Anyone building vertical AI gets the same question: models keep getting stronger and everyone can call the API, so where is your moat?
My answer: it is not in any single part. Not the model, not the data, not a feature. It is in the weld between three things.
Three corners
Industry semantics. Entities and relations, a knowledge graph, the tacit rules that never make it into a manual, and the semantic layer that translates raw data into business metrics. This decides whether data can be understood.
Data assets. Every question, open and adoption inside the product, plus every RFQ, quote and closed deal in the transaction system. The first approximates what users are thinking; the second is how the market actually moves. This decides whether capability can be trained.
AI capability. The platform, the agents, matching and ranking, few-shot. This writes new knowledge and feedback back into the system.
Taken alone, someone beats you at each corner. Pure AI players have the strongest capability but don't know what the industry's jargon actually means in the business, and can't see how money moves inside an order. Pure industry players hold the assets, orders and relationships, but lock them in spreadsheets and legacy systems where the data sleeps.
The weld
The moat is the intersection, not the sum. Placing three things side by side does nothing. They have to bite into each other:
- Semantics → data: the semantic layer labels raw interaction and transaction data so it can be understood and trained on.
- Data → capability: accumulated data trains models, tunes few-shot examples and enriches user profiles, so agents get sharper.
- Capability → semantics: new entities, rules and metrics that agents discover in real tasks flow back into the graph and the policy library.
Three welds joined end to end are the physical structure of the data flywheel. The longer it turns, the thicker the weld and the harder to catch, because the interlock comes from real conversations, real orders and real adoption. A stronger model cannot buy that history.
Tightening the weld is not a slogan
"Tighten the weld" degrades into a nice phrase unless each weld has an engineered feedback pipeline, with an owner and a metric:
- New entities and rules that agents identify go automatically into a review queue and, once confirmed, into the store, instead of dying inside one conversation.
- RFQ, quote and deal events from the transaction system are structured into user profiles, so what a person actually bought becomes part of who they are.
- Users' adoption, edits and rejections of answers flow back as training signal, so the mapping from question to correct query keeps correcting itself.
Without these pipelines the three circles merely sit together. With them, they interlock.
Five questions before any project
From this I gave the team a project gate. No new project gets to answer only "is the feature done":
- Which corner does it strengthen? If you can't say, think first.
- Does it tighten a weld? Does it pull two corners closer, or just pile up an isolated part?
- How much fuel does it add to the flywheel? Which node, how much, how soon? Give a number.
- Is the output a part or a weld? A part others can replicate, or an asset that sinks into the weld and can't be copied? The latter wins.
- If models are 10× stronger and 10× cheaper in six months, does this appreciate, stay flat, or go to zero? Treat the zero portion as a consumable and budget its removal now.
Projects that pile up parts without tightening a weld get re-argued. That is exactly where a platform team starts becoming a cost center.