Freshworks and the Agentability Gap: What 100 SaaS Homepages Reveal About AI-Ready Design

The software industry stands at an inflection point. As AI agents increasingly interact with web applications on behalf of users, a fundamental question emerges: How well can these agents actually operate the software we've built? A comprehensive audit of the top 100 SaaS products provides a sobering answer, with implications for vendors like Freshworks and across the entire industry.

Understanding Agentability

Agentability measures how well AI agents can operate software, scored on a scale of 0-100 across eight Agent Factors Engineering (AFE) principles. These principles evaluate whether software interfaces provide the structure, clarity, and predictability that agents need to navigate tasks reliably. Unlike human users who can interpret visual context and ambiguous interfaces, AI agents depend on machine-readable structures, explicit state information, and consistent patterns.

The eight principles assessed include machine readability, transparency, shadow-UI avoidance, sensible defaults, user control, appropriate chunking, status visibility, and clean handoffs. Each principle addresses a specific dimension of agent-friendly design, from the technical foundation of semantic HTML to the architectural decision of how to communicate system state.

The Tier Distribution

The audit categorized products into four tiers based on their agentability scores. At the top, 22 products achieved Agent-Ready status with scores of 45 or above, demonstrating that agent-friendly design is achievable with current web standards. The majority—54 products—fell into the Developing tier with scores between 35 and 44, suggesting partial implementation of AFE principles but significant room for improvement.

Further down, 17 products landed in the Lagging tier with scores between 20 and 34, while 7 products were classified as Agent-Blind with scores below 20. This distribution reveals that while a fifth of products have made meaningful progress, the industry as a whole remains in the early stages of agentability maturity.

The Transparency Crisis

Transparency, in the context of agentability, refers to how well a product communicates its capabilities and requirements to AI agents. This includes declaring authentication methods, exposing rate limits, documenting available actions, and providing machine-readable service descriptions. Without this information layer, agents must resort to trial-and-error approaches or rely on brittle heuristics that break with interface updates.

The average transparency score across all products was just 5 out of 100, making it the weakest principle by far. This gap is particularly significant because transparency serves as a foundation for agent operation—without knowing what's possible and what's required, agents cannot plan or execute tasks reliably.

The Shadow-UI Problem

Shadow-UI refers to interface elements rendered entirely through JavaScript without semantic HTML fallbacks, including dynamic overlays, custom dropdowns, and framework-generated components that lack proper ARIA labels or role attributes. While these elements often provide polished user experiences for humans, they create opacity for agents that rely on DOM structure to understand interface state and available actions.

The prevalence of shadow-UI reflects a decade of frontend development practices optimized for visual presentation rather than programmatic access. Single-page applications, in particular, often generate complex component hierarchies that obscure the underlying data model and user actions from agents attempting to parse the interface.

Relative Strengths

Not all principles showed universal weakness. Machine readability averaged 71 out of 100, indicating that most products do provide basic semantic HTML structure even if they layer problematic patterns on top. This foundation is encouraging—it suggests that improvements in other areas may not require complete architectural overhauls.

User control averaged 48, chunking 45, and status visibility 47, all clustering around the middle of the scale. These moderate scores suggest that products already implement some agent-friendly patterns, perhaps inadvertently, through practices that also benefit human users. Clear status indicators and appropriate information chunking serve both audiences, creating alignment between human UX and agent operability.

Sensible defaults averaged 35, while clean handoffs averaged 40, indicating partial implementation. These principles address how products handle state transitions and cross-context workflows, areas where intentional design for agents could yield significant improvements without degrading human experience.

What This Means for Product Teams

For engineering and product teams across the SaaS industry, including those at companies like Freshworks, these findings present both a challenge and an opportunity. The challenge is clear: current design and development practices do not adequately support AI agent operation. The opportunity lies in the fact that few products have solved this problem, making agentability a potential differentiator as agent-mediated interactions become more common.

Practical steps begin with the two weakest principles. Improving transparency requires adding machine-readable documentation of capabilities, authentication requirements, and operational constraints. This might take the form of structured metadata in page headers, standardized API descriptions, or agent-specific documentation endpoints. None of these changes need to impact the human user interface.

Addressing shadow-UI requires auditing frontend components for semantic structure and ARIA annotations. Custom components should expose their state and available actions through standard HTML attributes and roles. Where JavaScript-heavy interactions are necessary, teams should ensure that equivalent information is available through the accessibility tree—a practice that simultaneously improves screen reader support.

Teams can benchmark their own products against these industry averages using tools from The Agentability Project. The Agentability Index provides detailed scoring breakdowns for major products, while teams can run a free audit of their own properties to identify specific gaps.

The Path Forward

The average score of 38.3 out of 100 is not a permanent ceiling. It reflects the current state of an industry that optimized for human-first design before agents became significant users of software interfaces. As teams recognize the importance of agent operability, scores will rise—but only through intentional application of AFE principles.

Products that address transparency and shadow-UI avoidance first will see the most significant score improvements, given how underserved these principles are across the industry. More importantly, they'll enable more reliable agent interactions, reducing errors and expanding the range of tasks agents can complete successfully on behalf of users.

The transition to agent-ready software is not a binary switch but a gradual process of architectural refinement. The data from this 100-product audit establishes a baseline and identifies the most critical gaps. Whether a product currently sits in the Agent-Blind tier or approaches Agent-Ready status, the path forward involves the same fundamental work: making software that is legible, predictable, and transparent to both human and machine users.