FreshBooks and the Agentability Gap in Modern SaaS

As AI agents evolve from experimental tools to production systems, a critical question emerges: can they actually operate the software that runs modern business? Our audit of 100 leading SaaS products suggests the answer is "barely." The average agentability score stands at just 38.3 out of 100, with systemic weaknesses in the architectural decisions that determine whether AI agents can successfully interact with web applications.

FreshBooks, the accounting platform serving millions of small businesses, sits squarely within this landscape. Its detailed agentability profile illustrates the specific technical barriers that plague the SaaS industry—and points toward concrete solutions.

What Agentability Measures

Agentability quantifies how well software accommodates AI agent operation across eight Agent Factors Engineering (AFE) principles. These aren't aspirational guidelines; they're measurable technical characteristics that determine whether an agent can perceive state, take action, and recover from errors when interacting with web applications.

The eight principles evaluate:

• Machine readability: Whether semantic HTML and ARIA provide unambiguous element identification
• Transparency: If machine-readable documentation exposes capabilities, schemas, and business rules
• Shadow-UI avoidance: Whether interactive elements exist in accessible DOM rather than canvas or obfuscated JavaScript
• Defaults: If sensible fallbacks reduce the decisions agents must make
• Control: Whether agents can disable interruptions and operate deterministically
• Chunking: If complex operations decompose into observable steps
• Status: Whether current state and operation results are programmatically exposed
• Clean handoffs: If graceful human escalation pathways exist

Each principle scores 0-100, with the overall agentability score representing the composite assessment.

The Industry Baseline

Our audit of the top 100 SaaS products by market presence reveals an industry largely unprepared for agentic interaction. The tier distribution tells the story:

• 22 products qualify as Agent-Ready (scoring 45 or above)
• 54 fall into Developing territory (35-44)
• 17 rank as Lagging (20-34)
• 7 products are effectively Agent-Blind (under 20)

The principle-level averages expose where the industry struggles most. Machine readability scores highest at 71, reflecting years of accessibility investment that incidentally benefits agents. But transparency averages just 5, shadow-UI avoidance scores 17, and defaults reach only 35. These aren't marginal weaknesses—they're architectural gaps that fundamentally limit what agents can accomplish.

The Transparency Crisis

Zero. Not "poor" or "minimal"—complete absence. These products offer no machine-readable documentation of their capabilities, no schema definitions agents can parse, no programmatic discovery of what actions are possible or what constraints apply.

For human users, this absence barely registers. We navigate through visual exploration, learn through trial and error, and accumulate contextual knowledge over time. AI agents lack this luxury. Without explicit capability exposure, agents resort to brittle heuristics: guessing at element purposes, inferring workflows from visual layout, maintaining hard-coded assumptions that break with each interface update.

The FreshBooks audit demonstrates what this looks like in practice. Like most accounting platforms, FreshBooks presents a complex domain model—invoices, clients, expenses, payments, taxes—with intricate business rules governing state transitions. An agent tasked with "create an invoice for Client X" must navigate not just UI elements but financial logic: required fields, tax calculations, payment terms, approval workflows. Without machine-readable documentation of these rules, the agent operates blind.

The Shadow-UI Problem

Shadow-UI refers to interactive elements that bypass standard DOM structures—controls rendered in canvas, functionality hidden behind obfuscated JavaScript, state managed in ways inaccessible to assistive technologies and agents alike. Modern frontend frameworks make shadow-UI trivially easy to create and difficult to detect.

For agents, shadow-UI represents an impenetrable barrier. An agent can parse semantic HTML, query ARIA properties, and traverse accessible DOM trees. It cannot interpret pixels in a canvas element or reverse-engineer minified JavaScript to discover hidden affordances. When critical functionality lives in shadow-UI, agent operation becomes impossible regardless of how sophisticated the AI model.

The low industry average of 17 on this principle reveals how pervasive the problem has become. What developers implement as performance optimization or rich interaction patterns systematically excludes programmatic access.

Where FreshBooks and Peers Stand

FreshBooks exemplifies the industry median. It demonstrates typical strengths—reasonable machine readability in core navigation, adequate status exposure for completed operations—and typical weaknesses, particularly in transparency and shadow-UI avoidance.

This isn't a criticism specific to FreshBooks. The patterns observed in its audit appear across financial software, project management tools, CRM platforms, and virtually every SaaS category. The issue isn't individual product decisions but industry-wide architectural assumptions formed before agentic interaction became a design consideration.

To explore how specific products compare, consult the Agentability Index, which provides scored assessments across all 100 audited products.

Guidance for Product Teams

The agentability gap creates both risk and opportunity. Products that remain agent-hostile will struggle as customers increasingly expect AI-driven automation. Those that systematically address AFE principles will differentiate on a newly relevant dimension.

Concrete steps product teams can take:

Address transparency first. Machine-readable API documentation, schema definitions, and capability manifests deliver immediate value. Start with OpenAPI specs for existing APIs, then extend to UI capabilities and business rules. This is tractable engineering work, not research.

Audit for shadow-UI. Inventory interactive elements that aren't accessible via standard DOM queries. Prioritize migrating critical workflows to semantic HTML with proper ARIA. Performance concerns can often be addressed without sacrificing accessibility.

Instrument status exposure. Ensure operation results, error states, and system status are programmatically queryable, not just visually indicated. This also improves human accessibility.

Design for determinism. Provide agent-facing modes that disable non-deterministic interruptions like modals, tooltips, and dynamic content insertion. The control principle scores 48 industry-wide; this represents low-hanging fruit.

Test with agents. Include agentic interaction in QA processes. Have agents attempt representative workflows and measure success rates. Failures indicate architectural barriers worth addressing.

For teams ready to benchmark their current state, run a free audit to generate a scored assessment across all eight AFE principles.

The Path Forward

The 38.3 industry average isn't destiny. It reflects architectural decisions made when agentic interaction wasn't a consideration. As that changes—as customers demand agent-operable software and competitive pressure mounts—products will adapt.

FreshBooks and its peers in the Developing tier have clear optimization paths. The Agent-Ready tier proves these principles are achievable with current technology. The question isn't capability but priority: when will product teams treat agentability as a first-class design concern?

The data suggests that moment is now.