Glenda Brendler

Signal AI V2: the autonomous agent transforming feedback into product strategy

Date

August 19, 2025

Contact me

The challenge: from raw data to accelerated strategic decisions

Managing user feedback is a critical point for any product-led organization. Traditionally, product, support, and sales teams expend significant resources merely collecting, organizing, and attempting to extract strategic value from massive and dispersed volumes of feedback.

V1 of the 'AI Feedback Agent' project (check out the explanatory video at the bottom of the page) demonstrated the feasibility of automation to centralize and categorize this data. However, the true gap persisted in interpretation: how to move beyond 'what' the user says to understand 'why' they say it and 'how' the product team can act concretely? V2 was conceived to close this gap, elevating the agent from a mere classifier to an autonomous strategic assistant, capable of generating actionable insights and accelerating the product decision cycle.

The solution: self-correcting architecture and strategic AI

Signal AI V2 materializes this concept into a robust automation, orchestrated via Make.com, integrating Google Sheets, OpenAI, and Supabase. Its mission is clear: ingest raw feedback from various sources and produce structured intelligence, ready to guide product strategy.

The 7-step architecture is designed with a focus on resilience, self-correction, and efficiency:

Google Sheets - Search Rows (intelligent trigger): performs an active search for feedback with empty status (new) or 'Processing...' status (failed). This logic ensures no task is lost, and failures are detected and automatically reprocessed.
Google Sheets - Update a Row (intelligent 'status locking'): each identified feedback is immediately marked as 'Processing...' in the spreadsheet. This temporary 'lock' prevents duplicate processing of the same task and provides visibility into its current status.
OpenAI (the agent's strategic brain): utilizes a carefully designed prompt (defensive version v2.1) that instructs the AI to act as an 'Experienced Product Analyst', extracting and structuring:
- Sentiment: the emotional valence of the feedback.
- Job to be Done: the user's fundamental need.
- Suggested priority: based on pre-defined impact and effort criteria.
- Design hypothesis: a concrete proposal to solve the problem.
- UX opportunity (How Might We...): a question that catalyzes ideation and solution exploration.
JSON - Parse JSON (output validation): validates and structures the AI's response into a clean and consumable JSON format, preparing the data for storage.
Supabase (secure data persistence): stores the enriched insights in a database, ensuring integrity, scalability, and performance. Row Level Security (RLS) policies are applied to ensure authorized access to the data.
Google Sheets - Update a Row (cycle finalization): after successful processing and storage, the original row in the spreadsheet is updated to 'Processed', completing the feedback lifecycle.

This architecture not only automates processing but ensures it does so reliably, resiliently, and strategically, converting raw data into actionable intelligence.

Resilience in action: overcoming challenges and senior thinking in automation

Building an intelligent and autonomous system is an exercise in engineering and product design that reveals the ability to handle complexities. This project served as a testing ground for critical problem-solving, demonstrating expertise in architecting robust solutions:

Trigger robustness and failure reprocessing: overcame the fragility of reactive triggers with a 'stateless' system (Search Rows) and self-correcting 'status locking'. This ensures that initially failed feedback, stuck in 'Processing...' status, is detected and automatically reprocessed until completion, eliminating task limbo and guaranteeing 100% processing.
Data security and RLS: resolved 401 errors in Supabase through the implementation and debugging of Row Level Security policies. This measure ensured secure API operation, protecting access to strategic insights.
AI consistency and defensive prompting: mitigated AI 'hallucinations' and invalid output formats (JSON) through aggressive prompt iteration. The v2.1 prompt ('Strict Rules' and few-shot learning) was defensively designed to enforce consistency, strategic quality, and the expected output format.
Resource management and rapid iteration: addressed a critical credit consumption issue on an automation platform (Make.com). Instead of paralyzing the project, the strategic creation of a new development environment allowed for a clean rebuild and accelerated project delivery, demonstrating strong risk management and continuity capabilities.

This project stands as a testament to the ability to translate complex requirements into functional, secure, and resilient technical solutions.

Impact and future vision

Signal AI V2 elevates the feedback process, providing product teams with a constant stream of strategic and actionable insights—far beyond mere categorization. This significantly accelerates the discovery, validation, and development cycle.

Next steps and optimization opportunities:

Failure pattern detection and management: implement a mechanism to identify and isolate tasks that repeatedly fail (e.g., after N attempts) into a 'dead letter queue'. This would prevent continuous resource consumption on irrecoverable feedback, routing it for targeted human review.
AI output schema validation: strengthen the architecture with explicit validation of the AI's JSON output structure and content. This would ensure that saved data always conforms to expected types and formats, even in the face of unexpected model behavior variations.
Cost and scalability optimization: evaluate transitioning to event-driven triggers (webhooks) instead of polling (Search Rows) in high-volume scenarios. This optimization would reduce operation consumption and costs, improving performance at scale.