How agent-ready is Brex?

Brex received an agentability score of 44/100 on its homepage, 52/100 on pricing, and 40/100 on documentation. These scores indicate that while basic machine parsing is functional, autonomous AI agents will encounter significant friction when attempting to navigate the site, complete multi-step workflows, or understand system state and reasoning.

The audit identified strong performance in machine readability (80/100), meaning structured data and semantic markup are generally accessible to agents. However, Brex falls short in transparency (0/100), shadow UI avoidance (30/100), and defaults (33/100). The complete absence of transparency signals—such as confidence indicators, source citations, or audit logs—means agents cannot assess result quality or trace system decisions. Low scores in shadow UI avoidance and defaults suggest the interface relies heavily on visual cues and requires manual input where intelligent pre-population would improve agent efficiency.

Moderate scores in control (48/100), status (50/100), and clean handoffs (50/100) indicate partial support for agent-driven tasks, but missing affordances like labeled pause controls and enumerated progress indicators force agents to guess at system state during long-running operations.

To improve agentability, Brex should focus on the following actionable fixes:

• Add transparency to all key outputs by including confidence fields in API responses and 'verified vs. best-guess' labels in the UI, enabling agents to assess result reliability.
• Expose activity and audit logs in machine-readable formats (such as JSON event logs) with a visible 'why this result' affordance so agents can trace system actions and decision paths.
• Implement enumerated multi-step progress indicators or activity-feed patterns for long-running tasks to give agents clear status visibility.
• Provide labeled pause, cancel, and stop controls with accessible names like 'Cancel import' rather than unlabeled icons, allowing agents to reliably interrupt operations.
• Pre-fill form inputs with sensible defaults wherever reasonable assumptions exist, reducing the number of decisions agents must make.
• Attach source references or input citations to generated outputs so agents can verify data provenance and link results back to originating information.