What 164 sessions in Hamburg told us about the gap between an AI that works and an AI that ships.
🚀 One day, three halls, one uncomfortable conclusion
On 20 May 2026, around 7,800 people filled the Hamburg Messe for the AWS Summit. 164 sessions. 237 speakers. Five theme areas: Industry Solutions, Emerging Technologies (AI), Migration & Modernization, Governance, Sovereignty & Security, and Empowering Builders. 🇩🇪
We did not go to collect service names. We went to pressure-test our own roadmap against what enterprises are actually running in production.
Our verdict is short:
Capability is solved. Delivery is not. 🎯
A year ago, the question in every AI track was “does this work?” This year, nobody asked it. The question had moved to “does this work in production, reliably, controllably, and affordably?”
That shift is the whole story. And it is where most AI programme will quietly stall in 2027.
1️⃣ Governance is the real product
The clearest signal came from Deutsche Bahn. 🚆
In EDIS, the central dispatching system behind more than 34,000 km of track and 20,000 trains a day, AI agents patch over 2,500 servers in production, with a human in the loop for release approval.
What makes it work is not the model. It is the governance:
- 🪪 Every agent has its own identity and an “Agent Card” that records ownership and liability
- 🔒 A zero-trust narrative: every agent is treated as compromised
- 👤 Human approval by design, not as a safety net bolted on afterwards
DB Systel intends to migrate 2,500+ use cases to agents in the next twelve months. You do not attempt that number without an ownership model first.
AWS shipped the tooling to match. Policy in Amazon Bedrock AgentCore (generally available since March 2026) lets you write rules in natural language, translates them into the open-source Cedar language, and enforces them at the AgentCore Gateway against every tool call, without touching agent code. Alongside it, Automated Reasoning Checks in Bedrock Guardrails convert a business policy into formal logic and use a solver to verify the model’s answer against it.
The one-line summary worth keeping:
Policy controls what agents do. Automated Reasoning checks what they say. ✅
2️⃣ Economics decides which AI features survive
The second signal was financial. 💰
Delivery Hero had a GenAI feature generating menu descriptions. Projected cost: 700,000 dollars per year, with four to six seconds of latency. Internally, the feature was already stopped.
Three unglamorous changes brought it back:
- 🎯 The right model, not the strongest one
- ⚡ Prompt caching
- 📊 Custom inference profiles for per-country cost transparency
Result: cost down 80%, latency down over 50%, and the feature now runs globally.
The line we keep repeating to clients: you do not deliver food in a Formula 1 car. Model selection is a cost decision, and cost is a product decision. If you cannot see AI spend per market, per model and per agent, you cannot defend the feature in a budget review.
3️⃣ Sovereignty stopped being a philosophy debate 🔐
For regulated industries this was the day’s second thread. The AWS European Sovereign Cloud has been generally available since January 2026, with its first region in Brandenburg, a separate cloud with its own accounts, billing and identity management, operated entirely inside the EU, with staff, technology, network and leadership in the EU.
SCHUFA CTO Dr. Klaus Kolitz put it plainly:
“Data sovereignty is not optional, it is existential. We don’t have to sacrifice innovation for sovereignty. We get both.”
The most reusable slide of the summit came from session SEC201, where AWS’s Michael Hanisch laid out four sovereignty patterns:
- 🛡️ Data Perimeter: protection against external access and insider exfiltration
- 🎭 Tokenization: sensitive data pseudonymized before the model ever sees it
- 📋 Governed Access: purpose-bound, revocable, auditable via AWS CloudTrail
- 🔌 Kill Switch: cryptographic or IAM-based access severance
Turn that list into your architecture checklist. It is the shortest path we have seen from “we’d like to use GenAI on sensitive data” to something an auditor will accept.
4️⃣ Migration is now an operating-model question 🔄
DKB rebuilt its core internet banking on AWS, 14 frontends, 40+ API services, 42 databases, eight terabytes of highly sensitive data, with near-zero downtime and a canary cutover from 1% to 100% of live traffic. Operating costs fell by more than 50%, and teams took full end-to-end ownership.
Their headline recommendation was not technical. It was a “program mindset” and consistent decision records from day one. 📝
Trade Republic made the same point from the growth side: six new markets in eight months, achieved by replacing heroics with repeatable processes, and processes with automation. Or, as AWS’s Dominik Richter framed it:
“There is no compression algorithm for experience.”
🧭 What this means for MeJuvante, and for you
We build and run AI in exactly this territory: European regulation, Indo-German delivery, financial services and healthcare, 25 years of it. So we are adjusting where we put our weight in 2026/27:
- 🤖 Agent governance before agent count. Identity, ownership and audit trail for every agent we deploy; our AI Business Operations Hub and MeJu Health Hub ship with that model, not as an add-on.
- 💶 Cost transparency as a delivery requirement. Cost per request, per market, per agent, per outcome, visible from day one, not discovered in month six.
- 🔐 Sovereignty-native architecture. Data perimeter, tokenization, governed access and a tested kill switch as our default for regulated clients.
- 🗣️ Reach before sophistication. Internal search and conversational analytics, our AI Intranet Chatbot and Talk2Data, still deliver the fastest measurable return in most organizations, because teams lose 30–40% of their time hunting for information.
💬 Over to you
Five themes filled three halls in Hamburg. Very few companies can fund all five.
Which one is actually funded on your 2027 roadmap?
🤖 Agentic AI · 📊 Data & Analytics · 🔐 Sovereignty & Security · 🔄 Migration & Modernization · 🛠️ Empowering Builders
Tell us in the comments, we are compiling what we hear into an open benchmark for European mid-market AI programme.
And if you would rather skip the reading: book a 30-minute roadmap review with our team. We will tell you which of the five to fund first, and which to postpone. 📅
Source: AWS Summit Hamburg 2026 Recap — AWS Germany Blog, 15 June 2026 · Session recordings on demand