NOTE: Demo visuals use either blurred real data or synthetic placeholders to protect customer privacy.
Cut the Broken Telephone
Today, a single customer question can pass from a rep to a help channel, to a program manager, to a product manager, then to an engineer, and back again. The route changes each time, and invites slow or inconsistent replies. Broad AI tools help with retrieval, but they also risk hallucinations, outdated docs, or tone mismatches.
- Documentation may be wrong or out of date.
- Messaging sometimes answers the question but not the intent.
- AI responses can be hard to trust or verify.
Teams spend time fact-checking instead of selling. A more reliable system needs clear sourcing, accuracy, and restraint when uncertain.
Code as Ground Truth
Andy’s prototype builds on the Nora framework. It checks product code directly to find the truth behind a question. It verifies across multiple repositories, compares product sources, and asks clarifying questions when the input is ambiguous. When confidence drops, it simply declines to answer rather than risk an error.
Key functions include the ability to:
-
Read live product code to generate answers from the source of truth
-
Cross-checks across multiple product systems
-
Uses clarifying questions to confirm context
-
Refuses uncertain answers instead of guessing
-
Benchmarked through SME review and human evaluation
The result is a cautious, verifiable assistant that behaves more like a trusted expert than a generic chatbot. Early tests show it can answer technical questions such as configuration and posture counts with high factual accuracy.
Raising Trust and Speed
Code-grounded responses reduce the need for engineering escalations and speed up deal cycles. Accuracy builds confidence across sales, product, and support.
- Higher precision on technical questions
- Reduced Slack traffic and rework
- Consistent tone for customer-facing replies
- Clear abstentions when the answer is unknown
In future iterations, Andy aims to deploy the bot internally, refine its tone for customer messaging, and expand coverage to new product areas and data sources.
Early Reflections from Peers
Early observers highlighted two breakthroughs: sourcing directly from code and explicitly flagging uncertainty. Both behaviors encourage trust and usability.
This stage marks a shift from experimental Q&A to dependable automation. As accuracy improves and tone improves, the bot moves closer to direct-to-customer use, helping Abnormal’s GTM teams deliver human-quality answers at machine speed.
