Sellavie.ai pipeline: purchase intent, handoff, and context loss
Stabilizing the AI sales pipeline after context loss, stock disclosure, handoff, and model-choice problems showed up at the same time.
By late March, the hard part was not adding more AI.
The hard part was getting the AI to respect the product state around it.
This phase worked on purchase intent, handoff, context loss, stock disclosure, and model updates. That combination says a lot about the product: an AI response is only correct if the surrounding business state is correct.
The issue
An AI sales pipeline can fail quietly.
It can sound confident while using the wrong product variant. It can forget a customer already asked for something. It can keep talking when a human should answer. It can hide stock uncertainty behind a smooth sentence.
The fix was not better wording. It was better boundaries:
- require clearer purchase signals before creating intent state
- preserve handoff as a real state that the AI cannot silently overwrite
- avoid clobbering admin model choices with defaults
- disclose stock limits honestly instead of optimistically
- keep memory useful without letting it loop on stale context
What this took
Most of this work was not model work. It was product work: understanding where the AI was being given too much latitude, adding the right constraints at the right layer, and writing tests that caught the specific failure patterns we had already seen in production.
The LLM is not the hard part of an AI commerce product. The hard part is the system around it.