Saudi Arabia is backing its AI ambitions with investments of up to $100 billion as Vision 2030 approaches its target year. That’s a serious commitment, and a bet that organizations across the Kingdom can turn that investment into business value.

Agentic AI raises the stakes. It promises something automation doesn’t: software that decides for itself what to do next. Reasoning models have made that promise more convincing, but reasoning and agency aren’t the same thing. A system reasoning well inside someone else’s workflow is still following that script. It only earns the label “agentic” once it can be determined its own next step and be held accountable for where that leads. Much of what is marketed as “agentic AI” still looks more like sophisticated automation.

That distinction becomes clear when you consider what these systems act on. What an agent can safely decide is limited by how much of reality it can see. In most enterprises, that picture is split across network records, cloud platforms, and security tools that don’t talk to each other. Ask whether they agree on what’s happening right now, and the answer is often no.

Hand an agent that fragmented foundation and it won’t pause to double-check. It will act, at machine speed, on incomplete information, without necessarily knowing what it’s missing. That’s the opposite of the digital sovereignty Saudi Arabia is building toward, where you can’t control what you can’t verify.

None of that is a reason to slow-walk agentic AI. It is a reason to build the right foundation first. Before giving agents more authority, the Kingdom’s enterprises need more than smarter models. They need trusted infrastructure where identity, telemetry, and network records reflect the same reality. Without that foundation, agentic AI runs on unverified data. With it, enterprises can scale autonomy with confidence.

That foundation is what makes staged autonomy possible. First come recommendations based on reliable data, then actions approved by a person, and finally limited authority to act independently. Without that foundation, an agent is still guessing with confidence.

A CIO evaluating agentic vendors this year has one clean test: ask what the system does the moment it’s confident and wrong. If the answer is “that hasn’t come up yet,” that’s an untested assumption on infrastructure it hasn’t been asked to survive.