Sellavie.ai purchase flow: the regression wall that made the agent safer
Hardening purchase intent, stale orders, handoff behavior, payment interpretation, and building a regression suite after real conversation failure patterns showed up.
This was the point where the AI sales agent stopped being judged by demo conversations and started being judged by ugly real ones.
The issue was not whether the AI could answer politely. The issue was whether it could avoid reopening old orders, misreading "I already paid," suppressing new enquiries, or creating checkout loops when the customer was asking something simple.
What the failure patterns looked like
Some of these bugs were subtle. The agent would detect purchase intent in a message that was actually a status check. It would restore a handoff state when the human had already resolved the issue. It would generate a follow-up promotion after a system tag got stripped from the message history.
None of these felt like crashes. They felt like the agent being slightly wrong in a way that eroded trust.
What changed
This hardening pass added a serious purchase-flow regression suite and tightened backend behavior around:
- purchase intent detection thresholds
- stale order blocking
- handoff state restoration
- low-signal follow-up suppression
- payment interpretation
- auto-promotion after stripped system tags
- customer context continuity
The real engineering challenge
The agent had to be helpful without becoming reckless.
If it creates an order too eagerly, the business gets noise. If it blocks a valid purchase, the business loses revenue. If it forgets context, the customer loses trust. If it keeps trying to sell after a human should have taken over, the product feels unsafe.
The work was about narrowing behavior, not expanding it. That is the harder direction with AI products.