Use cases by role

SOC 2 auditors expect access to be attributable and logs to be unalterable after the fact, a bar most agent deployments miss because agents share credentials and write to logs an admin can edit.

Agencies face procurement rules, jurisdiction-specific state AI laws, and federal frameworks at once, with a higher bar for defensible evidence when an oversight body asks what an agent did and why.

Teams adopt agents and MCP servers faster than procurement or security can track, and every unmanaged connection is an unaudited tool surface the organization does not know it carries.

Agents that summarize records, schedule care, or route messages handle protected health information under a regime with little tolerance for uncontrolled exposure, yet are often deployed with the same broad, static access a human employee would have.

A support agent that reads account data, issues refunds, or escalates cases is processing personal data and taking consequential actions — but was shipped for speed, not for the audit evidence a regulator or enterprise customer will ask for.

Financial-services agents touching trading, account, or customer data need attributable, revocable access, but non-human identity programs were built for static service accounts, not for agents that spawn, act, and must be individually revocable without downtime.

Once agents call agents across org boundaries, the question stops being "is this authenticated" and becomes "should this specific external agent reach this specific internal one, for this task, right now" — which a static partner API key cannot express.

A coding agent with shell and filesystem access uses every capability it is granted, including ones nobody meant to leave open — and pipeline credentials are usually broad, long-lived, and shared across jobs.

A marketplace inherits the security posture of every agent it publishes, and buyers need a way to judge an agent's trustworthiness before granting it access.

Chained review agents compound each other's errors, and privilege and confidentiality mean some decisions cannot be fully automated.

Internal developer platforms let any team stand up an MCP server in minutes, but nothing stops that server from becoming an ungoverned, unaudited path into production systems.

A reasoning loop, a retry storm, or one misconfigured agent can turn a normal day's spend into a five-figure surprise, and finance usually finds out when the invoice arrives.