Wednesday 3 June 2026 - Estrel, Berlin - several parallel tracks
Day-two notes, kept to what's useful to us. Most of this is pitched at big enterprises, so for each talk I keep the principle that survives being shrunk to a small team and drop the apparatus.
Linda Kohl: how to avoid agent sprawl
Databricks - Accelerate Stage - 09:40
Sprawl is what you get when everyone wires up their own agents. Her image for it: agents don't wait for permission or for clarity, they just act (like teenagers, but with API access), and a pile of them is a startup with no CEO, no org chart and an unlimited credit card.
The failure modes she named: low-quality reasoning on enterprise data, too many vendors, no way to measure quality, no governance. Each one lands on a different owner (who owns the agents, what data did one touch, what did it cost, how do I stop it deleting a dataset).
Her fix, minus the platform: strong data foundations, an outcome metric per agent, a quality gate on every agent and version, continuous cost/inference monitoring, and retiring agents aggressively, because sprawl compounds when nothing gets deleted.
A small team gets most of that without buying a platform: one list of which agents exist and who owns them, a couple of eval gates, real cost tracking, and actually deleting the dead ones. Cheap to set up early, expensive to retrofit.
Guido Vetter: scaling AI beyond pilots
Bain & Company - end-to-end process redesign - 10:00
His diagnosis of why pilots stall: they cost money long before they return any, so the spend is the only part anyone can see. And fragmented use cases (a pilot here, a pilot there) never add up to a result; his push was to stop collecting them.
His ladder: rules-based automation (5-10% gains) -> AI use-case optimization (20-30%) -> end-to-end process redesign (up to 80%, mostly from taking the human checkpoints out) -> fully AI-driven enterprise. The jump that matters is from speeding up steps to redesigning the whole process. (Day one's "human in the loop" flips here: redesign the process, then humans handle the exceptions.)
Concrete point: an automotive-supplier engineering process where redesign freed up 80% of the people's capacity. The graveyard he listed: tech bolted on with no redesign, AI in a silo, pilots with no specificity, wait-and-see.
This is exactly our work, and a small company has the edge: one process owner in the room, redesign quoting or invoicing end-to-end in a Shape Up cycle (fixed time, variable scope), ship the 80% version, do the highest-value process first, then the next. No transformation program required.
Alexander Schellinger: the missing link in healthcare AI
Siemens Healthineers - connecting payers and providers - 10:40
His frame: there's plenty of AI inside single departments, but every step between them is a break, so the intelligence never adds up. Locally intelligent, systemically blind (his phrase, more or less).
Three levels: local intelligence (one department, one imaging algorithm) -> organizational intelligence (across departments in one hospital) -> system-level intelligence (across providers and payers). The value, he argued, sits in connecting the existing tools to the workflow rather than in a smarter point tool: individual reports -> real-time intelligence, manual coordination -> orchestration, retrospective -> predictive.
Swap "hospital" for any small company and it's the same shape. The break between a CRM, an ERP and three spreadsheets is the small-business version of departments that can't talk. The win is wiring what's already there into one flow, not buying a fourth island that's smarter than the other three.
Grisha Pavlotsky: acceleration drift
Miro - 10:40
His claim: "AI adoption" is the wrong problem. He walked an insight through a team: a marketing director debates it with her AI, lands on "trust" over "speed," writes a brief; the copywriter's agent then leans, fast and confidently, into the wrong insight; the designer's does the same. Each person is faster, each "silo of 1" hits its own local maximum, and the result drifts away from the original intent. That's the acceleration drift of the title: the faster the individuals move, the further apart the team gets.
His test for it is brutally concrete: count the handoffs. Past about seven, he argued, the final outcome stops resembling the initial intent.
His fix is rebuilding the workflow, not adding copilots: the org chart becomes a workflow chart; middle management becomes a monitoring layer, then disappears; handoffs are a design failure, not a feature; reduce them by at least 3x; feedback loops close in hours; every decision sits on a real-time intelligence layer. And his version of the line that kept coming up all day: the bravest leadership decision is what not to automate.
The local-vs-global-maximum picture is the one I'd use to explain why copilots aren't enough: every person reaches their own local max and the team still drifts. And "count the handoffs, cut them 3x" is a test I can actually run on any process we rebuild.
Pascale Schäfer: AI on 150-year-old rail infrastructure
Deutsche Bahn (DB InfraGO) - 11:00
No hype: plenty of self-awareness about the punctuality reputation (50,000 trains a day, old infrastructure, no buffer; they feel it every day), and a refreshingly low-key take on the thing everyone else was selling: they consider themselves at the very beginning, and AI is not the new 42, the answer to every question.
What's shipping, by her account, is all unglamorous and document-shaped: generative-AI checks on construction-approval paperwork (a wrong submission to the Eisenbahn-Bundesamt (the federal railway authority) costs three months), the same components reused on contract change-requests (25,000 of them, ~1M hours of checking), object recognition on forward-facing train video for vegetation near tracks. And the nice twist: pointing that same vision pipeline back at their own records. Everyone says you need good data to do AI; they also use AI to fix the data.
The operating model, as she described it: an "AI discovery team" gets six weeks to challenge an idea, kills a high share of them, and drops anything with no owner. Stopping a topic is counted as a success there. They started in 2020 with the data and the platform, not a product (homework first; she'd rather drop the "artificial" and just call it intelligence), then reuse one component (risk assessment, documentation) across products and swap the backend without users noticing.
The most directly copyable talk of the day. Data and a reusable foundation first, one component reused across tools, kill low-value ideas in weeks not quarters, point AI at your own messy data to clean it, and decide the value before building. None of it needs DB's scale: it's the same discipline a small team can run, just without a six-week committee.
Marie-Helene Ametsreiter: what Europe can learn from China
General Partner, Speedinvest - scaling AI startups - 11:40
The numbers that frame it: Europe has 5 of the world's top 10 universities and 0 of the top 10 AI companies; 27,000 new deep-tech startups last year but only 6% of global AI investment; 88% of corporate-startup AI pilots never reach production. Her diagnosis: the bottleneck isn't talent or ideas, it's demand. The fix is for big companies to become a "customer zero": buy, not just admire.
An off-take agreement (a commitment to buy before the product exists) or a real procurement contract, she argued, helps a young company more than a cheque. Her China contrast: a chain only works if every link does, and China, in her telling, engineered the whole chain while Europe optimizes each link in isolation. Her connective tissue from lab to exit: corporate joint funds (industry peers plus state co-investing, procurement written into the deal). And a reality check for corporates hoping startups will just show up: the best ones won't come to you by themselves.
We're not a VC and we're not raising money, so what I keep is the customer-zero idea: one real customer running the thing in production teaches you (and protects you) more than any demo. The other half I keep: 88% of pilots dying in procurement is a big-company disease. A five-person company can decide, start and be using the thing within a week.
Marc-Andreas Albert: AI as a colleague
CEO, Webedia - org hierarchy & the future of skills - 12:00
His claim: AI hits two things, the skills we keep and the hierarchy that organises them. On skills, the danger he names is the skill ladder. We all started as juniors doing grunt work and climbed: grunt work -> basic -> intermediate -> expertise. In his picture AI swallows the bottom rungs, leaving a "??" gap where juniors used to climb. And skills atrophy anyway: use it or lose it.
On hierarchy: it exists to route control, and he thinks it falls. His "intelligence layer" (owners / directly-responsible individuals, builders / individual contributors, player-coaches) builds on Jack Dorsey and Roelof Botha's From Hierarchy to Intelligence: put the intelligence in the system, people on the edge. His own framing, the "four technology effects" (task displacement, tool effect, tutor effect, task creation), leads to the bet: invest in the tutor effect and in creating new human work, not just displacement. His closer echoed thyssenkrupp the day before: the hard leadership decision is not what to automate, it's what to leave alone.
The skill ladder is the one we keep coming back to. On a small team there's no big org to absorb the missing rungs: if AI does all the grunt work, your juniors never become seniors. So the question isn't only "what can we automate" but "what learning are we automating away", and where to deliberately keep people on the lower rungs (or build the tutor effect in) so they still climb. Automate the toil, protect the apprenticeship.
Jens Polomski: from user to operator - the 5 layers of AI-native work
Founder, snipKI - ~12:00
This one is basically our pitch. The frame: from Excel sheet to app, same purpose, a real interface. Software is getting smaller, more personal, more temporary: the classic build is for millions of users, months of work, pixel-perfect and scalable; the new one is for one person or one team, hours of work, functional and temporary.
The progression to climb is Question -> Task -> Initiative: ask how to write a good pitch and AI answers; tell it to research and write the pitch and AI acts; at the top there's no question at all, the system recognises the moment and reaches out. His own proactive setup: a 10-minute briefing before every meeting, auto-summaries to Notion and to-dos to Asana after, and a 7:30 daily content alert in his voice from the news and his last 50 posts.
Why now: dictation beats typing roughly 3x (~150 vs ~50 wpm; 1,000 words in 5 minutes vs ~20), and voice-to-text is past 95% accuracy (Whisper, SuperWhisper, Wispr Flow, MacWhisper). Per the agenda he also shipped a viral no-code app to 160,000 users, as a marketer.
This is the clearest statement of the job: replace the spreadsheet and the one-person workaround with a small, real, functional app, in hours not months. His "temporary, for one team" software is exactly the right-sized internal tool, and his proactivity patterns (pre-meeting brief, self-following-up meetings) are product ideas worth stealing.
Deborah Hüller: what happens when AI actually works?
Partner, IBM Consulting - what comes after adoption - 12:30
Her premise: AI has stopped being a niche technology and become a general-purpose utility, so the interesting question is no longer whether it works but what happens when it does. The answer she keeps seeing: work doesn't shrink, it shifts. Routine down, complexity up. Her three paradoxes, paraphrased:
more available knowledge raises the value of judgment;
removing the routine work makes expertise harder to build;
the further the technology goes, the more the human side matters.
That second one is Marc's skill ladder again, from a different stage. On IBM's own "Client Zero" transformation (running all this on themselves), her read was that it was never really a tech project: technology was necessary but never sufficient, and the change came when people, process, culture and tech moved together.
The expertise paradox is the practical warning for a small shop: automating the routine removes exactly the reps people used to build mastery on, and we have no bench to hide that. So protect a few rungs on purpose, and treat any AI rollout as a people-and-process change, not a tool install. "Necessary but not sufficient" is the line to keep.
Luise Mohn: walk the AI talk - from vision to impact
Director of Analytics & Data, Cosnova (essence / Catrice) - ~12:50
The anti-top-down talk. The failure mode she's watched play out: AI done top-down and tech-led (which model, which feature), people overwhelmed, and everyone quietly back in their spreadsheets within a few months, just because it's easier.
Cosnova, by her account, did the opposite: bottom-up, not tech-led. The foundation is a Champions Program, 64 internal champions so far (roughly 6-7% of employees): people with domain expertise, domain authority and technical curiosity, chosen through an internal application process and sitting decentrally across functions. They surface the friction in their own workflows (where does your time actually go, what stops us delivering the quality the consumer wants); a small central AI team turns those into a use-case pipeline and prioritises by impact (their operating model is a central/de-central x business/technical matrix). A live demo built a Catrice "Campaign Creator Palette" concept in ChatGPT. The point she closed on: AI amplifies people who know their business, so free them up to spend more time on the consumer and on creativity. One result cited: a copywriting task down ~80% in time.
Start from the people who know the work, not the tech. I have watched the back-to-spreadsheets-within-months failure happen too, and the champions pattern (a few domain people who spot friction, a small team that ships) works at any size. In a small company the champions and the team are the same handful of people, which makes it easier, not harder.
Jakob Freund: every enterprise process is already legacy
CEO, Camunda - "the great re-engineering" - 14:30
His image: bolting AI onto how you already work isn't transformation, it's a caterpillar with a jet engine strapped on. We'd all like the butterfly; what most companies are doing, he argued, is the jet-engine caterpillar. The uncomfortable conclusion he draws: every process in the enterprise is, by definition, legacy, because it was designed for a world where humans did the reading, routing and deciding.
He cited Barclays' global chairman calling business-process re-engineering the most important thing his organization will do over the next 2-3 years, and reframed the goal as continuous re-engineering: every process has a number (cycle time, cost, error rate, revenue impact) that today only moves when you run a project. What if it moved every week? Then the product pitch (Camunda's ProcessOS / agentic orchestration), including a tidy "safe Customer Data Agent" where the delete step routes through a human approval gate.
"Every process is legacy" is the same drum as Bain and Miro, hit hardest. The transferable part isn't the platform, it's the question to put to any client process: if you designed this today, knowing what AI can do, would you build it this way? And the safe-agent pattern (a deterministic gate on the destructive step) is exactly how to let agents act without trusting them blindly.
Florian Matusek: AI is not enough - outcomes over tech
Director AI Strategy, Genetec (physical security) - 15:10
He opened with a cautionary tale: a client called urgently, needing to export his video-surveillance data to feed it to an AI, budget no issue. Which AI? Undecided. To help the operators do what? Hadn't asked them. What problem? Not defined. The board had simply decided the company needs to utilize AI. His counter: start with the problem, and actually define it before anything else.
Three lessons: define the problem, get the buy-in of the organisation, then choose a tool you can trust. Technology last, not first. That, in his telling, is why AI on its own is never enough.
"The board said do AI, so let's export everything and feed it to an AI" is the exact anti-pattern we walk clients out of. Problem first, buy-in second, tech last, and a tool you can trust over the shiniest model. None of it needs scale; it's the discipline that makes a small project actually land.
Firas Ben Hassan: agentic AI in the enterprise - people, tech, culture
Head of the Agentic AI Solutions Hub, Allianz Tech - 15:30
His open: 95% of his ~40 slides (text and images) were generated by AI agents, saving ~6 hours, with (he reassured us) the human in the loop to validate. He ran the frameworks (the 2023->2026 arc from hype through FOMO to value; a five-level maturity model from simple chat to self-driving agents; chatbots and copilots are not agents, because generative AI creates while agentic AI acts; the prototype -> enterprise-GenAI -> agentic-workforce arc), but his real subject was change management, delivered bluntly. On motivating adoption with "use AI or lose your job": he dismissed that as nonsense from the stage, in considerably blunter words.
He also described, openly, the opposite of the human-in-the-loop slogan: a claims process he said is live in production in Australia (seven-plus agents and an orchestrator handling food-spoilage claims end to end) with no human approval step. Full automation, he argued, is where the productivity actually shows up, not in an agent that pings you "please approve" at every step.
His point on adoption: people are frustrated mostly because they have no time, so make "learning is working" real and embed it in the job. Two more that travel well: AI leaders prioritise agentic use cases by reusability, not per-case ROI (build once, reuse many; the returns follow at scale), and bring the safeguarding functions (security, privacy, governance) into the build as champion co-constructors rather than training people at the end. And on running the change: understand the pain first, and make the people who feel it contributors to the design, not just executors of it.
The honest parts are the keepers: don't motivate with fear, make learning part of the work, and design governance in from the start rather than bolting it on. Reusability over per-case ROI is the Deutsche Bahn lesson again: build one reusable component, not a pile of one-off pilots.
Gordian Braun: voice agents in production
Head of Growth Europe, ElevenLabs - 16:00 - I missed a few minutes in the middle
His case for why voice is different: it carries tone, timing, emotion and trust on top of the words.
The opening visual: typewriter -> old PC -> laptop, and the one thing that never changed is the keyboard. Voice agents are in production now (Deutsche Telekom, Klarna, Lidl, Freenet...), usually starting with inbound customer support but spreading fast: the Deutsche Telekom team realised there was far more to do than automate support (e.g. inbound sales qualification instead of waiting for a sales rep to call a form lead). He ran a live e-commerce "shopping concierge" demo (Alexis, the ElevenLabs apparel concierge). Product claims: 65 ms real-time latency, emotion- and context-aware speech recognition and synthesis, strong turn-taking and interruption handling, 11k+ voices, 70+ languages. His vision: leave the typewriter's keyboard behind and bring human interaction and technology interaction into the same pace.
Voice is becoming a real interface layer, not a gimmick. The interesting pattern isn't "automate the call centre" but what teams find once a voice agent is live (sales qualification, internal flows). For a small company the bar just dropped: off-the-shelf voice agents with decent turn-taking let a tiny team answer the phone in a way that used to need a call-centre budget.