From Large Language Model to ROI.

Read the full DMS Talk transcript

‍

DMS Talk – From LLM to ROI

Using AI in Retail and Consumer Goods

Oliver Nitz (DMS):
Welcome to our DMS Talk – I believe this is already the tenth one.
From LLM (Large Language Model) to ROI (Return on Investment): Using AI in Retail and Consumer Goods.
A long title – and the key question: Generative AI – higher, faster, further?
AI now affects nearly every value chain. Where do we stand today? AI has become indispensable in many workflows. But what about tangible performance and practical solutions? What can we actually use it for?

We’ll explore these questions today – together with my guest.
My name is Oliver Nitz, I’m with DMS and I’ll be your moderator today.
We at DMS have been active in retail for 20 years, expanding physical spaces with digital solutions – primarily digital signage, frequency and motion analytics, and much more. That’s exactly where we use AI: to analyze large data volumes and help clients make informed decisions.

Please welcome Norbert Hillinger, AI & Retail Associate at Retail AI GmbH.
A long title – what does “AI & Retail Associate” actually mean?

Norbert Hillinger (Retail AI):
Thanks for the invitation, Oliver. I’ve been working with Retail AI for a few months now. As an AI & Retail Associate, I get to look across departments within the corporate startup and, after a sort of trial phase, focus where I can deliver the most value.
As a trend and futures researcher with lots of training experience, I find it particularly exciting to see how generative AI is taking effect in retail. And yes – I fully expect to “extend my stay.” AI is here to stay.

Who is behind Retail AI?

Oliver:
What exactly are you working on?

Norbert:
Maybe a step back: Behind Retail AI is the Markant Group (Switzerland) – known in retail as a platform connecting manufacturers and retailers (offering pricing, payment, and inventory services, among others). Markant acts as a technical service provider – not logistics from A to B, but the digital/technical infrastructure behind it.

To address artificial intelligence, Markant founded a corporate startup – Retail AI, based in Frankfurt (launched early 2024). We’re still small (<10 employees) and work across four pillars:

1. Product Factory – develops AI applications, especially generative AI for retail and consumer goods, ideally built on the Markant platform.
2. Labs – a “sandbox” for cross-innovation with other startups (e.g. combining AI with AR/VR, blockchain, etc.). Innovation often happens at the intersections.
3. Alliance – a think tank or safe space where practitioners (from the DACH region) exchange experiences with AI (challenges, opportunities).
4. Academy – focuses on upskilling and workforce transformation: training programs to build AI competence within organizations.

We’re not non-profit – the goal is to learn, build, bring to market, and thus create both value and business.

Hype or Substance?

Oliver:
AI is influencing every value chain. Where do we stand – and what actually works today?

Norbert:
AI itself is old – but generative AI at this scale is very new. Since late 2022 (with ChatGPT), the pace of innovation has been enormous. Many refer to the Gartner Hype Cycle – expectations are sky-high. A reality check is important: not everything that’s demonstrated is already robust. Yet we often underestimate how much becomes possible in a short time.

A practical tip: AI directories (such as There’s an AI for That) list thousands of tools – with new ones emerging daily. In retail, there are use cases along the entire value chain.

Examples Across the Value Chain

Norbert:

• Procurement: Negotiation bots (e.g. “Nelson” at Deutsche Telekom) conduct purchasing dialogues on terms and conditions.
• Category Management: With generative AI, you don’t just ask “How’s butter performing in Eastern Austria?” – you also get actionable insights (“Sales +6%, but Kerrygold weaker – suggested actions …”).
• Store & Checkout: Computer vision and sensors enable self-checkout and “pick & go” concepts (à la Amazon Go).
• Marketing/CRM: Personalized content, AI chatbots for service and consulting, content generation (text/image/video/audio), live-shopping automation.

Service & Chatbots – Threat or Opportunity?

Oliver:
Customer service: There are smart bots – but sometimes people just want to talk to people.

Norbert:
Both are valid – it’s a question of values and strategy.
Klarna is a standout example: its bot handles about two-thirds of all inquiries with comparable quality, but much faster – equivalent to hundreds of FTEs and saving tens of millions annually.
Other companies deliberately focus on human excellence (like Tony Hsieh’s legendary Zappos example). What matters is alignment with the brand.

Internal AI: Unlocking Knowledge, Staying Secure

Oliver:
What’s happening internally – beyond website bots?

Norbert:
A huge topic. Many organizations are building internal LLM assistants (“copilots”) trained on their own data: intranet, manuals, reports, knowledge bases.
Example: dm drogerie markt uses an internal model with several thousand users and tens of thousands of prompts.

Possible strategies:

• Taker: Use external models (e.g. OpenAI, Claude, Llama) – but check data protection!
• Maker: Build your own model (common among large players, e.g. banks).
• Shaper: Combine a base model with own data (RAG/sandbox), EU-based hosting or on-premises – often the middle ground.

The key is enablement. Without AI literacy (prompting, source validation, governance), the potential remains underused – that’s where our Academy steps in.

Data, Bias & Compliance

Oliver:
What about data protection, copyright, and bias?

Norbert:

• Data protection: Minimize data use; employ RAG architecture; use EU-cloud or on-prem hosting; control logging and training.
• Copyright/Terms: Check usage rights for each tool (e.g. Adobe Firefly – trained on licensed assets, explicitly approved for commercial use).
• Bias/Ethics: Training data reflects human bias (see the Apple Card credit limit debate) – so governance and quality assurance are essential.

For video/image data, anonymization tools exist (e.g. brighter AI, Berlin): faces and license plates are blurred, analytics remain possible – GDPR-compliant.

Content & Commerce – What Already Delivers Measurable Impact

Norbert:

• Content recycling & scaling: Automatically generate short clips, product descriptions, and snippets from live-shopping shows; turn webshops into shopping apps within minutes.
• Post-production: Automatically remove objects from videos (e.g. Runway), distribute auto-formatted content across channels – hours instead of days.
• Service/Audio: Text-to-speech in multiple languages and voices – quick iterations.
• Personalization: AI-driven newsletters with improved tone and relevance – often achieving higher KPIs than manual ones.

Agents & Workflows – The Next Stage

Norbert:
We’re moving from single prompts to agents that manage workflows: one prompt is refined, triggers appropriate models/tools, and produces multi-format output.
Soon, “digital twins” (personal shopping/service agents) will negotiate bot-to-bot – creating new spaces of experience and efficiency.

Practice at DMS (Oliver)

Oliver:
We use AI, among other things, for visualization (e.g. pop-up concepts), GPS-based motion analysis (3-meter accuracy, origin, paths, target group clusters), and content scoring (automatic evaluation/optimization of advertising assets for screens and social).
Data protection is always an issue – many clients prefer non-visual sensors. AI helps us analyze volumes, optimize control, and iterate faster.

Regulation & Outlook

Norbert:
Please don’t view AI solely as a technology topic. It’s about interaction:

• Social: skills and culture
• Tech: models, data, MLOps
• Eco: business case, energy/CO₂
• Reg: e.g. EU AI Act – risk-based governance

Data centers need clean energy (think fusion or green power); synthetic data will become more important (as “natural” training data runs short); bias management remains essential.

Big players (e.g. Nestlé, Schwarz Group/Kaufland) are deeply involved, but small companies also surprise with fast, bold pilots and quick time-to-value.

‍