment: everyone, vendors included, is learning on the job.
• His best warning: don't lie to yourself. Build your strategy on optimistic PowerPoint and your agents get fed written-down lies, amplified the moment you start scaling.
• The org claim: stop drawing org charts, model how decisions actually flow (protocols beat org charts; agents can "swarm").
• And the humane bit: treat people as collaborators with agents, not as people about to be replaced by them. Leaders still owe direction, alignment and commitment.

His closing thought, paraphrased: leadership is not a role, it's a commitment to kindness and curiosity.