Vertical Knowledge First, AI Second
There is a common way for AI deployments to die: the deck looks great, the system ships, and then nobody uses it. The post-mortem blames the model, or the staff. But most of the time the root cause sits earlier — the people who designed the solution never actually understood the industry.
My own path ran in the opposite direction. Logistics management degree, years spent living in incoterms, HS codes and clearance desks first, then a turn to AI in 2021 — four shipped product lines since. That order defines how we build: every engagement starts by understanding how the workflow really runs, before deciding where AI belongs. And where it doesn’t.
Why the order matters:
Industry knowledge is wildly underestimated. A model person can learn the surface of HS classification in two weeks. They cannot learn which product descriptions trigger classification disputes, or which categories carry historical landmines at declaration. That knowledge lives on the field, not in documentation.
AI amplifies your understanding — and your ignorance. Models are compliant: ask the wrong question and it will confidently hand you a wrong answer. In customs, that means wrong HS codes, wrong duties, sometimes penalties. An AI tool built by someone who doesn’t know the field industrializes error.
The floor doesn’t trust outsiders. The last mile of deployment is human. Operators only fold tools into their own workflow when they feel the builder speaks their language. That’s not attitude — that’s trust.
So we gave ourselves a rule: the model is never allowed to invent classifications. Our customs tool Urgenth reads a product description, queries the official HTS database, and selects only from real candidate codes — inventing a code is explicitly forbidden. That constraint isn’t a technical compromise. It’s what industry knowledge looks like when written into product architecture.
Vertical knowledge first, AI second. That order is the whole company.
DeepTill is a forward deployed engineering company: engineers in your field, shipping AI that runs inside real workflows.