Demand Planning in Retail — What You Really Need to Know.

Transcript of the DMS Talk – For Reading

DMS Talk – Demand Planning in Retail with AI

Oliver Nitz (CMO, Digitale Mediensysteme):
Hello and a warm welcome to DMS Talk number 7. It really is already the seventh – incredible how time flies.
Welcome! I’m Oliver Nitz from Digitale Mediensysteme GmbH. We’ve been active in retail for 19 years and work closely with retail companies, banks, corporations, and chain stores – expanding their spaces with digital technologies: above all digital signage, in-store radio, frequency measurement, and other topics.
We do digital with purpose.

You can join the discussion in the chat today: please enter your name first, then your question.
All microphones are muted; so please contribute via chat. We’re happy to address your questions.
The talk is being recorded and will be made available online afterwards so that people who don’t have time this evening can watch it. I’ll send the link to the recording to everyone by email – feel free to share it.

Today’s topic: demand planning – and also demand optimization.

Please welcome Erik Weisz, who literally rushed in two minutes ago – CEO and founder of Circly GmbH. He’ll introduce himself and the company in a moment. The firm has won multiple awards, including from CASH, Handelsblatt, and other rankings – the list is long. So: he must be doing something right.

Today we’ll learn from him how demand planning works, for example at Nah&Frisch and other brands.

Oliver: Hello, Erik!
Erik Weisz: Hi, good to see you.
Oliver: Tell us – what exactly do you do?

Erik Weisz (Circly GmbH):
In short: we use artificial intelligence to forecast demand quantities – meaning which product I will need, in what quantity, and when, e.g., on the shelf in retail.

Before we came to market, this was often complex in practice: demand-planning tools have existed for years, but many still worked with Excel spreadsheets, averages, or regression analyses – relatively simple equations that explain relationships only to a limited extent. Accuracy was rare.

I’m a serial entrepreneur. In my first company we learned that independent shop owners and restaurateurs struggle with planning – and when they do it, it’s often in Excel. You really noticed it during the pandemic: think of the run on toilet paper. You could see how irrational demand can be. With better planning, from the second lockdown at the latest – certainly by the third – you could have controlled things much more precisely. That’s exactly what we did.

Oliver: With what – and how exactly?
Erik:
You need 2–3 years of sales data and the master data (product information) to reference the history. We also annotate the data – e.g., lockdown periods – so that a “data sandwich” is created:

• Internal data (sales, item master),
• plus exogenous features such as weather, paydays, inflation, promotions (own & competitors’), and, where relevant, special factors (for a drugstore, for instance, birth rates).

This allowed us, in the third lockdown, to recognize patterns from the previous two and make forecasts – and to warn retailers early: “Something is happening here – increase your stock.”

Oliver: How much more accurate is it?
Erik:
Across all customers (evaluation over nine months, all product groups), we were just under 90% forecast accuracy.
For context: many shop owners working with averages achieve around 60–65%.
Of course there are fast movers (very accurate) and slow movers (e.g., white shoe polish) with fewer data points – we use different models there.

An example: we work with the wholesaler KASTNER and 156 Nah&Frisch locations. Many shop owners used to plan by gut feeling; the top 15–20% looked at prior years and took averages. But Easter isn’t the same date every year. We integrate our models directly into the IT infrastructure – shop owners scan the EAN code on a mobile device and immediately see sales recommendations.

Important: more features are not automatically better. We let up to 15 factors run, but the system weights them and ultimately uses the most important ~5 – the rest drop out. This avoids error propagation.

Oliver: Why Circly – and not the “big” providers?
Erik:
Many standard “AutoML” offerings sound good, but often end up producing straight lines (averages) – too crude for production/merchandising.
The large tier-1 suites are excellent, but expensive, project-heavy, and resource-intensive (training, dedicated team).
We specialize in production & retail and have ten base models – chosen by horizon (short-term peaks to long-term trends) and data situation (sparse vs. rich). Bananas need a different network than white shoe polish.

We’re quick to integrate (our fastest customer in 8 hours, the slowest in 2.5 days) and price-wise we’re at about 30% of what large competitors charge.

Oliver: How did you get into this?
Erik:
Through a hackathon. I’m a business lawyer by training and was actually looking for a developer – and there we discussed socially relevant AI applications. From gastronomy I knew the planning problem.
That’s where I met my co-founder. Our first pilot customer was Austrian Post (staff/package planning). From two models, we now have ten – all based on open-source research but developed in-house and specialized in demand forecasting.

Our data runs – whenever possible – in Austria (server housing, green energy), with failover to Germany/Helsinki. We do not use US hyperscalers unless a customer explicitly requests it. Data protection is a central point.

Oliver: On your website I read −25% returns, >90% accuracy, −80% inventory (tied capital). Is that correct?
Erik:
Yes, those are real customer results (varying by industry and assortment). In one case, we estimated around 90 tons of CO₂ saved – through fewer write-offs, less overproduction, and optimized inventory.

Oliver: What happens if I have two branches and call you?
Erik:
First a quick check (do you have the data? retail vs. wholesale? systems?). If it fits, we schedule a meeting, share a data list (columns), and get initial data as CSV/export or via API.
We ingest, train, and deliver the first forecasts – fed back into your system or as an export. Nightly runs update the suggestions. Most see usable results within a week.

Oliver: And data sharing between Nah&Frisch locations?
Erik:
No – no data sharing between customers or branches. Models can learn from each other at a meta level (e.g., model weighting), but raw data remains strictly separated.

Oliver: Shelf sensors, Amazon Go, etc. – is that the future?
Erik:
Not in the short term. Compute costs are high, margins are low – many Go-stores have been closed again. More sensible right now is a “semi-closed” merchandise management system: wholesale vs. retail outflows, dynamic safety stock (replenishment lead time), cash-flow optimization rather than full sensorization.
What’s exciting are price elasticities and dynamic pricing via electronic shelf labels – with the goal of lower markdowns, higher margins, and less waste.

Oliver: Sounds like digital signage at the POS: display time- and discount-appropriate messages?
Erik:
Exactly. If we see that product X is behind plan, you can trigger campaign-driven notices at the entrance or on the shelf – linked to the forecasts. That’s more efficient than slapping on blanket 50% stickers shortly before closing.

‍