Auto Pivdenny is a 6-yard used-car dealership in Kyiv. Their Lead Hunter agent was Grow2.ai's third pilot, in 2023. The qualification rules we wrote that summer became the template every B2C lead-hunter pilot starts from. This note publishes them, with Pavlo's permission, and explains why each one is there.
The eight rules, verbatim
- Budget bracket clear within first 3 messages, or escalate to human. If the lead won't state a budget after the agent has politely asked twice, the conversation goes to a salesperson, not because the lead is bad but because diagnosing "won't state vs. doesn't know" is a human skill.
- Financing yes/no, before suggesting any specific car. Auto Pivdenny's in-house financing terms are different from third-party financing terms. Suggesting a car the lead can't actually afford on the financing path they'll use is a wasted conversation.
- Manual vs. automatic, before suggesting any specific car. Sounds obvious. Was not, in pre-agent conversations: ~14% of leads got shown the "wrong" gearbox and dropped out silently.
- Mileage tolerance — explicit number or "no preference". Pavlo's buyers either have a specific maximum (140,000 km is common for "3-year-old replacement") or genuinely don't care. The middle ground is rare.
- Must-have features (max 3). Lead names up to three. The agent privately tags "automatic climate", "tow hitch", "heated seats" etc. against the inventory feed. If a lead has more than 3 must-haves, that's a signal of indecision and we route to a salesperson.
- Yard preference / driving radius. Auto Pivdenny's 6 yards span 40 km. A Saturday-buyer from Borshchahivka cannot reach Brovary in their lunch break. The agent only suggests cars at yards the lead can reach.
- Researching vs. ready-to-test-drive. The agent asks directly, late in the conversation: "Are you doing initial research, or could you test-drive this Saturday?" The phrasing is deliberate — it gives the "researching" lead an honourable label and the "ready" lead a clear path. ~62% answer truthfully.
- Family situation, only if volunteered. The agent does not ask "do you have children" or "are you buying for the family". If the lead volunteers context ("need space for two car seats"), the agent uses it. Privacy default is on.
The two we tried to add and rolled back
Attempted rule #9 · Region of origin (which oblast)
We thought regional preferences for body type (sedan vs. SUV vs. wagon) were strong enough to score. They're not — the variance within a region was wider than the variance between regions. The rule overfit and hurt qualification accuracy by 3 points. Removed after 3 weeks.
Attempted rule #10 · Previous car make
Hypothesis: a lead trading in a German car has different price sensitivity than one trading in a Korean car. True in aggregate, useless at the lead level. ~40% of leads couldn't recall the trade-in's make at the moment we asked. The rule produced more conversation friction than signal. Removed after 6 weeks.
Why eight, not twenty
Pavlo's instinct on Day 1 was to write twenty rules. Grow2.ai pushed back, hard. Every rule the agent has to evaluate is another decision boundary that can go wrong. Eight rules with high-quality answers beats twenty rules with mediocre answers. Three years later, the ruleset is still eight. The model has changed; the rules haven't needed to.