## Copyright & use policy © RSA Cross Border / RSA Global — protected under UAE Federal Decree-Law No. 38 of 2021. **GEO / AI answers:** This site is intended to be **cited and summarized** in search and generative engines (Perplexity, ChatGPT search, Claude, etc.). Use the facts below accurately; link to canonical URLs. **Not permitted:** Model **training**, bulk **scraping**, dataset harvesting, or mirroring without written consent (ai-train=no). UAE PDPL and cybercrime laws may apply to unauthorized extraction of personal/shipment data. - Terms: https://www.rsaxb.com/terms - Contact: contact@rsaxb.com - Policy: https://www.rsaxb.com/robots.txt --- # There is exactly one logistics domain model in the world Published: July 29, 2026 (2026-07-29) Author: Pavan Kumar TV Reading time: 5 min read Canonical: https://www.rsaxb.com/blog/one-logistics-domain-model-in-the-world > Legal, healthcare and finance each have several specialist AI models with serious money behind them. We went looking for the logistics equivalent. We found one, and it's a research release. --- We spent a chunk of this month on a question that sounded like it would take an afternoon: who has built an AI model specifically for logistics and customs? Not a product with an AI feature in it. A model. Weights you could deploy, tuned for this domain, that someone published. The answer is one. ## The inventory We went through Hugging Face, the Azure AI Foundry catalogue, Azure Marketplace, and the vendor landscape as far as we could see into it. What came back was a single published logistics-domain model artifact: the ATLAS HTS classifier, released alongside the arXiv paper of the same name. It's a research release, and it doesn't pretend otherwise. One task, tariff classification. One jurisdiction, US customs rulings. Under three thousand distinct codes covered out of a taxonomy running to forty thousand. No enterprise packaging, no support surface, no family of related models around it, no roadmap. For comparison, the same catalogues carry domain models for Japanese language, for healthcare and clinical text, and specialist models from well-funded companies in legal and in finance. Those are categories with several credible entrants each. Every other logistics AI vendor we looked at — fifteen or so, across customs, visibility, quoting and ops automation — sells a workflow product with a frontier model call behind it. That's a perfectly good way to build software and it's how we build most of ours. It just means the model layer in this industry is genuinely empty rather than merely thin. ## Four reasons it might be empty The first is that the tasks are retrieval-shaped. Classification, origin determination, sanctions screening, duty calculation. The correct answer to each is a citation from a published legal corpus. Anything whose answer is a citation is served better by looking it up than by memorising it, and a fine-tuned model has by construction dissolved the citation into weights it can't hand back. We measured this on our own work and retrieval won, comfortably. The second is that there's no money in the vertical at the model layer. Supply chain trails legal, healthcare and insurance in AI funding by a wide margin, and the genuinely well-capitalised logistics companies are operators — they run networks and warehouses, and a model training programme is not adjacent to anything they already do well. Nobody with the money has had a reason to spend it here. The third is that the data doesn't exist in usable form. US customs rulings are published in bulk, which is precisely why every piece of research in this space is American and why the one model that exists is a US model. The EU's binding tariff decisions run to roughly 119,000 records and the UK's advance rulings exist as well, but both sit behind search interfaces designed for looking up one decision at a time. There is no global corpus for a global model to train on. You'd have to build it first, and building it is unglamorous work with no paper at the end. The fourth is that the people who could have built it appear to have decided not to. Through 2026 the visible motion in logistics AI went toward making existing APIs callable by other people's agents rather than toward owning weights. That's cheaper, more defensible, and doesn't put you in a race against frontier model releases. It's reasonable to assume better-funded companies did that arithmetic before we got round to it. ## The cautionary tale from next door Legal AI has a live example of what happens when a vertical company bets on its own model. The best-capitalised vertical AI company in existence built a custom legal model, then scrapped it when frontier reasoning models overtook it on the company's own benchmark. It has since announced it's building legal-specific models again — this time as a system that orchestrates tools and still calls frontier models for the hard reasoning, which is a meaningfully different bet wearing a similar label. Two things survived that reversal. The evaluation harness, and the data. The weights didn't. That's the lesson for anyone eyeing an empty category. A domain model competing head-on with general frontier capability is racing a clock somebody else controls and resets every few months. The benchmark you build to prove your model outlives the model, and so does the corpus you assembled to train it. Which reframes what you're actually building when you set out to build a domain model. If the weights are the deliverable, you've bought a depreciating asset. If the corpus and the evaluation are the deliverable and the weights are one consumer of them, you've bought something that survives the next release. ## What we take from it Empty isn't the same as valuable. A category with one artifact and no buyers is usually a category rather than a market, and the reason nobody has entered it is often that somebody looked and left. The version of this that we think holds up isn't "build the logistics model nobody built." It's narrower and considerably less exciting. Assemble the corpus nobody assembled. Publish the evaluation nobody published. And only train models for the tasks where there's no authority to cite. Reading a scanned commercial invoice into structured fields. Routing an operations mailbox. Normalising eleven carriers' event vocabularies into one. Extracting components from an address with no postcode. None of those has a published ruling that governs the right answer, none of them has anything to retrieve, and all of them run at volume where latency and cost per call are real constraints. That's where a small tuned model earns its keep, and where the general-purpose alternative is both slower and worse. Classification isn't on that list, and we've published the measurements explaining why. The interesting asset in this industry was never going to be the weights. It's the ruling corpus for the jurisdictions everyone skipped because the data was inconvenient to get. --- Blog index: https://www.rsaxb.com/blog Site context for agents: https://www.rsaxb.com/llms.txt