Implementation Guide

Governing AI Agents in ServiceNow Now Assist

Governing AI agents in ServiceNow Now Assist requires enterprises to define agent identity independent of end users, scope permissions at the tool-call level rather than the application level, and add runtime enforcement and audit logging that go beyond static role-based configuration. Native platform controls alone do not reliably close these gaps.

Four Governance Gaps to Address

Before implementing controls, enterprises should confirm where their Now Assist deployment stands on each of the following:

Agent Identity

Determine whether agents act under their own identity or inherit user or service account permissions.

Permission Scoping

Restrict agent access to specific tools, tables, and workflows rather than broad application roles.

Runtime Enforcement

Add a control point that evaluates agent actions before execution, not only at configuration time.

Audit Logging

Capture the agent's decision path, not just the resulting system action.


Where Runtime Enforcement Fits

  1. 1

    Beyond Static Configuration

    Static configuration, such as roles and ACLs assigned when an agent is deployed, defines what an agent is permitted to do in principle. It does not evaluate whether a specific action, at the moment the agent attempts it, is appropriate given context, sensitivity, or prior activity. Runtime policy enforcement requires an intermediary control point that can intercept an agent's proposed tool call or action, evaluate it against policy, and approve or deny it before execution completes.

Why Now Assist Introduces a New Governance Problem

ServiceNow Now Assist agents differ from traditional workflow automation because they make autonomous decisions about which tools, tables, or integrations to invoke in response to a request, rather than executing a fixed sequence configured in advance. This shifts governance from a design-time question, what was this workflow built to do, to a runtime question, what did this agent decide to do in a specific instance and was that decision appropriate. Enterprises evaluating Now Assist deployments need to treat this as an instance of the broader enterprise AI agent governance problem rather than a platform-specific configuration task. The underlying risks (unclear identity boundaries, permission scope that is too broad for the task, and insufficient visibility into agent decisions) apply to any agentic system operating inside a system of record. ServiceNow's specific implementation details should be confirmed directly against its official product documentation before an enterprise finalizes its governance architecture.

Agent Identity: The Foundational Decision

Before permissions or logging can be addressed, an enterprise must establish how a given Now Assist agent's identity is defined. In general enterprise identity practice, an autonomous agent can operate under one of several models: a dedicated service identity scoped to that agent alone, a shared service account used across multiple agents, or a delegated identity that inherits the permissions of the human user who triggered the interaction. Each model carries different blast radius implications if the agent is compromised, misconfigured, or simply makes an incorrect tool-call decision. A dedicated identity per agent allows for precise permission scoping and clean audit attribution. A shared service account or inherited user identity makes it difficult to answer a basic governance question after an incident: which agent, acting under what authority, took this action. Enterprises should confirm which identity model Now Assist uses in their specific deployment before assuming that existing user or role-based access controls provide adequate separation.

Permission Scoping Beyond the Application Level

Standard ServiceNow access control, built on ACLs and role-based permissions, is designed to govern what a human user can see or do within the platform. Agentic systems require a finer-grained model because a single agent may hold broad application-level access while only needing to invoke a narrow set of specific tools, tables, or workflow steps for any given task. Least-privilege enforcement for AI agents generally means scoping permissions at the level of the individual tool call or action, not the application or module. Enterprises should map every tool, workflow, and third-party integration an agent is capable of invoking, classify each by data sensitivity and system-of-record impact, and then determine whether existing ServiceNow role configurations are sufficient to restrict the agent to only what a given task requires. Where an agent's role grants access broader than any single task needs, that gap represents unmanaged risk regardless of how well the underlying workflow itself is designed.

Auditability: Logging Decisions, Not Just Outcomes

Standard system logging typically records the resulting action, such as a record update or workflow trigger, but does not necessarily capture why an agent chose to invoke a particular tool or take a particular path to reach that outcome. For an autonomous agent, the decision path is the material fact an auditor or incident responder needs: what options were available, what the agent selected, and what triggered that selection. Enterprises should establish logging requirements, covering action, acting identity, justification, and outcome, before deployment rather than attempting to reconstruct this after an incident. This is consistent with how identity governance treats human accounts, where access reviews and activity logs are expected as a baseline control, and the same standard should apply to agents that are granted comparable or greater system access.

Where Native Controls Fall Short

Platform-native access controls are generally built around static, role-based configuration applied at deployment time. They do not inherently address the runtime question of whether a specific agent action, in a specific context, should be permitted, nor do they typically capture the reasoning behind an agent's tool selection. Third-party integrations invoked by an agent introduce an additional trust boundary that may sit entirely outside the platform's native governance scope. Where these gaps exist, enterprises need a supplementary policy layer that can enforce least privilege at runtime and produce audit records suitable for compliance and incident response. This is the specific area where runtime governance tooling, including Trussed AI's runtime policy enforcement, agent identity, agent permissions, and audit logging capabilities, is intended to address gaps that static platform configuration does not close on its own.

Frequently Asked Questions

Does ServiceNow's native role-based access control fully govern Now Assist agent permissions?

Role-based access control governs what an agent's identity is broadly permitted to touch within the platform, but it does not typically enforce granular restrictions on individual tool calls or evaluate the appropriateness of a specific action at runtime. Enterprises should confirm current capability against ServiceNow's official documentation.

Should an AI agent share a service account with other agents or workflows?

Shared service accounts make it difficult to attribute a specific action to a specific agent during an audit or incident response. A dedicated identity per agent generally supports clearer accountability and more precise permission scoping.

What is the difference between design-time configuration and runtime enforcement for AI agents?

Design-time configuration defines what an agent is permitted to do in principle through roles and permissions. Runtime enforcement evaluates a specific proposed action against policy at the moment it is attempted, before execution completes.

Address Runtime Governance Gaps in Now Assist Deployments

Where native platform controls stop at static configuration, runtime governance adds identity verification, least-privilege enforcement, and audit logging for AI agent actions.

Explore Runtime Governance