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: 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/monthinfrastructure: $400/monthpinecone (when they finally upgrade): $70/month 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 A 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 B 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 company C why this matters: 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 usershave zero moat when openai releases chatgpt with pluginscant explain what happens when openai raises pricesrealize you built a $10M business on someone else's infrastructure that could disappear tomorrow 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: 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 quantizationhad fallback strategies if their LLM provider changed pricingwere honest about their stack in investor docs 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: 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 projectyour actual innovation might be UI/UX, workflows, or domain expertiseif your entire business dies when gpt-5 launches with a better system prompt, you dont have a businessinvestors are getting smarter about this, technical due diligence is real 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.

I Reverse-engineered How 23 'AI-first' Companies Actually Build Their Products

About Author

Comments

TOPICS

THIS ARTICLE WAS FEATURED IN

Related Stories

The Problem With Persistent AI Memory: It Doesn’t Forget Context

Challenges in Building Natural, Low‑Latency, Reliable Voice Assistants

The Real Fix for React Native + pnpm: Hoist Everything

The Noonification: How Often Do NFTs Pass The Howey Test? (1/13/2023)

Darwin's Hybrid Intelligence to Align AI & Human Goals for Startups & VCs

The Noonification: White Man (11/26/2022)

The Problem With Persistent AI Memory: It Doesn’t Forget Context

Challenges in Building Natural, Low‑Latency, Reliable Voice Assistants

The Real Fix for React Native + pnpm: Hoist Everything

The Noonification: How Often Do NFTs Pass The Howey Test? (1/13/2023)

Darwin's Hybrid Intelligence to Align AI & Human Goals for Startups & VCs

The Noonification: White Man (11/26/2022)

Light-Mode

Classic

Newspaper

Minty

Dark-Mode

Neon Noir

Minty

HN StartUps