The Model Isn't the Moat
Decagon runs 90% of its customer-support workflow on fine-tuned open-source models. Not the frontier ones. And not to save money: "on the specific task we want them to do, they actually outperform the large smart state of the art models," one of its co-founders said on a16z last week. The company you would most expect to worship frontier models has concluded the model is not where the value lives.
That is the software story right now. The best model does not win the enterprise. The wrapper does: trust, guardrails, provable ROI, and politics. Decagon's version of the argument is that the moat is everything around the model, permissioning, testing, compliance tooling, the stuff that makes an agent deployable inside a bank. Model capability already exceeds what enterprises will permit. The constraint is trust.
ServiceNow's Neil Sparkes made the CFO version of the same point. Efficiency claims no longer clear the bar; boards want to know what line item AI replaces, benchmarked against real workflow data. His frame for agents is a hiring one, with a twist: "the difference is you can hire a 1000 of those people in an hour and get them all stood up and trained and knowing what to do." They still need onboarding, access, and guardrails. And the gap is widening between companies doing that work and the laggards: "they're not aware how much they're getting left behind."
Cerebras sells a chip it claims runs inference roughly 10x faster, and its partnerships chief Alan Chhabra spent an hour explaining why that isn't enough. Cracking the hyperscalers is "extremely political." His line: "You can't just like walk in and be like, hey, I have a faster widget. Why don't you use it?" The way in was co-architecting around AWS's own silicon rather than against it.
And Checkr's COO Lindsey Scrase gave the pure go-to-market version: enterprise deals are year-plus, career-risk trust decisions, and "just because you've scaled with that company doesn't make you enterprise ready." Her most clarifying stat about cold outbound in 2026: "I get 100 emails a week from people trying to sell me something. I don't read any of them." Checkr's highest-ROI channel is still events. Human to human.
Sources: Jesse Zhang and Ashwin Srinivas (Decagon), a16z, "How Enterprise AI Really Gets Deployed," Jul 31, 2026. Neil Sparkes (ServiceNow), The Customer Success Pro, "AI Adoption in Customer Success: What ServiceNow Is Actually Doing," Jul 29, 2026. Alan Chhabra (Cerebras), Revenue Builders, "Why Partner Leaders Must Master the AI Landscape to Win Hyperscaler Partnerships," Jul 30, 2026. Lindsey Scrase (Checkr), In Depth, "What startups get wrong about enterprise," Jul 30, 2026.
Fine-tuned open-source models already beat frontier models on narrow tasks; the durable moat is compliance tooling, CFO-grade ROI proof, and trust.