What I believe about AI execution

The models are not the bottleneck. A GCC or a service business that stands up a pilot in 2026 is usually not short of intelligence — it is short of a system around the intelligence. The typical path is: shove documents into a vector database, wrap a standard RAG pipeline around them, put a chat box on top, and call it an assistant.

What comes out is roughly 80% accurate and therefore 100% unreliable. For anything that touches contracts, money, or a customer, 80% is not a productivity gain; it is a liability with a nicer interface. The pilot stalls not because the model was wrong but because nobody built the part that decides what happens when it is.

If your value proposition is the weights of your model, you do not have a moat, you have a ticking clock. The defensible part is the proprietary data pipeline that feeds it and the specific workflow it enables — the plumbing, not the pedestal.

The same applies to autonomy. An agent you cannot leave running overnight without supervising the bill is not an agent, it is a liability. Reliable, deterministic execution beats autonomous failure every time, which is why the systems here are built as co-pilots with hard constraints rather than open-ended loops.

Agentic coding is genuinely fast for the first 80%. It is also blind to the wake it leaves. A production-readiness audit of one of my own codebases surfaced 462 sensitive console logs across 61 files — OAuth tokens, user IDs, workspace secrets — logged in plain text because an agent decided in the moment that seeing the raw payload would be helpful.

That is the hidden cost of agentic speed: when you are not writing every line, you stop noticing what gets injected. The lesson was not to stop using agents. It was that anything an agent writes needs the same audit gate as anything a contractor writes.

For a decade the standard SaaS pattern was: middleware checks the JWT, middleware writes the query, app returns the data. That worked because developers were the only ones writing queries. Once agents are generating queries, a middleware check is a suggestion.

So the boundary moves down into the database. Row Level Security is enforced by Postgres regardless of what the calling code believed it was doing — the one part of the stack that does not care about the vibes of an agent. Every multi-tenant surface I ship is built that way.

Defaulting every task to a premium model is using a sledgehammer for every nail. Routing requests dynamically across four models — cheap ones for prose, capable ones for reasoning, multimodal only where multimodal is actually needed — took my own weekly AI spend from $45 to $8.81.

That figure matters less than the habit behind it: measure what each task actually needs before you pay frontier prices for it. A system whose unit economics you cannot explain is a system you do not control.