Accelerate Tomorrow AI Summit - Estrel, Berlin - 2-3 June 2026
The same handful of ideas surfaced talk after talk, on both days. Each theme below is my paraphrase of what the speakers argued (with links to the full write-ups), read through the one lens that matters to us: what a smaller company should do about it.
Human in the loop, but governed
The one-super-agent pitch is gone; what replaced it is many narrow agents, and a real fight over how much human stays in the loop.
Johann Strauss (Dell) warned that AI doesn't fail loudly: it fulfils every KPI you gave it while going, quietly and logically, in the wrong direction.
Christian Gondek (thyssenkrupp) described building with many narrow agents talking to each other, each constrained, each with a human in the loop, instead of one super-agent.
Linda Kohl (Databricks) on why sprawl is dangerous: agents don't wait for permission or clarity, they just act.
Deepak Alse (ProSiebenSat.1) drew a useful line: agency is not autonomy. An agent acts, but a human still picks the goals.
Counterpoints: Guido Vetter (Bain) argued for taking the human checkpoints out once a process is redesigned (people keep the exceptions), and Firas Ben Hassan (Allianz) described a claims process he said is live in production without a human approval step.
Govern from day one: who owns which agent, what it can touch, what it costs. Cheap as a habit, painful to retrofit.
Pilots are not value: make it pay or kill it
Guido Vetter (Bain) was blunt about why pilots stall: AI mostly shows up as money going out, not value coming back, and fragmented pilots never add up to a result.
The "Cautious by Design" talk (BioNTech): caution doesn't eliminate risk, it only moves it into your future.
Pascale Schäfer (Deutsche Bahn) described killing ideas on purpose: AI is not the answer to every question, and a stopped topic counts as a success.
Marie-Helene Ametsreiter (Speedinvest) put a number on it: 88% of corporate-startup AI pilots never reach production.
A real customer running it beats a clever demo; decide the value before building, and delete what doesn't earn its keep.
Where it can't be wrong: deterministic & auditable
The EPAM with First Derivative session: for real decisions AI has to be predictable, repeatable and auditable. "Probably right" is not enough.
Till Behnke (Rulemapping) showed "Law as Code": a rule-based architecture that follows legal logic instead of predicting the next word.
Dr. Jochen Kokemüller (Bosch) on governance at machine speed: his thesis was that it may take agents watching agents to scale safely.
For the regulated bits, wrap the model in deterministic controls (policy-as-code, gates) rather than trusting it to behave.
Connect the silos: the data foundation is the real advantage
Alexander Schellinger (Siemens Healthineers) described healthcare as locally intelligent, systemically blind: plenty of AI inside single departments, almost none across them.
Pascale Schäfer (Deutsche Bahn) turned the data-quality cliché around: everyone says you need good data to do AI, and you can also point AI at your data to clean it up.
Linda Kohl (Databricks) on hygiene: retire agents aggressively, because sprawl compounds when nothing gets deleted.
Wire the CRM, the ERP and the three spreadsheets into one flow instead of adding a fourth island. That is the actual job.
The skill ladder: don't automate away how people learn
Marc-Andreas Albert (Webedia) on the rungs juniors climb: AI eats the grunt work at the bottom of the ladder, and the hardest leadership call is what not to automate.
Deborah Hüller (IBM Consulting) made the same point from another stage: removing the routine removes the reps people used to build expertise on, and the human side matters more as the technology advances.
A small team has no bench to hide the missing rungs: automate the toil, but protect the apprenticeship on purpose.
From hierarchy to intelligence (and acceleration drift)
Grisha Pavlotsky (Miro) named the failure mode acceleration drift: the faster each person moves with their own AI, the further apart the team drifts. His test: count the handoffs; past about seven, the outcome has little to do with the original intent.
Deepak Alse (ProSiebenSat.1) put it as protocols over org charts: model how decisions actually flow, not the boxes on the chart.
Small and flat is the advantage: fewer routing layers to dismantle, so reorganise the work around the agents directly. Count the handoffs.
Demand beats tech: be (or find) customer zero
Marie-Helene Ametsreiter (Speedinvest) asked corporates to become a "customer zero": a real procurement contract helps a young company more than a cheque, and the best startups won't come to you by themselves.
Guido Vetter (Bain) pushed in the same direction: away from fragmented use cases, toward whole processes.
Find the customer zero: one client running the thing in production beats any demo. (And small companies don't die in procurement; they can just start.)
European sovereignty
Dr. Jochen Kokemüller (Bosch) framed sovereignty as a hedge: on US-controlled infrastructure, someone else could read your data, or switch your agents off.
Johann Strauss (Dell) framed it as a design decision, not a location: spread your bets across providers instead of marrying a single (US) vendor. The Paradox talk echoed it.
Our instinct already: independent, self-hostable, EU infrastructure over hyperscaler lock-in.