AI Query Risk Check

Project Summary

AI Query Risk Check is a lightweight chatbot prototype that helps workplace users determine through whether a planned AI query is likely low risk, needs revision, or should be reviewed before use.

The demo version was built in approximately 4 hours, including basic QA. It uses a structured Voiceflow intake workflow, an LLM-assisted prompt risk interview, and an API-driven guidance lookup to give users practical, plain-language guidance.

Why I Built It

More employees are using AI tools in daily work, and organizations need simple ways to help people pause before sharing sensitive, confidential, regulated, or decision-related information. However, many organizations are not creating or using clear AI governance policies and employees are turning to using personal accounts with corporate information.

This bot is not designed to approve AI use or replace internal policy. It is an informational triage assistant that helps users identify risk signals, consider safer prompt approaches, and decide whether a planned query should be revised or not used.

Architecture

The prototype uses a hybrid design. A deterministic intake workflow collects consistent variables such as policy status, organization type, query purpose, data type, and output use. That workflow sets a minimum risk level before the user enters the more flexible prompt interview. The variables serve two purposes:

• They provide the LLM with consistent context without requiring the user to restate the same information
• They establish datapoints that can be used to define further governance requirements and update the guidance cards to match common request types over time.

The Prompt Risk Interview playbook then asks for a high-level description of the planned AI task. The LLM can raise the risk level based on the user’s description, but it cannot lower the minimum risk set by the workflow. This keeps the governance-critical logic predictable while still allowing natural conversation.

After the risk interview, Voiceflow sends a JSON payload through a Make webhook to simulate a full API POST integration. Make receives the structured lookup key, searches a Google Sheets guidance table, and returns JSON containing the matching guidance card, safer prompt template, and next step.

This API pathway demonstrates structured data passing, external system integration, next-best-action retrieval, and scalable content maintenance outside the bot. GPT-5 nano is used for both agents because the classification task is bounded by deterministic intake variables, controlled lookup keys, and external guidance content. As the capabilities of the chatbot expand, the model may need to be upgraded based on the complexity of the tasks.

Future enhancements and adjustments may include:

1. Moving off of voice flow to a custom bot framework using pure JSON and Netlify or Cloudflare for LLM hosting and processing.
2. Enhancing feedback to just-in-time prompt planning and refinement, with a heavy focus on data privacy and security best practices.
3. Integrating with organizational AI registries and usage logs to provide more accurate risk assessments.
4. Improving the prompt risk interview to handle more complex queries and provide more tailored guidance.
5. Adding a supplemental agent to create additional guidance cards on demand based on common user requests, with the card content flagged for human review before being added to the guidance table.