What AI cannot see: account state, cost, and policy

The single most important thing to understand about AI for cloud work is what it cannot see. The model works from text you give it and patterns it learned in training. It has no live connection to your environment unless a specific tool provides one, and even then it sees only what that tool exposes.

Three blind spots cause the most trouble:

• Your account state. AI does not know which resources already exist, what your current state file says, or what a change would actually create, replace, or destroy. It can guess the shape of your infrastructure; it cannot read it.
• Real cost. A model can tell you that a larger instance type generally costs more, but it does not know your committed-use discounts, your negotiated rates, your current spend, or how many of a resource you are about to multiply. It will happily suggest something expensive with total confidence.
• Your organization policy. Naming standards, required tags, allowed regions, approved modules, and security baselines live in your company, not in the model. Generic best practice is not your policy.

These gaps are not bugs you can prompt away. They are structural. That is why the safe pattern is always the same: let AI draft and explain, then let deterministic tools that do see your account — a plan or diff, a cost estimator, a policy scanner — tell you the truth before anything changes.