So i spend way too much time looking at how companies claiming to be "AI-powered" or "built with AI" actually implement their tech. sometimes clients ask me to audit or rebuild their systems, sometimes founders reach out for architecture review before they burn through their series A.. And theres this pattern thats honestly hilarious.
These companies raise $5M+ selling investors on "proprietary AI algorithms" and "advanced machine learning infrastructure" but when you look under the hood its literally just OpenAI API calls wrapped in a react app.
out of 23 companies i audited that explicitly marketed themselves as AI companies:
87% are just making API calls to OpenAI/Anthropic with zero actual ML infrastructure. their "proprietary AI" is a $0.002/request GPT-4 call with a custom system prompt
91% have no vector database, no embedding pipeline, no fine-tuning, nothing. one company raised $3.2M for an "AI research assistant" that was literally just ChatGPT API + web scraping + markdown formatting
74% store conversation history in postgres with no optimization. one app i saw was doing full table scans on 300k rows for every single chat message. their "AI is slow" - no bro your database design is slow
100% of the ones claiming "we fine-tuned our model" had never actually fine-tuned anything. i asked to see their training data and one founder said "we're planning to do that next quarter"
the math on the OpenAI wrapper problem:
company markets as: AI-powered legal research platform with proprietary algorithms actual stack: nextjs + openai api + pinecone (free tier) engineering team: 4 people what they actually built: a nice UI and some clever prompts
cost breakdown per 1000 users at moderate usage:
- openai api: $12k-18k/month
- infrastructure: $400/month
- pinecone (when they finally upgrade): $70/month
margin problem: theyre charging $29/user/month but openai costs eat 41-62% of revenue before any other costs
i saw one company spending $47k/month on OpenAI API calls while charging customers $38k in MRR. they were losing $9k/month just on AI costs and the founders had no idea because they never looked at their AWS billing vs revenue.
the worst part is the promises vs reality:
company A - "we use advanced neural networks trained on millions of legal documents" reality: gpt-4 api + embeddings from openai + they scraped justia for free legal docs when i asked about their "training": "we engineered really good prompts"
company B - "our proprietary AI model understands context better than GPT" reality: gpt-4-turbo with a 15k token system prompt they spent 3 months perfecting their "better than GPT" claim was literally just... using GPT with more context
company C - raised $4M for "AI that writes production-ready code" reality: gpt-4 + github copilot api + custom react components for the IDE actual innovation: they made a good UI and focused on devops workflows problem: they told investors they built the AI themselves
why this matters:
look im not saying API wrappers cant be good businesses. some of the best products are just excellent UX over existing APIs. stripe is kind of an API wrapper over payment processors. but they dont claim to have "proprietary payment processing algorithms"
the problem is when you lie to investors about your tech stack and then:
- burn VC money on OpenAI costs that scale linearly with users
- have zero moat when openai releases chatgpt with plugins
- cant explain what happens when openai raises prices
- realize you built a $10M business on someone else's infrastructure that could disappear tomorrow
what actually works for AI companies:
the 3 companies out of 23 that had real tech did this:
- fine-tuned models on domain-specific data (actual training, not prompts)
- built retrieval systems with hybrid search (keyword + vector + reranking)
- optimized inference costs with model distillation or quantization
- had fallback strategies if their LLM provider changed pricing
- were honest about their stack in investor docs
one founder told me "we spent 6 months on our RAG pipeline and it gives us 40% better accuracy than raw GPT on technical docs. that's our moat." thats real differentiation
if your'e building an AI product right now:
- using openai API is fine, lying about it is not
- "prompt engineering" is not a moat, its a weekend project
- your actual innovation might be UI/UX, workflows, or domain expertise
- if your entire business dies when gpt-5 launches with a better system prompt, you dont have a business
- investors are getting smarter about this, technical due diligence is real now
happy to answer questions about AI architecture, what actually constitutes technical differentiation, or how to build AI products that dont collapse when openai changes their pricing.