The transparency gap: what most top SaaS products get wrong for AI agents

The software industry faces a structural problem as AI agents become capable of operating applications autonomously. While most products have invested in making their interfaces visually appealing to humans, they have largely ignored the machine-readable signals that autonomous agents require to function effectively.

A recent audit of the top 100 SaaS product homepages reveals the extent of this oversight. Using the Agent Factors Engineering framework—a methodology that scores products across eight principles affecting agentability—the analysis found an average transparency score of just 5 out of 100. More striking: 83 of the 100 products scored zero on this principle.

What agentability measures

Agentability refers to how well AI agents can operate software products. The Agent Factors Engineering framework assesses this capability across eight principles, each scored from 0 to 100. These principles include machine readability, transparency, shadow-UI avoidance, defaults, control, chunking, status, and clean handoffs.

The overall audit found an average agentability score of 38.3 across all products. The distribution broke down into four tiers: 22 products rated Agent-Ready with scores of 45 or above, 54 classified as Developing with scores between 35 and 44, 17 marked as Lagging with scores from 20 to 34, and 7 considered Agent-Blind with scores below 20.

While the overall scores suggest moderate attention to agent-friendly design, the distribution across individual principles reveals significant gaps. Machine readability scored highest at 71, indicating that basic semantic HTML and structured markup are relatively common. However, transparency and shadow-UI avoidance emerged as critical weak points, averaging 5 and 17 respectively.

Understanding transparency in the AFE framework

Transparency in Agent Factors Engineering refers to how clearly a product communicates its available capabilities, actions, and API endpoints to autonomous systems. For a human user, discovering what a product can do involves visual exploration, reading documentation, or trial and error. AI agents lack this intuitive discovery process and instead require explicit, machine-readable declarations of what operations are possible.

A transparent interface might include structured schema definitions that describe available actions, clear API documentation linked from the interface itself, or semantic annotations that indicate interactive elements and their purposes. Without these signals, an AI agent must resort to guessing, which increases error rates and reduces reliability.

The homepage audit specifically examined whether products provided any machine-readable indication of their core capabilities or available actions. This could take the form of JSON-LD schema markup describing the product's functions, clear links to API documentation with structured endpoint descriptions, or semantic HTML that explicitly labels key interactive components.

Why 94 products score under 20

The data shows that 94 of the 100 audited products scored under 20 on transparency. This near-universal gap suggests that transparency for agents is not yet a design consideration for most product teams. The scores indicate that most homepages provide minimal to no structured information about what the product can do in a format that an autonomous system could parse and understand.

This stands in sharp contrast to the machine readability scores, where the average of 71 indicates that basic semantic HTML structure is relatively common. The discrepancy suggests that while teams implement foundational markup for SEO and accessibility, they have not extended this thinking to capability declaration.

The low transparency scores correlate with another weak principle: shadow-UI avoidance, which scored an average of 17. Shadow-UI refers to interface elements that are generated dynamically or hidden behind interactions that agents cannot easily predict. When combined with low transparency, products create environments where agents must navigate interfaces designed exclusively for human visual processing, without any guide to what actions exist or how to access them.

The compound effect of low transparency

Transparency does not operate in isolation. When products score low on transparency, the deficiency compounds with weaknesses in other principles. An agent attempting to operate software with poor transparency must rely more heavily on other signals—status indicators, predictable defaults, and clear chunking of complex operations.

The audit found moderate scores in these compensating areas: defaults averaged 35, chunking 45, and status 47. While these scores are higher than transparency, they remain in the developing range. Control, which measures how easily agents can direct product behavior, averaged 48—the highest score among action-oriented principles, but still indicating significant room for improvement.

Clean handoffs, which measures how well products facilitate transitions between human and agent control, averaged 40. Without transparency about available actions, clean handoffs become nearly impossible, as agents cannot communicate clearly to humans what they were able to accomplish or what actions remain.

What product teams should do

Addressing the transparency gap requires concrete changes to how products expose their capabilities. Teams can start with several practical interventions that do not require redesigning entire interfaces.

Add structured capability declarations

Implement JSON-LD schema markup that describes the product's core functions and available actions. This structured data should be embedded in the homepage and key landing pages, providing a machine-readable map of what the product can do.

Create agent-accessible API documentation

Ensure that API documentation is linked from the main interface using clear, semantic HTML. The documentation itself should include structured endpoint descriptions that agents can parse to understand available operations, required parameters, and expected responses.

Use explicit semantic labeling

Label interactive elements with ARIA attributes and semantic HTML that describe not just what the element is, but what action it enables. A button labeled semantically as "create-new-project" is more transparent than one labeled generically as "submit."

Expose capability metadata

Consider adding a machine-readable manifest or capability file that enumerates the product's functions. This could follow emerging standards for agent-computer interfaces or use existing formats like OpenAPI specifications.

Test with actual agents

The most direct path to improving transparency is testing with AI agents. Teams can run a free audit to establish a baseline score, then iterate on capability declarations while measuring improvements in agent task completion rates.

The path forward

The transparency gap represents an opportunity rather than an indictment. Most products score low because transparency for AI agents has not yet entered the standard product development checklist. As autonomous systems become more capable and prevalent, the products that invest early in clear capability signaling will have a significant advantage.

The broader agentability scores, available in the Agentability Index, show that many products have already addressed machine readability basics. Extending this foundation to include transparency requires a shift in perspective: thinking of the interface not just as a visual presentation layer, but as a communication protocol between the product and any system—human or artificial—that might operate it.

For the 83 products that scored zero on transparency, the improvement path is clear and the starting point is straightforward. Adding structured capability declarations and explicit semantic labeling costs little but delivers immediate value for agent operation. As the ecosystem matures, transparency will likely shift from optional enhancement to competitive requirement.