Robert Ta argues that most organizations fail with AI because they lack governance and standardization. Fewer than 10% of companies see sustained value from AI deployments, and 60% have no formal security policies. Leading firms like Stripe and Uber succeed by building unified infrastructure layers that route tasks to appropriate models and control costs.
If you're an AI builder or tech executive, you're probably looking at AI adoption. And what we're finding is AI adoption without standardization creates more problems than it solves. And every engineering org is really adopting AI tools. Few of them have answered the structural questions that determine if it compounds or if it just creates chaos. This points to a study done by BCG. Less than 10% of orgs can point to a GenAI deployment delivering measurable sustained value at scale. 56% of CEOs report AI produced neither revenue growth nor cost reduction in the past 12 months, and 60% of orgs have no formal AI security policy governing employee usage of public LLM APIs. So your IP might just be leaking out the door. The organizations pulling ahead— Stripe, Uber, DoorDash— at least to our research based on the public information out there, they do one thing really well. They invest in infrastructure around their harnesses and a knowledge layer or context layer. Those are very different things than just worrying about who foots the bill for my tokens. A few things here from those companies, but they've essentially seen this emerging pattern out from public information for enterprise architecture, where if you're a developer with Cloud Code or with Cursor, your CI pipeline might call for adversarial code reviews or whatever, or you have an internal app, a product feature that calls an LLM model via API. What we're seeing is these enterprises are putting in the screws and tightening them around all model calls go through one abstracted layer. Where we can inject rate limiting, model routing, cost tracking, et cetera. So the whole idea is if I have 100% of work needing to be done, maybe 20% of that is just some unit tests. Maybe I throw that into a self-hosted open source model like DeepSeq Ollama. Maybe for planning or orchestration across many repos in a highly subjective, complex domain set of problems, I have a frontier intelligence model like Opus or Fable-5 or, you know, in the future Mythos, whatever comes out. Maybe that does the orchestration and the planning. And then maybe I have specialized models, maybe I create them for my own super unique case. But either way, we're taking the casino out of the— you spend your tokens and we're saying, let's put something there that optimizes what you get out of the casino.
