The problem
By design, every custom application JDQ builds and supports reports its warnings, validation messages and errors back to a JDQ monitoring mailbox, even when the application has already caught and handled them. Many of these reports come from new versions in development and user acceptance testing (UAT), where finding issues early is the whole point. That gives us a complete picture of how our software behaves, but across more than ten applications it adds up to a large volume of routine reports. Only about 1% point to something a developer needs to act on, such as a genuine bug. All large software has some; ours runs to hundreds of thousands of lines of code. Reviewing the stream by hand meant those few important reports could wait among the routine ones, minor issues sometimes took weeks to surface, and JDQ developers and our clients' QA testers carried more of the testing burden than they needed to.
What we built
An AI agent now reads each report as it arrives, judges its priority and counts how often it recurs, separating the few that matter from routine warnings and validation messages. For each important one it documents the problem, proposes likely causes and root causes, and opens an issue for the right developer or project manager in Project Watch™, JDQ's own end-to-end project and business management app. It uses generative AI with retrieval (RAG) over our project knowledge and MCP connections to the Project Watch™ databases, and runs on JDQ's own AI server so client data stays private and secure. During UAT, testers no longer have to write up every problem by hand, because the agent may already have documented it. JDQ developers and client QA testers can spend their time confirming the software does what the business needs. The agent works at the Supervised level of our trust ladder: a developer reviews every diagnosis before anything is changed or a client is contacted. It is running in development on real reports from client systems and is now being rolled out. JDQ developers built the agent itself with AI coding agents.
Results
Lessons
We were surprised by how well the agent separates the important from the routine. Measured results, and the mistakes it makes along the way, will be added as the pilot continues.