On 5 November we attended the KI_TAGE at the MMZ in Halle (Saale), a conference right on our own doorstep. Eight talks were on the programme, and the day was hosted by Michael Carl and Mathias Harrassowitz from Carls Zukunft. New tools were hardly the focus. Almost everyone talked about adoption within the company and about what ultimately pays off. Let's get started, here is our take.
Who was on stage
The keynote was given by Wolf Brüning, who has worked at OTTO for 15 years and leads UX design there. He spoke about the democratisation of software development and was the declared optimist of the day. Stefan Groschupf joined via video from San Francisco. Born in Halle, he founded Datameer and now runs Centrum AI. Speaking for the city was Lord Mayor Dr. Alexander Vogt. Jacqueline Köppen and Yannick Büchle came from Hypoport hub SE, Dr. Stefan Hennig from ACP and Eric Tessenow from Kieback & Peter. Johannes Kanthak and Andreas Kerndt reported on their client projects.
What we took away most was from Hypoport and from Johannes Kanthak.
Hypoport: 2,200 employees and every training session fully booked
Jacqueline Köppen and Yannick Büchle have been handling the introduction of AI at Hypoport for two and a half years. It started in May 2024. Back then, employees were using ChatGPT at work, which was not compatible with data protection, so IT set up its own system with HypGPT. Now there are licences for ChatGPT Enterprise, Claude and Microsoft Copilot. The accompanying training sessions were fully booked every time.
Even so, AI is used in the company less than expected. Köppen and Büchle named two reasons. Someone who has been trained returns to a team where the others have not. And with a 60-hour week, managers have no time to deal with it, precisely in the areas where AI would bring the most benefit. The AI champions in the subsidiaries changed little about this either. They were supposed to "do something with AI" and bring everyone along, on top of their actual work.
Hypoport is therefore now starting at the management level. Management chooses one of four stages: support tasks, redesign processes, hand standard cases over to AI, or reshape the business. Beforehand, management had worked every week for two months on an AI vision, which in the end consisted mainly of text. For their area, managers fill out a canvas with eight fields and present it to the group every two to four weeks. One of the fields asks about the effect on the result. The two of them accept "I don't know" as an answer there.
The programme is called SPARK. On the side, the company runs an "AI November" with 25 short talks by colleagues, all remote.
We liked that here two people, after two and a half years, said what didn't work. That rarely happens at conferences.
Johannes Kanthak and the four kinds of work
Kanthak sorts work into knowledge, judgement, relationship and manual labour. Under knowledge he includes everything that is done over and over according to known rules, such as quotes, research or document capture. In the future AI agents will handle that, and money can no longer be made with it. What remains is knowledge that is nowhere on the internet. Judgement gains value because someone has to stand behind the decision, and manual labour on changing building sites stays anyway.
He had two examples with him. A trade business has the customer's house measured by a drone. Meanwhile the tradesperson talks with the customer and types in their wishes. Along with the PDF, the customer receives a website with a 3D model of the existing state and the plan, each component clickable and with a price.
The second example was an architecture firm that needs two to five weeks for a concept. The firm anonymised the data from old projects and put it into a database; an AI agent generates the first sketches from it. It sells access to this database by licence to other architecture firms, technically running via an MCP server.
On retail, Kanthak said that the standard segment is shrinking and only premium and economy are growing. He mentioned Galeria with its repeated insolvencies. Against online retail and against AI agents that shop on behalf of customers, an assortment without a unique selling point has no chance. In our view the statement made is about right, and it is likely to affect many online shops.
In passing he mentioned what he works with himself. It is Codex and Claude Code, with four licences at 200 euros each.
The bicycle retailer and the shipping confirmation
Andreas Kerndt demonstrated a voice assistant, using invented data from a bicycle retailer. A customer calls and asks where his parcel is. He gives his order number and postcode and learns that the helmet is on its way and the rear light will follow in three working days. The link to the tracking comes by SMS. Very important: such an assistant must identify itself as a machine on the phone, something Kerndt pointed out specifically.
He calls his approach AOA, which stands for analyse, optimise, automate. Before every project he has people count for ten working days what they spend their time on. For optimising he had an example from online retail. If the shipping confirmation is worded more clearly, hardly anyone calls about the parcel any more, and then no voice assistant is needed for that question.
An example from Dr. Stefan Hennig of ACP went in the same direction. A logistics company wanted to have AI handle the appointment scheduling for furniture deliveries and found in the process that the phone calls cost it hardly anything. What was expensive was the goods standing in the warehouse without an appointment. In the end the solution was to call in the evening, when the recipients are at home.
Kerndt also told of an employee of a client who thought little of AI. The man then built a small administration application for his association and has been promoting it within the company ever since.
Who actually checked this?
Eric Tessenow is responsible for information security at Kieback & Peter. He asked the room for a show of hands twice. First he wanted to know who had last created a binding document such as a contract or a guideline with AI, and many hands went up. On the second question, who checked the document afterwards, there were considerably fewer.
Tessenow treats an AI like a new employee who is being onboarded and only gradually gets access to documents. Two of his experiences stuck with us. His AI at some point took unpublished drafts for established practice in the company; since then he strictly separates the actual state from everything that is only being tried out. And an AI that checks its own result, as expected, finds little. That is why, for him, checking runs separately from creation.
And the city of Halle?
Lord Mayor Dr. Alexander Vogt stuck to what has to come first in the administration. The electronic file with document management is to be introduced from 2027. He sees AI where incoming orders and forms are automatically sorted and forwarded to the right place. He made it quite clear that digitalisation counts as a voluntary service for municipalities and is therefore not funded.
Our recommendation to our clients:
We took a lot away from the talks. Among other things, we would pass on these two simple practical tips to our clients:
Over ten working days, keep count of which questions come in to your customer service. Delivery status and returns are probably near the top. A portion of that is likely to be resolved with a clearer shipping confirmation, which needs no AI.
Also clarify who checks what an AI has written at your company. That applies to quotes as well as to product texts.
Comments
No comments yet.
Write a comment