The phrase enterprise AI deployment hides a five-layer security stack that most pilot programs only encounter once they begin negotiating a procurement review.

What the operating teams treat as table stakes — gateway, identity, prompt provenance, retrieval boundary, and red-team logs — the project teams are still discovering.

Across the dozen rollouts we reviewed, a pattern emerges: the firms that ship are those that wrote the security architecture before the pilot scoring began.

"Procurement now asks five questions before any model touches production data — and only one of them is about the model."

Cloudflare AI Gateway is doing the heavy lifting in five of those rollouts, paired with MCP-bound vendor documentation that scopes what each agent can read.

The remaining work is operational: who owns the runbook when an agent fails, who updates the policy when a model version changes, who is paged.

Reporting compiled across 10 interviews with operating teams and procurement leads in the ai cohort. Names anonymized at the request of the firms involved.