AI agent tool adds on-screen guides to direct users where to click in SaaS products

This tool addresses a common gap in AI-driven support for SaaS products: the inability to guide users through UI workflows when full automation isn’t feasible. Instead of defaulting to RAG-based help center searches or generic web results, both of which often return outdated or irrelevant information, the Assist API generates on-screen guides tailored to the user’s current view. This reduces the cognitive load of mapping text instructions to UI elements, a frequent pain point for users navigating complex applications.

The implementation requires a browser-based agent to crawl and map the application UI, which introduces setup costs. Engineers must provide a test account (e.g., staging or preview environments) for the agent to log in and explore the product. The agent then builds a model of the UI, including permissions and feature flags, which the Assist API queries to generate guides. While this process automates documentation, it also means the tool’s accuracy depends on the completeness of the UI crawl and the frequency of updates, which can be triggered via CI/CD or scheduled runs.

The tool’s utility is constrained by the scope of the UI it can map. If the browser agent misses workflows or fails to account for dynamic UI elements (e.g., conditional rendering), the generated guides may be incomplete or incorrect. Additionally, the tool’s reliance on a test account means it may not capture real-world user states or edge cases, such as permission variations or custom configurations. For products with rapidly evolving UIs, the scheduled or CI/CD-triggered updates will need to keep pace to avoid stale guidance.

From an integration standpoint, the tool is designed to slot into existing AI agent frameworks, as demonstrated with the Vercel AI SDK. This lowers the barrier to adoption for teams already using such frameworks, but it also means the tool’s effectiveness is tied to the broader AI agent’s capabilities. For example, if the parent agent misinterprets the user’s query, the Assist API may generate a guide for the wrong task. The tool’s three-tiered response system (guide, text, or rejection) provides flexibility but also requires the parent agent to handle cases where the tool cannot assist.