Owning Workflows, Not Models
Kavin Stewart, Partner at Tribe Capital, on Reddit's 10x opportunity
The real moat in AI apps is moving up from model wrapper to workflow owner. If a company only adds a chat box on top of one model API, the model vendor can copy the feature, raise prices, or deprecate the interface. The safer position is owning the surrounding job, like search grounding, file retrieval, memory, approvals, and distribution, because those pieces are harder for any one model provider to yank away overnight.
-
This is the same kind of dependency risk that showed up in earlier platform waves. Facebook game studios could grow fast on borrowed distribution, but the platform owner controlled ranking, invites, and access. In AI, the choke point is model APIs, context limits, tool access, and pricing.
-
The interview points to why that risk looked temporary even in 2023. Google already had a major structural asset in search indexing, and Anthropic was expanding context windows quickly. Today, Gemini offers Google Search grounding as a native tool, and Claude offers very large context windows and memory features, so developers have more credible substitutes.
-
That shift changes what gets funded. Thin wrappers are easiest to replace. Products that plug into a company’s documents, approvals, CRM, codebase, or customer support queue are stickier, because ripping them out means retraining workflows, not just swapping one model endpoint for another.
The market is heading toward multi model, tool rich AI stacks where the winning apps look less like standalone chatbots and more like software that quietly runs real work. As model capabilities converge and built in search, memory, and retrieval keep improving, durable value will sit in distribution, proprietary data, and workflow integration rather than the underlying model alone.