An AI sales agent can ask good questions and still be unhelpful if it recommends a property or car that does not exist. A reliable design separates conversation from inventory truth.
In SyntaxLab's lead-qualification demo, the model reads an enquiry and returns structured criteria such as location, budget, type, condition, and timeline. Application code matches those criteria against a fixed set of sample listings, scores the lead with a visible rubric, assigns an agent, and prepares a next step.
Use the model to understand the enquiry
People describe what they want in flexible language: “two bedrooms near the metro, ready before summer, under this budget.” A language model can turn that into fields the application can use. Structured output makes those fields easier to validate and render than a paragraph.
The model should not be asked to invent available stock. Its job is to interpret the request and identify missing details.
Let application code own matching and price
The demo filters and scores a known inventory using criteria such as property type, area, budget, and condition. It returns only matching records from the catalogue. A fixed scoring function calculates clarity, urgency, engagement, and budget fit; the UI exposes the breakdown.
This is a useful boundary for sales systems: AI understands the request; business logic determines which products are actually available and what they cost. If the catalogue changes, the matching layer can use live inventory APIs without asking the model to memorize it.
Make the score an aid, not a verdict
Lead scores can help a team prioritize, but they reflect chosen assumptions. A short response time or a stated budget may be useful signals, yet neither proves that a prospect will buy. Show the score inputs, allow staff to correct the record, and monitor whether the rubric supports the team's process.
Keep the handoff concrete
Useful output includes a short summary, matched inventory, unanswered questions, owner assignment, and a suggested next step. A human should be able to review those details before they become a customer-facing promise or CRM commitment.
Project evidence
The SyntaxLab software-lead flow is implemented in syntaxlab-backend/src/leads.ts; consumer property and car matching appears in src/leads-consumer.ts. The current demo uses seeded example leads and sample inventory. It does not establish a live MLS, dealer feed, CRM connection, or sales-performance result.