Automate the standard, keep the exceptions human: 9 takeaways on AI in supply chain and logistics

9 takeaways on AI in supply chain and logistics

1. Adoption is a leadership task. It starts with AI literacy at C-level 

The clearest pattern of the evening had nothing to do with technology. The organisations making progress are the ones where leadership understands and uses the tools themselves, acts as a visible example, and creates safe space to experiment. 

Carola Appel described the approach at Carl Kühne KG, a company with 300 years of history and, as she put it, data of a similar vintage. The programme started two years ago with two people and some external support. The C-level team went through AI deep-dive training themselves as participants rather than sponsors. Today around 70 employees have volunteered as AI pioneers alongside their day jobs, 35 creative use cases have been prioritised, the first are in implementation, and the company is now building a dedicated organisation for it, deliberately named Data, AI & Value Orchestration, because the point is value rather than activity. 

Two details are worth noting. The pioneers span all age groups, not just the young talent segment. And the framing throughout was curiosity and playfulness rather than pressure, which, as several participants pointed out, is precisely what makes people willing to surface problems leadership cannot see from the top. 

2. Not everything should be automated. Business impact decides 

Knut Alicke made the case against automation as an end in itself. Manual processes survive where there are exceptions, where context is required, and where the cost of a solution outweighs the value it creates. 

The small franchise partner running a single store was the recurring example. Equipping them with licences, hardware and an interface to place an order is not a sensible investment. But the receiving end of that process can be automated regardless: the partner keeps calling, and the processing behind the call becomes digital. The same logic applies in the warehouse, where a three-year-old still outperforms a picking robot on badly defined items, and where automated truck loading remains a genuinely hard problem. 

This is the discipline that separates AI programmes that scale from those that stall: matching the complexity of the solution to the actual economics of the case. 

3. The next step is exception-based, not human-free 

Tchibo's automated document checking illustrates the target picture. With most goods arriving from overseas, the verification of invoices, bills of lading and related documents had occupied a substantial number of people. Following the introduction of a transport management system for coffee shipments, the team built an internal AI solution that compares documents against the expected values, flags where they diverge, and explains why. 

The result is not the removal of people from the process. It is the reallocation of their attention: staff now work exception-based, reviewing the cases that actually require judgment. As Alida Tiemann put it, that saves a great deal of time, and, as the discussion noted, produces more accurate outcomes with the same team. 

The principle generalises. Automate the standard case, surface the deviation with its reason, and keep human judgment where context and experience decide the outcome. 

4. Many automation problems are really standardisation problems 

Digital delivery notes were the evening's most-cited example of a process that should have been solved by now and has not been. 

Tchibo digitalised delivery notes for its own stores ten years ago. For its depot partners, around 10,000 of them, stacks of paper are still printed at the forwarder and handed to the driver every week. Jack Wolfskin faces the same problem from a different angle: a dealer structure ranging from single-store franchise partners on the Baltic coast to Amazon, Zalando and Sportscheck, with an equally wide range on the inbound side, where orders still arrive by phone and email. 

A participant from the industry side made the underlying point sharply. Internally, a company can mandate its own standard. Externally, every transaction involves at least two and often four parties, and no individual player, not even a very large one, can impose a format on the rest. The technology for digital delivery notes exists and has for years; what is missing is critical mass. His counter-example was a South African farmers' market a decade ago, where every stallholder used the same payment app with a laminated QR code and near-zero investment, because everyone had converged on one standard. 

Shared data and process standards are not a side issue in end-to-end digitalisation. They are frequently the binding constraint. 

5. AI is creating capabilities that were not previously viable 

The most interesting opportunity is not making an existing process 20% faster. It is doing things that were not economically or technically accessible at all. 

One participant described driver analyses that were previously impossible: linking hundreds of thousands of articles to identify the most profitable product, with the most profitable customer, from the most profitable supplier. Carola Appel reported the same effect at Kühne: data sets that had never been connected, including external market data alongside internal figures, now feeding sales analysis agents that answer questions the organisation had been circling for years. 

The same shift applies to forecasting. Dennis Trautmann described a store in Munich being overrun during an AC/DC concert at the Olympiastadion in heavy rain, with no rain jackets left. The signals, weather, event, footfall, all existed; nothing combined them. As Knut Alicke observed, the value of combining them is not a perfect number but an early indication that something is about to happen, which an experienced planner can then act on. 

6. Concrete use cases are already delivering measurable operational value 

Alongside document checking, Tchibo has built an AI-based returns forecast for its Czech returns site, with a roughly 10% error rate, which determines whether returned articles go back to stores or straight into the online shop. In a business that launches a new theme world every week, speed of re-availability is directly commercial. 

Jack Wolfskin started applying AI to customs classification four years ago. The domain is unexpectedly intricate: whether a polo shirt buttons left or right determines whether it is classified as menswear or womenswear, with different tariff numbers; shoes are classified by material composition, where 20% versus 21% changes the outcome. Colleagues previously cut products open on their desks to determine it. Today, product data flows into the classification system and the large majority of items are classified automatically and written back into the ERP. Invoice auditing followed: carrier invoice data, which is anything but standardised across German forwarders, is now normalised into a single monitoring system and audited without manual intervention, raising accuracy on parcel volumes that could never be checked line by line. 

The common thread is the digitalisation of expert knowledge. In each case, the capability previously sat with one or two individuals, and the dependency on those individuals was itself the risk. 

7. Forecasting improves. AI does not remove randomness 

Kai Hoberg offered a useful corrective to the belief that AI resolves all forecasting problems. Some things will not be predicted well, and it is worth being precise about which. 

Highly granular demand, a specific jacket, in a specific size, in a specific colour, in a specific store, over the next six weeks, involves small numbers and a large random component. The same applies to spare parts demand for individual service technicians across a six-figure parts catalogue, where no one knows which machine will fail. Aggregated quantities are considerably more tractable, particularly when the relevant influencing factors are included. 

The practical conclusion is to reset the ambition. The goal is not the perfect forecast but earlier signals, better decisions and faster reaction. 

8. Change and adoption are harder than the technology 

This was the point on which the room was most unanimous. Several participants described the decisive factor as belief in the project, sustained long enough through implementation to get the feedback needed for fine-tuning. In a traditionally structured, less IT-affine organisation, that is a significant burden, and no tool compensates for its absence. 

Dennis Trautmann added the piece that is often left unsaid: many employees assume that AI arriving means they are leaving. That assumption is a natural brake on any willingness to experiment. The counter-message has to be explicit: your day-to-day tasks become more efficient, and we want your time for things more valuable than screening newsletters every morning. Where that message is credible, the appetite to learn follows on its own. 

Structures help it scale. Tchibo has established a monthly community of practice in logistics where teams present use cases to each other; Kühne's pioneers meet every four weeks, present cases and are recognised for them; another participant described an AI Forge initiative pairing nominated people from each plant with external support, which grew to 30 or 40 participants and, more importantly, changed what people imagined was possible. 

9. Access to technology is itself becoming a competitive factor 

The sharpest observation of the evening concerned tooling. In OMMAX's own workshop formats across the industry, effectively all participants report that they do not have capable AI tools available in their corporate environment, which means they also have no basis for judging what would change if they did. 

The consequences run in two directions at once. Innovation slows, and people route around the restriction. Knut Alicke described the resulting shadow IT as the real risk: employees whose corporate environment offers them nothing simply use consumer tools privately, with company data. Several participants noted that rolling out Copilot and declaring the AI question answered leaves the same gap, because it sets expectations at the level of a chat assistant rather than showing what an agentic workflow or a self-built internal tool can do. 

The resolution is not prohibition. It is IT as enabler and governance partner rather than bottleneck: a secure sandbox, clear rules, and permission to try. Shadow AI is the flip side of missing enablement structures.