The question that keeps coming up in AI investing and AI hiring is the same: where does defensibility actually sit? A better model can be trained with enough compute. Institutional memory requires time, customers, and stakes — things money can’t shortcut.
There’s a useful framework that maps to five archetypes of AI-era businesses. Each survives model commoditization for a different reason. Understanding them changes how you evaluate AI companies, AI career moves, and AI transformation projects.
The first archetype is Knowledge Compounders. These are businesses whose data improves through usage. Medal.tv accumulates two million daily gaming clips that become training data. Mercor applies human-in-the-loop verification that creates ground truth the models couldn’t generate alone. The flywheel: more users generate more data, better data improves the product, better products attract more users. You can’t buy your way into this position — you have to earn it through accumulated interaction. Banks, interestingly, sit on decades of proprietary transaction data that no AI lab can replicate. Whether they can operationalize it is a different question.
The second is Workflow Commons. These are shared templates that capture how work actually happens inside organizations. n8n built a library of seven thousand community-contributed workflows — automation patterns that encode real operational knowledge. ComfyUI, Roboflow. The moat isn’t the software itself; it’s that the accumulated workflow library becomes the thing organizations depend on and can’t easily migrate off. The institutional knowledge is in the templates, not the code.
Third: Reality’s Gatekeepers. These sit between AI and real-world consequences. Stripe, Plaid, Deel, Column. They’re toll booths: small fees on massive volume, with positions that AI can’t disintermediate because they’re the trust layer the AI depends on to touch the real world. There’s a cautionary tale embedded in the 2024 Synapse collapse — an intermediary that sat between fintech apps and banks. When it failed, $85M in customer funds got stuck because nobody could reconstruct whose money was where. Mercury survived because it had migrated to Column, a bank built to serve fintechs directly. The company that owned the rails won; the middleman imploded.
The fourth archetype is Marketplaces. Trading floors for AI-era participants — OpenRouter routing LLM traffic, Composio integrating tool calls. These have a real risk baked in: they’re bets on fragmentation. If models converge to two or three dominant providers, the value of routing between them collapses. Marketplace moats require that the problem they solve remains interesting.
Fifth: Vertical Transformers. These replace entire workflows in regulated industries — Harvey in legal, Sierra in customer success, Rogo in financial research. The value isn’t purely the AI capability; it’s that they’ve taken on the compliance and liability that comes with operating in the domain. Same output quality at a fraction of the cost, but with accountability built in. The model is commodity; the domain wrapper is the business.
What makes this framework useful is that it forces a specific question: when the underlying model improves, does the business position strengthen or weaken? Knowledge Compounders get better as the models that process their data improve. Reality’s Gatekeepers become more valuable as more AI flows through them. Marketplaces get more volume as AI adoption grows, but face convergence risk. Vertical Transformers can be disrupted if a frontier lab decides to build the vertical themselves.
For the financial services AI transformation specifically, the most interesting archetype is the third — Reality’s Gatekeepers. Banks already own the rails. The open question is whether they’ll realize that their existing infrastructure, designed for human callers, is the bottleneck, and whether they’ll rebuild it for an AI-first world fast enough to matter.
P.S. The original framing is Tina He’s, from a piece on boring businesses that will dominate the AI era. I’ve added the banking application layer. The insight that landed hardest: own the rails, don’t rent them.