Implementation Guide

How to Govern AI Agents That Approve Expense Reports

Governing an AI agent with expense approval authority requires four concrete controls: a distinct machine identity separate from any human user, permission scopes with explicit dollar and category thresholds, runtime policy enforcement that can block a non-compliant approval before it executes, and immutable audit logs capturing the agent's inputs, policy checks, and final decision. Without all four, the agent operates as an unaccountable financial actor rather than a controlled system component.

Where Policy Enforcement Should Sit

  1. Two architectural models, two risk profiles

    There are two architectural models for enforcing expense-approval policy, and they produce materially different risk profiles: enforcing rules before an approval executes, or only reviewing decisions after the fact. The remainder of this guide explains why that distinction determines whether an agent is truly governed.

Four Controls for Expense-Approval Agents

Every expense-approval agent should be evaluated against the same four structural controls, independent of how accurate its decisions appear to be.

Agent Identity

A non-human credential distinguishable from any human approver.

Permission Scope

Explicit thresholds by amount, category, and vendor.

Runtime Enforcement

Policy checks that block actions before execution.

Audit Trail

Logged reasoning, checks, and outcomes for every approval.

Audit Trail Requirements

A defensible audit trail for an expense-approval agent needs to satisfy the following, at minimum:

  • The agent's distinct identity is recorded on every approval action
  • The specific policy rules checked at the time of approval are logged
  • The inputs used to reach the decision are captured, not just the outcome
  • Logs are stored in an immutable or tamper-evident format
  • Escalations and their resolutions are logged as part of the same trail
  • Logs are structured in a format accessible to internal and external auditors

Why Expense Approval Is a Useful Governance Test Case

Expense approval is a bounded, financially consequential decision, which makes it a clear lens for examining agentic AI governance. Unlike open-ended AI assistance, an expense-approval agent takes a discrete action with a dollar value, a vendor, a policy category, and a downstream accounting entry. That specificity is what makes the governance requirements tractable. An enterprise does not need to solve for every possible agent behavior; it needs to define what this agent is allowed to approve, under what conditions, and how that decision is recorded. Organizations deploying agents for this function without addressing identity, scope, and audit are effectively giving a piece of software signing authority over company funds without the controls that would apply to a human employee in the same role.

Agent Identity: Separating the Approver From the Human It Acts On Behalf Of

A recurring architectural gap is agents that inherit a human user's session, API key, or role rather than operating under their own identity. This makes it impossible to determine, after the fact, whether a given approval was made by a person or by an autonomous process. Agent identity should be issued as a distinct credential, tied to a defined role and permission set, and visible as such in every log entry and downstream system record. This is not a cosmetic distinction. When an approval is later disputed, audited, or investigated, the first question is who approved it. If the answer is ambiguous between a human and an agent acting under that human's identity, the control has already failed, regardless of what policy logic the agent applied.

Permission Scoping and Approval Thresholds

Least-privilege access is a well-established security principle, and it applies directly to expense-approval agents. In practice, this means defining explicit boundaries: a maximum dollar amount the agent can approve unilaterally, a list of expense categories it is authorized to act on, and vendor or merchant restrictions where relevant. Anything outside those boundaries should not be approvable by the agent at all; it should route to a human. The scoping decision is not a one-time configuration. Permissions should be reviewed on a defined cadence to catch privilege creep, where an agent's effective authority expands over time through configuration changes, new integrations, or accumulated exceptions that were never revisited.

Why Runtime Enforcement Matters More Than Detection

Post-hoc review has a place in any control environment, but it cannot substitute for runtime enforcement when the action in question moves money. If an agent's approval only gets checked against policy after the transaction has been recorded and potentially paid, the organization is relying on remediation rather than prevention. Runtime policy enforcement means the system evaluates the proposed approval against defined rules, such as threshold limits, category restrictions, or duplicate-submission checks, before the action is committed. If the action would violate policy, it is blocked or automatically escalated to a human approver instead of completing. This distinction is one of the clearest differentiators between an agent that is genuinely governed and one that is simply monitored.

Runtime check, not just a log entry

If a policy violation is only visible in a report generated after payment, that control is detection, not enforcement. Enforcement stops the action before it completes.

Escalation Paths for Ambiguous Cases

Not every expense report will fall cleanly inside or outside the agent's defined authority. Ambiguous cases, such as a borderline amount, an unusual vendor, or a policy exception request, need a defined escalation path to a human reviewer rather than allowing the agent to interpret its way to a decision. This is an extension of segregation-of-duties principles from traditional finance controls, adapted for a system where the reviewing party is not a person. The organization should document what triggers escalation, who receives it, and what happens if no human responds within a defined window, since an unresolved escalation should not default to automatic approval.

Evaluating Your Current Architecture

Organizations already running or piloting expense-approval agents should assess their architecture against the four controls described here rather than against the agent's approval accuracy alone. An agent that makes correct decisions most of the time but operates under a shared credential, has no enforced approval ceiling, and produces logs that only show a final status is not governed; it is simply performing well so far. The relevant evaluation questions are structural: can you name the agent's identity independent of any human user, can you point to the specific threshold beyond which it cannot act, can the system block a non-compliant approval before it completes, and can you reconstruct the full decision path for any given approval on demand. If any of these cannot be answered directly, that is the gap to close before expanding the agent's authority further.

Bring Runtime Governance to Your AI Agents

Trussed AI provides runtime governance and security for enterprise AI agents, including agent identity, permission enforcement, and audit logging for agents that take consequential actions.

Request a Demo