Brex’s ‘Invisible’ AI: Can a Mesh of Autonomous Agents Really Automate Finance?

Brex’s CTO James Reggio claims the company aims to use AI to make itself ‘effectively disappear’ through an ‘Agent Mesh’ of role-specialized agents. But for finance managers at mid-sized enterprises, the question remains: how can they validate Brex’s 99% automation claims using this system?

The Agent Mesh relies on event-driven architecture with plain-English communication between agents and no central coordinator. To test compliance checks and payment initiations, a finance manager could simulate scenarios where agents must autonomously verify receipts against policy rules and initiate payments.

For example, uploading a receipt that violates a company’s expense policy should trigger an automated rejection without human input. The system’s LLM-based evaluations act as a ‘judge,’ auditing agent decisions to ensure alignment with corporate guidelines.

When errors arise—like a receipt not matching a policy—the plain-English communication between agents is meant to resolve conflicts. Instead of rigid code-based error messages, agents exchange contextual explanations (e.g., ‘This receipt lacks required tax details’).

However, Reggio admits the system ‘remains a bit of a technology where we don’t entirely know the limits of it,’ suggesting real-world edge cases may still require manual intervention.

The absence of a central coordinator complicates debugging. If a reimbursement request fails, there’s no single log to trace the breakdown.

A finance manager would need to analyze message streams between agents to identify where communication faltered. Brex reports 99% automation for customers using its AI tools, but the lack of third-party benchmarks leaves room for skepticism about how much of this is truly autonomous versus human-assisted.