Zauber

𝗧𝗵𝗲 𝗾𝘂𝗼𝘁𝗲-𝘁𝗼-𝗯𝗼𝗼𝗸𝗶𝗻𝗴 𝗮𝘂𝘁𝗼𝗺𝗮𝘁𝗶𝗼𝗻 𝗺𝘆𝘁𝗵 📦 "We just take our existing quote from TMS or QMS and turn it into a file." I hear this in some conversation about file creation automation. Technically it is correct. Operationally it is misleading, and the gap is where the actual work happen. Take a QMS spot quote example: It contains for instance, two containers at price X USD, various line items plus some routing information. What is missing is various additional shipping details that happen to exist only later in the process. So you get roughly 𝟮𝟬 𝘁𝗼 𝟯𝟬%, best case around 50% in our experience. Which is exactly why a TMS quote usually produces a booking or equivalent and not a shipment. 𝗪𝗵𝗮𝘁 𝗶𝘀 𝗺𝗶𝘀𝘀𝗶𝗻𝗴 𝗶𝘀 𝘁𝗵𝗲 𝗽𝗮𝗿𝘁 𝘁𝗵𝗮𝘁 𝘁𝗮𝗸𝗲𝘀 𝘁𝗶𝗺𝗲. Full addresses for all parties, shipper and consignee, weights and volumes, which on LCL and air almost always differ from what was quoted, commodities, products, quantities, release types and counterpart details. It gets collected by email from the shipper or consignee, reconciled to the shipping instructions, and entered by hand. Many point automations claim to create the file from the quotation. In practise, it easily leaves 10-15 minutes additional work lot manually, sitting between email communication, TMS and document handling. We think this broader and call it 𝘀𝗵𝗶𝗽𝗺𝗲𝗻𝘁 𝗰𝗿𝗲𝗮𝘁𝗶𝗼𝗻 𝗮𝗻𝗱 𝗰𝗼𝗺𝗽𝗹𝗲𝘁𝗶𝗼𝗻, including the SOPs that sit behind it. The difference is between a record that exists and a shipment operations can actually work with. As an outlook on the practical implications: AI point solutions will lose relevance beyond the initial deployment, when you move more towards end2end automation, because they lack context or the actual history. If you have the email history on the booking, you complete the shipment better. If you also run pickup and delivery, you already have the right accruals to run AP or AR invoicing much smoother with an AI agent. The value compounds end to end, and a single-step tool cannot get there. If you are in charge of operations or IT at a forwarder, the question worth asking your team: of the data points needed to create a shipment in your TMS, how many actually arrive with the quote? Happy to share what we see across live deployments, just reach out.

This week we hosted the second edition of the Zauber AI Dinner, this time in Frankfurt. Everyone around that table is somewhere in the middle of adopting AI, and we appreciate how openly forwarding leaders exchanged experiences and insights, on what delivers and where they are stuck. With development moving as fast as it currently does, shared learning is worth a lot. The last guests left after midnight, so we take it the second edition landed well. Thank you to everyone who joined, and a special thanks to Dr. Florian Grillo-Bartels and Tobias Wölfel from McKinsey & Company for their impulse on AI in logistics. #FreightForwarding #AI #Logistics #Zauber

Every stellar candidate already has attractive options, which puts the burden on us to show that our problem is worth their time. That is how this evening is set up. The companies present, the room asks the questions. Together with Olivia & zoran from five degrees and with our friends from Point Nine we are hosting the Berlin edition of their Founder Showcase. A hand-picked group of engineers and operators meeting some of Europe's fastest-growing companies. Super honored to be part of this with Zauber and looking forward to showing the room what we’ve been building. Spots are limited, hope to see you there! (link in the comments)

GPT-6 Astra, c𝗼𝗺𝗽𝘂𝘁𝗲𝗿 𝘂𝘀𝗲 🖥️ and AI forwarding automation 🚢 OpenAI released GPT-6 Astra just before the weekend and we were among the first to test it. The capability I find most exciting is the degree of interactivity: the model can operate a computer and work through a browser the way a person does swiftly. On OSWorld 2.0 (a computer use AI benchmark) it scores 72.6% against 65.7% for its predecessor, at roughly half the time per task. Why does that matter for AI-powered operations in logistics? The majority of time is spent on e-mail and TMS, so this is where primary agentic AI automation happens, but 10 𝘁𝗼 𝟮0% 𝗼𝗳 𝗼𝗽𝗲𝗿𝗮𝘁𝗼𝗿 𝘁𝗶𝗺𝗲 𝗮𝗹𝗼𝗻𝗴 𝗮 𝘀𝗵𝗶𝗽𝗺𝗲𝗻𝘁 𝗶𝘀 also 𝘀𝗽𝗲𝗻𝘁 𝗼𝗻 𝘄𝗲𝗯𝘀𝗶𝘁𝗲𝘀 or computer applications. For instance, checking milestones on HHLA or LBCT terminal websites every other morning, confirming or placing bookings in local portals or uploading documents into customer interfaces is part of the job today. A closer look at milestone monitoring For years the industry argued about whether visibility solutions are worth the money. Tracking is a market standard. Yet, operators still open the terminal site, because of data quality and/or latency challenges. Worse, where an API writes back into the TMS, the next push overwrites the ETA the operator just corrected. This can easily mean mathematically more than three full-time people doing nothing but looking things up. In the end it is a question of agent efficiency for this use case 𝟭. Double costs pre-AI: once for the tracking API, once for the human doing the checks again. That is how visibility spend turns into a cost that is simply there and no longer questioned. 𝟮. Seize computer use selectively: Computer use is is token intensive in any case, and Astra lists at 2.5x the per-token price of GPT-5.6 Sol. Pointing the most powerful model at a 90 second lookup cannot be the solution; therefore, wherever APIs can deliver sufficient quality and latency, use these. However, there will be gaps. Here AI can help (see 3.) 𝟯. Efficient AI logistics agents needed: Anyone playing with the latest OpenAI or Anthropic model in-house will get a computer-use agent visiting terminal websites working in an hour live (just did it myself earlier, try it with Astra impressive). The point is deploying it globally, at scale, at a quality and security level you can defend and at a cost per task in the cents range that can materially undercut the costs of humans. So, we invest in the agentic infrastructure around the models: the part that gets the task done efficiently and globally deployable. On the other hand, we do not believe it will be the widespread agentic tool of this use case, because it is  a. resource intensive on provider side, b. expensive and will likely remain so, and  c. permissions not solved (as this is designed for humans, not agents using human-designed interfaces)  Happy to share what we learned.

It's my pleasure to announce Zauber's Chief Technology Officer, Otmane "OT". When building this company, one of our core premises was to find the best people in the world at the intersection of logistics and technology. With OT, Zauber has not only won a technology leader who could cite the DCSA carrier booking data model in his sleep, but also an incredibly strong technologist who keeps pushing the boundary of AI adoption for us as a company and for our product. This has gone so far that none of us is sure any more whether we're talking to his OpenClaw Friday or to him. Or that he single-handedly spends as much money on coding agents as our customers do revenue with us (almost!). Most importantly, OT also embodies and shares our deep commitment to customer excellence and the willingness to do whatever it takes, while bringing a good laugh to the office in the morning. OT, we're looking forward to the next chapter together, even though it will mean 10x more craziness: being pulled into customer calls by Fabian, rescuing a deployment on short notice with Christopher, or sparring with Dennis on how to raise the bar even higher. Keep shipping!

“It works” is no longer a high bar for software & agents. With vibe coding, a functional agent can be built in hours or days. That is a huge step forward and something every company should use. However, the real work starts when the agent moves from a demo into a business-critical process. In freight forwarding, that means live shipments, customer communication and operations that do not stop at 6 p.m. McKinsey’s Q1 2026 B2B survey found that, depending on the use case, only 3–11% of surveyed companies had fully implemented agentic AI at scale. At the same time, roughly 80% preferred to source agentic AI externally rather than build entirely in-house. Those numbers make sense to me and match what I see every day. In-house teams can now build useful prototypes quickly. Putting them into production however adds a much longer list of requirements: 1. 𝗕𝘂𝘀𝗶𝗻𝗲𝘀𝘀 𝘃𝗮𝗹𝘂𝗲 & 𝗥𝗢𝗜: Cheaper prototypes make it tempting to postpone the business-case discussion. At scale, model, integration, maintenance and exception-handling costs add up quickly. Strong ROI backing remains essential so that AI actually moves your company's P&L. "What to build" actually becomes more important as building gets cheaper. 2. 𝗥𝗲𝗹𝗶𝗮𝗯𝗶𝗹𝗶𝘁𝘆: Queues and retries, fallbacks and redundancy when models hit rate limits or connected systems fail become a must. Operations need to keep running, including outside business hours, and require concepts such as engineer-on-call and automatic health monitoring. 3. 𝗚𝗼𝘃𝗲𝗿𝗻𝗮𝗻𝗰𝗲 & 𝗼𝗯𝘀𝗲𝗿𝘃𝗮𝗯𝗶𝗹𝗶𝘁𝘆: Teams need traces, evaluations, alerts, audit logs, approvals and human handovers. Every action needs to remain visible, explainable and controllable. 4. 𝗦𝗲𝗰𝘂𝗿𝗶𝘁𝘆 & 𝗰𝗼𝗺𝗽𝗹𝗶𝗮𝗻𝗰𝗲: Enterprises need clarity and control over hosting locations, subprocessor management, permissions, retention and tenant isolation. These are just non-negotiable requirements when agents handle sensitive operational data. 5. 𝗠𝗮𝗶𝗻𝘁𝗮𝗶𝗻𝗮𝗯𝗶𝗹𝗶𝘁𝘆: Workflows need versioning, testing and rollback as models, APIs and processes change. Otherwise, every model upgrade becomes a production risk, and keeping up with the continuous developments in AI becomes a burden. 6. 𝗜𝗻𝘁𝗲𝗴𝗿𝗮𝘁𝗶𝗼𝗻𝘀: Agents need to handle operational exceptions, in our case across Outlook, TMS, RMS and other custom systems. Integrating the happy path is relatively easy, but only integrations that provide feature parity with the relevant actions available to end users give agents the capabilities required for reliable daily operations At Zauber, we focus on proven logistics use cases that can materially reduce cost-to-serve and improve service quality. We offer ready-to-run enterprise agents, that are easy to customize, in a governed, scalable environment. That is our bar at Zauber: reliable in daily operations, with economics that hold at full volume.