Franchise sales qualifier + financial model bot (80 % pre-qualification)

A bot that answers arbitrarily deep financial-model questions from prospective franchisees / customers: "can I afford this in city X?", "what does my P&L look like month-3?", "what if I open with half the staff?". Uses a long-context LLM (200K minimum; 1M is dramatically easier) to keep the full P&L model + the prospect's local inputs + the unit-economics deck in one conversation. 80 % of pre-qualification is bot-driven; humans handle closing.

Franchise / deep-B2B sales typically require a senior manager 3–6 hours per prospect to walk a P&L. The bot replaces those hours, scales to nighttime / weekend prospects, and gives a prospect a consultation deeper than a junior manager could give in their first 3 months. One franchise founder tested this against their own bench: bot's answers were measurably better than 80 % of his junior salespeople's.

1. Encode the financial model as structured data (Excel + JSON). Every assumption is named.
2. Encode unit economics, regional adjusters, regulatory variants.
3. Encode 30–50 real prospect transcripts from past sales — these become the few-shot exemplars.
4. System prompt: bot's role (consultative, never closes), knowledge boundaries, when to escalate to human.
5. Conversation logging: every prospect chat saved with timestamps. Used both for follow-up and for model fine-tuning.
6. Hot-handoff to human when prospect signals readiness (named phrases: "I'm in", "send me the contract", "what's next?"). Bot doesn't close.
7. Weekly review: which question types did the bot get wrong? Patch the model + the prompt.

• Don't let the bot close. Closing requires reading social signals the bot can't reliably do. Closing-by-bot causes mismatched-expectation churn 2 months later. Hard-cap the bot at qualification.