Why I stopped letting LLMs do math (and built a deterministic pricing layer instead)

Hey everyone,

Like a lot of you, I’ve been building AI agents for local businesses and contractors. But I kept hitting the exact same wall: Pricing Hallucinations.

I set up strict pricing rules, tax rules, and margin floors in the system prompt. If the user asks a simple question, it works. But the absolute second a prospect starts negotiating—"Can we remove the premium filter?" or "Can I get an off-peak discount?"—the LLM starts guessing numbers. It calculates taxes on the wrong base price, ignores mandatory fees, and invents discounts.

The realization: AI generates prices based on probabilities and patterns. Pricing requires deterministic math (1+1 always equals 2). Better prompt engineering cannot fix this.

So, I built a solution: Quotix. It’s a deterministic pricing engine (rules-first node architecture) that handles the math, while the AI just handles the conversation. The AI is the salesman; the engine is the calculator. They never cross over.

I just recorded a deep-dive video where I show the exact prompt failures, explain why the LLM breaks, and do a walkthrough of the node canvas I built to fix it.

I am building this in public and would love the feedback of the builders in this group, since you guys are actually in the trenches deploying these systems for clients.

1. The Video Breakdown (The Problem & The Architecture):https://youtu.be/mX5NLB5xSg4?si=UK-GbezuM3pCm4el

2. The Live Demo (Please keep in mind work in progress): https://thegrowthmark.com:8090/engine/823d012fde358c02fc9a7fd464dd3f4a

Would love to hear your thoughts and feedback!

3 comments