Saviynt Introduces Zuma for Apps to Govern AI Agent Actions

Saviynt has introduced Zuma for Apps, a solution that evaluates the actions AI agents take inside business applications such as SAP, Salesforce, Oracle, and ServiceNow. It is part of Zuma, the Saviynt Enterprise AI Identity Security Platform, and customers can reach the capabilities initially through a private preview program.

Saviynt calls the approach “action governance.” Where access governance shows what an identity can do, action governance examines what an agent did, on whose behalf, and through which chain of agents, then asks whether the action should have been allowed.

AI Agents Now Run Inside Business Applications

Most agent security watches the model, code, and platform where an agent was built. But the business transaction happens in the applications that run the enterprise, and agents embedded in those applications, such as SAP Joule, operate inside the application’s own perimeter. Saviynt says tools outside the application do not see what those agents execute.

AI-driven actions already reach production systems. In the SAP Secure Together Community’s 2026 global survey on segregation of duties (SoD) risk management in SAP environments, which had 100 respondents, 38% of deep-dive respondents reported AI-driven actions reaching production SAP transactions. Among the 21 organizations running agents in production, only two have fully enforced SoD guardrails.

Saviynt’s position is that SoD checks built around one identity at a time miss conflicts that form across a chain of agents. Vinit Shah, Saviynt’s senior vice president of product management, gives a hypothetical example. A person asks an agent to pay a known vendor. When one sub-agent cannot find the vendor, it creates a new one using the person’s access, and a second sub-agent pays it. Each entitlement looked appropriate on its own, but a routine payment became a high-value payment to a new vendor.

“The business risk was invisible until the transaction was done,” Shah writes.

Zuma for Apps Maps the Agent Chain and Checks Each Action

Zuma for Apps starts by mapping the chain behind an action: the humans who can invoke an agent, the authority delegated to it, and the sub-agents and tools it can call. The map also covers activity and transaction data, which Saviynt attributes to 15 years of connector work across SAP, Salesforce, and Oracle.

Zuma for Apps then evaluates SoD across the chain and across applications, instead of one identity at a time. When it finds a conflict, such as vendor master data maintenance combined with invoice approval, it traces the conflict to the invoking human and the agents involved. It also expresses realized and potential exposure in financial terms.

The check also runs before the transaction completes. The Agent Access Gateway and Intent-Aware Runtime Authorization evaluate who is behind an action, what authority was delegated, and the agent’s declared purpose, and the outcome is allow, block, or escalate. Zuma’s Trust Center adds signals such as scope drift and credential hygiene, and a supervisory agent escalates to a human reviewer when risk warrants.

Saviynt says it registers as the Model Context Protocol (MCP) server inside Joule Studio. The Joule agent then has no direct path to SAP systems, and every skill invocation passes through Saviynt’s policy engine first.

Zuma for Apps Extends Saviynt’s Application Access Governance

Zuma for Apps builds on Saviynt Application Access Governance (AAG), the company’s product for managing access and detecting SoD violations in applications such as SAP. Saviynt says existing AAG customers get the same control plane extended to agents.

Zuma for Apps also sits within Zuma, which Saviynt launched earlier this year to find the AI agents running in a company, name a person accountable for each, and check what they do while they work. Zuma covers agents built on platforms such as AWS Bedrock, Google Vertex AI, and Microsoft Copilot Studio. Zuma for Apps adds the application layer, where those agents carry out business transactions.

Zuma for Apps is available through the private preview and is not part of the 30-day Zuma free trial. Saviynt is inviting organizations that deploy or evaluate AI agents to share use cases and feedback before general availability.

What This Means for SAPinsiders

Agent deployment is running ahead of agent controls. Survey results show agents reaching production SAP transactions while few organizations running agents in production have fully enforced SoD guardrails. Saviynt is building Zuma for Apps through a private preview with customers that are deploying or evaluating agents.

Saviynt extends identity security to what agents do. Zuma for Apps evaluates actions before a transaction completes. Saviynt built it on an application access product centered on who can reach which systems, bringing identity controls into the business process, including payments and invoices.

Governance for SAP agents can come from outside SAP. Saviynt reaches SAP Joule through Joule Studio and works with Salesforce and Oracle through connectors, all under one control plane. In Saviynt’s design, customers with agents in several applications govern them from one place.