From Cloud Builder to AI Orchestrator

When Thomas Nowicki moved from the consulting firm Accenture to Handelsblatt, a special challenge awaited him: to technologically lead a highly regulated, historically grown media house into the future.

“I wanted to see what happens when you lift a complete system into the public cloud and simultaneously build new AI functionalities,” says Nowicki. The entry was traditionally through infrastructure and migration – but with the AI boom, his role quickly changed.

Together with Microsoft, he developed the first prototype of the now productively deployed Smart Search within a few days – a Retriever-Augmented-Generation system (RAG) that semantically analyzes editorial content and intelligently answers user questions. “That’s when I realized how much infrastructure is actually AI,” he says. From this insight, a dedicated AI team was eventually formed.

Handelsblatt's AI Strategy: Enhancing Products, Not Replacing Them

Many companies start their AI journey for efficiency reasons. However, at Handelsblatt, the focus is on the readers:

How can AI make a journalistic product more valuable?
How does it improve the user experience?
How can the subscription be made more attractive – especially for professional decision-makers?

Nowicki puts it succinctly:

“AI does not define our products. But it changes how we build products.”

There is a central principle: No AI-generated articles.
AI assists in research, analysis, and production – but the content remains journalistically created and verified.

Smart Search: From Full Text to Real Meaning

The first major milestone was the introduction of the new, fully semantic search. Previously, the search was based on classic full-text annotations. Today, the system recognizes meanings in context – through vector databases, embeddings, and semantic similarity scores.

The result:

• significantly more relevant search results

• increasing usage statistics

• positive feedback from editorial staff and readers

A special learning: Users previously found content more often through Google than through the internal search. That has completely reversed.

Text-to-Speech: Every Article as High-Quality Audio

Another project that has a great reach is the automatic audio narration of all articles.

Instead of relying on generic browser reading functions, Handelsblatt built its own system – together with the startup AudioStack.

Today, articles are:

• automatically narrated after publication

• audio-optimized

• delivered in a compressed format

• made available on the web and in apps

The usage is high. “Many readers prefer to listen rather than read,” says Nowicki – a trend that the media industry is increasingly sensing.

Build or Buy? How Handelsblatt Decides

Not every AI system is built by the editorial team. The criteria:

1. Competence – can the team realistically handle it?

2. Speed – can it be built in market-conforming time?

3. Economics – is in-house training more worthwhile than SaaS?

4. Quality – are there providers who have already solved the problem better?

Training speech synthesis on their own is for example unrealistic. “If OpenAI rents a nuclear power plant for six weeks to train a model – we cannot do such a thing.”

Therefore, the principle is: SaaS first. If the product can later be sensibly internalized, the option remains open.

No Hallucinations: AI with Quality Gates

Since journalistic integrity is central, Handelsblatt has implemented extensive security layers.

For example:

• Semantic thresholds: Only content that is close enough to the user context is taken into account.

• Internal relevance vectors: Articles from the editorial team are prioritized – agency material is only included in a structured manner.

• Multimodal validation: A second LLM checks whether the answer is substantively justifiable.

• “No answer” as a valid response: If no suitable source exists.

The goal: No hallucinations.

Internal Challenges: Expectations, Misunderstandings, Architecture

Nowicki identifies three recurring themes:

1. Expectation management

AI can do a lot – but not everything.
“One must not oversell it,” he says.

2. Involving the editorial team early

Data Stewards test new systems professionally.
What is not plausible is filtered out immediately.

3. Technical architecture

A well-thought-out pub/sub structure enables reuse, scalability, and rapid implementation of new AI features.

How Success is Measured

Handelsblatt measures AI features strictly based on data:

• Click behavior

• Use of Smart Search

• Time spent

• Scroll depth

• Use of audio

• In-product feedback

A central learning: A large part of the users click on content directly from the Smart Search – a strong engagement signal.

The Future: Handelsblatt as an AI-Powered Media House

For the upcoming years, Nowicki sees great opportunities:

• AI-powered analysis of financial data

• automatic indicators for corporate risks

• data-driven journalism

• intelligent linking within articles

• sentiment analyses

• context-based story scouting

Long-term, so his vision:

“Every company will become an API. And we have to keep track.”

The Desired Headline for 2030

In closing, Hahnenkamp asks:
What headline would you like to read in five years about AI in Handelsblatt?

Nowicki laughs – and responds:

“That we continue to be the leading medium for finance and economy.
And maybe someday also the leading medium for AI.”