A quote that looks fair — and isn’t
A few months ago, the founder of a US SaaS startup reached out to us. He’d gotten a quote from a well-known dev shop to build an AI agent for customer support and technical help. The price? $9,500.
“It looked fair,” he told us. “We checked two or three vendors, and this was one of the lower prices.”
We ran a CODEFAIR Scan on the quote. What we found changed the company’s entire trajectory.
The architecture they proposed — and why it’s dangerous
The vendor proposed a standard RAG (Retrieval-Augmented Generation) architecture:
- Customer asks a question
- The system pulls the customer’s entire conversation history
- On top of that, it injects entire technical support manuals (roughly 100,000 tokens)
- Everything gets sent to the Claude 4.6 model API for analysis and a response
Sounds technical? Here’s what it means in dollars:
Monthly cost breakdown
| Parameter | Value |
|---|---|
| Conversations per day | 1,000 |
| Tokens per conversation (input) | ~100,000 |
| Cost per million tokens (Claude 4.6) | $2.50 |
| Cost per conversation | $0.25 |
| Daily cost | $250 |
| Monthly cost (API only) | $7,500 |
In under two months — usage costs overtake the original development cost. And in the first year? $89,000 on API alone — before server costs, support, or updates.
The vendor quoted a “fair price” for development. But the OpEx baked into the architecture killed the entire venture’s economics.
What CODEFAIR recommended changing
1. Semantic caching at the vector layer
Repeated questions (like “what are your hours?” or “how do I reset my password?”) can be answered from cache with zero API calls.
Estimated savings: 35–45% of calls
2. A local SLM for simple questions
A Small Language Model like Phi-3 Mini runs locally on the server and handles simple questions with no API cost.
Estimated savings: an additional 20–30% of calls
3. Focused RAG: chunking + reranking
Instead of injecting 100,000 tokens, a focused reranking mechanism sends only the 2,000 most relevant tokens.
Savings: 98% of input-token cost per conversation
The result: monthly cost from $7,500 to $300
| Before | After | |
|---|---|---|
| Development cost | $9,500 | $10,300 |
| Monthly API cost | $7,500 | $300 |
| Annual cost (API) | $89,000 | $3,600 |
| 3-year savings | — | ≈ $257,000 |
Development cost rose by $800. OpEx cost dropped by 96%.
What you can learn from this
A question every founder must ask before signing:
“How much will it cost me to run this system per month, with X active users?”
If the dev shop can’t give you a detailed answer — that alone should be a red flag.
A careless AI architecture doesn’t look like a problem in the quote. It looks like a low, attractive price. The problem only shows up 60 days after launch, when the first API bill arrives.
CODEFAIR analyzes tech quotes and catches cost traps before you sign. Upload your proposal for a scan »