The checklist that decides who wins

SOC 2 Type II, SSO/SAML, data residency, RBAC, audit logging, sandbox isolation, BYOC. Not one item concerns model quality. Enterprise agent procurement has arrived, and it does not care which model you use.

The stated minimum for enterprise agent deployment is now SOC 2 Type II, GDPR and HIPAA alignment where applicable, SSO/SAML, data residency, sandbox isolation, audit logging, RBAC and bring-your-own-cloud. Every item predates AI by a decade. None is about capability.

What that list means for who wins

A better model does not close a SOC 2 gap. A startup with the strongest agent on the market and no audit report loses to an incumbent with a weaker product and a complete compliance file — and loses in procurement, where no benchmark is consulted.

This is the ordinary fate of every enterprise software category, arriving exactly on schedule. It is also why the model layer commoditising matters less to incumbents than it should: they were never selling the model.

The one genuinely new requirement

Sandbox isolation. Everything else on that list exists because software touches data; isolation exists because an agent generates and executes code paths nobody wrote or reviewed. It is the item most specific to this technology and the one most likely to be waived during a pilot, because it slows things down and the pilot is small.

Where the checklist and the risk diverge

A complete compliance file does not tell you what your agents are doing. 82% of executives are confident their policies cover unauthorised agent actions; 14.4% of organisations ship with full security approval — and it is entirely possible to satisfy every item above while the second number stays where it is.

Which is why the gateway pattern is the more consequential development. Certification is a statement about how an organisation intends to behave. A gateway is a record of what its agents actually did, including what they were prevented from doing.

The advice for buyers

Use the checklist — it filters out vendors who cannot operate at enterprise scale, and that is a real service. Then ask the question it does not: when this agent invokes a tool in my environment, what stands between the request and the execution, and who reads the log?

Northflank — Enterprise AI coding agent deployment in 2026 → · Airtable — Best Enterprise AI Agent Platforms for 2026 → · Lyzr — Understanding Enterprise AI Agents: The 2026 Guide → · ITNET — Securing enterprise AI agents: agentic workflows 2026 →