I Analyzed 100 AI Startups: Here's What They're Building

I spent the last six weeks tracking down 100 AI startups that launched in 2024. What I found wasn't what I expected.

Everyone's talking about how AI is eating the world, but after digging into the trenches scrolling through Product Hunt launches, GitHub repos, Y Combinator batches, and LinkedIn announcements I realized we're not in an explosion of diversity. We're in an explosion of convergence. And that convergence is telling us something important about where this market is actually heading.

How I Found Them (And Why This Isn't Perfect)

Let me be upfront about my methodology because it's flawed, and that matters.

I started with Product Hunt's AI category from January to October 2024, filtered for startups with actual traction (200+ upvotes or visible GitHub activity). Then I layered in Y Combinator's S24 and W24 batches, cross-referenced with Twitter announcements from VCs, and added companies I'd encountered through developer communities and conferences.

The biases here are obvious: I'm more likely to catch B2B/developer-focused tools than consumer apps. English-speaking, US-centric startups are overrepresented. Stealth companies aren't here. And my own work at Weam.ai where we're building AI infrastructure means I probably noticed infrastructure plays more than vertical apps.

Still, even with these limitations, patterns emerged. Strong patterns.

The Categories: Where Everyone's Building

Here's the breakdown of what these 100 startups are actually doing:

Agent Platforms: 28%
The largest category by far. These are platforms that let you build, deploy, or orchestrate AI agents. Think "we're the Zapier for AI agents" or "we're the operating system for autonomous workflows."

RAG/Search/Knowledge Management: 22%
Everything from enterprise search tools to document Q&A systems to "ChatGPT for your data" implementations. This was the most crowded space, which surprised me given how commoditized basic RAG has become.

Developer Tools: 19%
Observability, testing, evaluation, prompt management, model deployment. The picks-and-shovels plays for AI development.

Vertical AI Applications: 18%
Purpose-built AI solutions for specific industries or use cases: legal research, medical coding, sales outreach, recruiting, content creation for e-commerce.

Infrastructure & Core Tech: 13%
Vector databases, inference optimization, model training platforms, data labeling tools.

The agent platform dominance is the story. At the beginning of 2024, "agents" felt like a buzzword. Now it's become the default mental model for how AI products should work. Whether that's justified or just groupthink is still an open question.

The Tech Stack Everyone's Using

Here's where things get really homogeneous.

Model Layer:

• OpenAI (GPT-4/GPT-4o): 67% of startups

• Anthropic (Claude): 31%

• Open-source (Llama, Mistral): 18%

• Multiple models: 42%

Most startups are model-agnostic now. They've learned from 2023's chaos that betting on a single provider is risky. The typical pattern is: default to OpenAI, offer Claude as an option, maybe sprinkle in some open-source for cost optimization.

Vector Database:

• Pinecone: 34%

• Weaviate: 16%

• Qdrant: 12%

• PostgreSQL + pgvector: 21%

• Others/custom: 17%

The PostgreSQL + pgvector adoption is the dark horse here. Developers are realizing they don't need specialized infrastructure for many RAG use cases. This is probably why standalone vector databases are pivoting to "AI data infrastructure" positioning.

Framework/Orchestration:

• LangChain: 38%

• LlamaIndex: 22%

• Custom/in-house: 26%

• Haystack, Semantic Kernel, others: 14%

What surprised me? The "custom" category is growing. Early-stage startups that launched with LangChain are ripping it out. Not because LangChain is bad it's incredibly useful for prototyping but because production needs are diverging from what general frameworks offer. This mirrors every platform evolution: scaffolding helps you start, but you eventually need control.

Deployment:

• Vercel: 29%

• AWS: 24%

• Self-hosted/VPC: 18%

• Modal, Replicate, Hugging Face: 16%

• Others: 13%

Vercel's dominance in the deployment layer is wild to me. It speaks to how many AI startups are essentially web apps with an AI backend, built by teams coming from the Next.js/React ecosystem.

Business Models: Three Patterns Dominate

1. Usage-Based Pricing (47%)
Pay per API call, per document processed, per agent run. This makes sense for PLG motion and aligns costs with value. But it creates unpredictability for customers and makes CAC recovery slower.

2. Seat-Based SaaS (31%)
Traditional per-user-per-month pricing. Works for tools that integrate into workflows (dev tools, collaboration platforms). Predictable but doesn't capture value from power users.

3. Hybrid Models (22%)
Base fee + usage, or tiered plans with usage caps. This is becoming more common as companies try to balance predictability and scalability.

The interesting thing? Almost no one is profitable. That's not unique to AI, but the burn rate is higher because of compute costs. A typical YC-backed AI startup I spoke with (anonymized) was spending $15-30K/month on LLM APIs alone at 1,000 users. Their revenue? $8K MRR.

The math only works if you believe you can either: (a) raise prices significantly, (b) reduce costs through efficiency/caching, or (c) get to massive scale quickly. Most are betting on (b) and (c).

Go-To-Market: The Playbook Is Converging

Nearly everyone is following the same GTM motion:

1. Build in public on Twitter/LinkedIn

2. Launch on Product Hunt (aim for top 5)

3. Developer community presence (Discord, Slack communities)

4. Content marketing (lots of technical blog posts)

5. Direct outreach to design partners

The problem? When everyone does the same thing, nothing stands out. The most successful launches I saw had one differentiator: a unique distribution channel or community they could tap into.

One startup (in the legal AI space) launched exclusively through a partnership with a law school's tech clinic. Another (dev tools) got early traction by solving a pain point for a specific open-source community and becoming the "official" solution there.

The lesson: if your GTM is "post on Twitter and hope," you're competing with 99 other startups doing the same thing this month.

What's Working vs. What's Not

Working:

• Narrow, specific use cases. The vertical AI plays with clear ROI are getting traction. A tool that automates medical coding for gastroenterology practices is selling. A general "AI assistant for healthcare" is not.

• Developer tools with free tiers. PLG motion works in AI tooling. Observability platforms that let developers start free and upgrade as they scale are growing sustainably.

• Workflow automation, not just chat. Products that actually do something (trigger actions, update databases, send emails) are more valuable than fancy chatbots.

Not Working:

• Generic RAG platforms. There are 15 companies building "ChatGPT for your company docs" and none have meaningful differentiation. The switching costs are low, and customers are price-sensitive.

• Agent platforms without clear use cases. "Build any agent you want!" sounds good but creates paradox of choice. The winners are those that ship pre-built agents for specific workflows.

• Consumer AI apps without retention mechanics. Viral launches mean nothing if users don't come back. Most consumer AI tools I tracked had week-2 retention below 10%. The exceptions? Tools tied to existing workflows (email, calendar, note-taking).

The Gaps I'm Seeing

After reviewing all these companies, here's where I think there are real opportunities:

1. AI Ops for Non-Developers
Most AI dev tools are built for engineers. But the people actually deploying AI in companies are often product managers, ops teams, or business analysts. There's a gap for "Retool for AI" that's truly no-code.

2. Multi-Agent Orchestration at Scale
Everyone's building single-agent systems. The tooling for managing dozens or hundreds of agents with different specializations, coordinating them, and debugging their interactions barely exists.

3. Cost Management & Optimization
As AI apps scale, compute costs become existential. I'm surprised there aren't more startups focused purely on reducing LLM costs through intelligent caching, prompt optimization, or model routing.

4. Privacy-First AI Infrastructure
With EU regulations and enterprise security requirements, there's demand for AI infrastructure that provably keeps data isolated. On-prem, end-to-end encryption, audit logs. It's not sexy, but it's necessary.

5. Evaluation & Testing
This is slowly being addressed, but the tooling is still primitive. How do you regression-test an LLM feature? How do you know if your prompt changes improved things across 1,000 edge cases? This is hard, important, and under-invested.

Contrarian Takes I Can't Shake

Take 1: The Agent Hype Is Overblown
I think we're calling too many things "agents." An LLM with tool use isn't an agent it's a chatbot with plugins. Real agency requires learning, adaptation, and autonomous goal-setting. Most "agent platforms" are just glorified workflow automation with LLMs. That's useful! But let's not pretend it's AGI.

Take 2: Most Startups Should Not Be Building Custom Models
The number of companies fine-tuning models when they don't need to is staggering. You don't need a custom model. You need better prompts, better context, and better UX. Fine-tuning should be a last resort, not a default.

Take 3: Vertical Integration Will Win
I used to think we'd see horizontal AI platforms dominate (like AWS for cloud). Now I think vertical integration owning the full stack for a specific use case will win. The startup that builds the best AI legal research tool won't use a generic RAG platform. They'll build custom retrieval, custom ranking, custom UI, all optimized for legal workflows.

What's Coming in the Next Six Months

Based on the patterns I'm seeing, here's what I think happens by Q2 2025:

1. Consolidation in RAG/Search
Half of these companies will either shut down or pivot. The market can't support 22 similar RAG platforms. Survivors will be those with unique data sources or vertical focus.

2. Agent Platforms Mature (or Die)
We'll start seeing real production deployments of multi-agent systems, or we'll see the whole category revealed as vaporware. My bet: both. A few platforms will get real traction in specific verticals (sales, customer support), while most generic platforms struggle.

3. Model Costs Drop, Margins Compress
As inference gets cheaper and open-source models improve, the unit economics of AI apps will compress. Startups won't be able to charge the same premiums. This will force everyone to move up-stack into workflow value, not just LLM access.

4. Enterprise AI Ops Boom
Large companies are realizing they can't just dump AI tools into production without governance, security, and compliance. The tooling for "AI at scale" in enterprises will become a major category.

5. Developer Fatigue Sets In
The pace of AI tool releases is exhausting. Developers will consolidate around fewer, better tools rather than trying every new framework. This will benefit incumbents who execute well.

Final Thoughts

At Weam.ai, we're betting on infrastructure making it easier for developers to build, deploy, and scale AI applications. But this analysis has made me question a lot of assumptions.

The biggest one? I thought we'd see more fundamental innovation. Instead, most startups are building variations on the same themes with slightly different positioning. That's not a criticism it's how markets work. The first wave of any platform shift is about execution and distribution, not invention.

But it makes me wonder: where are the truly novel ideas? Where are the teams building things that don't fit into these categories?

If you're building something weird, something that doesn't fit the agent/RAG/dev-tool/vertical-app boxes I want to hear about it. Because the most interesting company in this list of 100 might not be in the list at all. It might be the one that makes us rethink the categories entirely.

What patterns are you seeing in AI startups? What am I missing?

---