Five positions I'd defend in a room. The rest of this site is downstream of these.
These are the five operating beliefs that shape how Shrey builds OsmoSync, writes the newsletter, and picks what to ship next. They exist on this page because the rest of the site is downstream of them. Every project, essay, and decision pressure-tests against these positions first.
Companies that bolt AI onto a stack designed for humans get the worst of both. The ticket queue still routes the way it did in 2019, the data still lives in seven places, and the model is asked to wave its hands at the gap. The compounding returns go to teams that rebuild the foundation around AI from the start. That means systems of record that agents can read and write, observability that treats a model call as a first-class event, and pricing that pays for outcomes the model produces rather than seats it sits inside.
Vendors who hold your data, your prompts, your agent definitions, and your evals own your product. When their model updates, your product changes. When their pricing tilts, your margin tilts. Operators ship inside the customer's stack, their data, their identity provider, their observability, because that is the only configuration where a customer is actually buying their own software. The rest is a service contract dressed up as a tool.
Most teams optimize for shipping the next thing. The rare ones optimize for the foundation that lets every next thing ship faster than the last. The choice between those two paths shows up nowhere on a roadmap, but it sets the slope of everything that follows. Foundations are unsexy: a real test bench, a real deployment story, a real way to roll back. Companies that pay for them in year one are the ones that look unreasonable in year three.
Seat-based and usage-based pricing reward vendor activity, not customer results. The future of B2B AI is outcome-priced because it has to be. Models will keep getting cheaper to run and faster to serve, which means usage curves go down even as value goes up. Anyone whose pricing is married to API tokens is signing up for a margin compression they cannot price out of. The companies that survive will charge for the meeting booked, the lead qualified, the ticket resolved.
The interesting decisions in AI right now are made by builders inside companies, not by buyers running RFPs. The person who knows whether a system is real is three layers below the one who signs the contract. I write and build for that person. They're allergic to slideware, generous with their time when the work is good, and the only audience that compounds. If your distribution depends on procurement liking you, your distribution is borrowed.
If you read this and disagree, I'd rather hear it than not.
Last updated: May 2026
AI OS by Shrey. One essay a week, written in the open.