Sales Lead Proposal System
A lead qualification and proposal engine that predicts purchase intent with semantic similarity rating, instead of asking an LLM to score leads directly.
Ask an LLM to rate a sales lead from 1 to 5 and it will give you a number without hesitating. The trouble is the distribution. Models pile up on two or three values, avoid the extremes, and quietly shift their scale when you rephrase the question. Any single score looks reasonable. A thousand of them don’t.
This system uses Semantic Similarity Rating (SSR) instead, a method introduced by Maier et al. in LLMs Reproduce Human Purchase Intent via Semantic Similarity Elicitation of Likert Ratings.
SSR doesn’t ask for a number at all. The model answers in ordinary language, and the rating is recovered afterwards by embedding that answer and comparing it against reference statements anchored to each point on the scale. The score comes from where the response sits semantically rather than from the model asserting a digit.
Two things made it worth building on:
- The distributions look human. The paper reports KS similarity above 0.85 against human response distributions, and about 90% of human test-retest reliability across 57 surveys and 9,300 responses.
- You get the explanation for free. The model writes prose before anything is scored, so every rating already has its reasoning attached. That matters when a salesperson wants to know why a lead landed where it did.
For lead qualification that’s the difference between an opaque number and something a person can argue with. The same output then feeds proposal generation downstream.
Credit: the SSR method is the work of Maier, Aslak, Fiaschi, Rismal, Fletcher, Luhmann, Dow, Pappas, and Wiecki. My work was implementing and adapting it for lead qualification and proposal generation.
innoscripta SE · 2025