The AI SaaS glut and where the gaps actually are

Every other Product Hunt launch is "ChatGPT for X". Most will be dead in six months. The wrapper gold rush is over — not because AI is done, but because the easy surface area has been strip-mined. What's left is harder, weirder, and more interesting.

The wrapper graveyard

Open any "AI tools directory" and count the duplicates. Twenty AI meeting note-takers. Forty AI resume builders. A hundred AI chatbots-for-your-docs. Most are a system prompt, a Stripe integration, and a landing page generated by another AI tool.

The economics are brutal:

• No moat. Your prompt gets reverse-engineered in a tweet. Your RAG pipeline is a LangChain tutorial.
• Margin compression. OpenAI cut prices ~80% in 18 months. Good for users. Lethal if your "product" was the markup.
• Platform risk. Every feature you ship is one OpenAI DevDay away from being a checkbox in ChatGPT.
• CAC is broken. Google and Meta ads for "AI [category]" are saturated with VC-funded competitors burning money.

If your pitch deck slide says "we use GPT-4 to..." and the next slide doesn't explain a serious data, distribution, or workflow advantage, you don't have a company. You have a feature.

What's actually saturated

Be honest about which categories are cooked:

• Generic chatbots and "chat with your PDF"
• AI writing assistants for blog posts and tweets
• Resume / cover letter generators
• Meeting transcription + summary (Otter, Fireflies, Zoom, Granola, Read, Fathom...)
• Generic image generators wrapping SDXL or Flux
• AI logo / slide / website builders
• "Notion but AI" productivity tools
• Browser-extension copilots for Gmail and LinkedIn
• AI SDR / cold email tools (this one is a bloodbath)
• Code completion for mainstream languages — GitHub owns it, Cursor owns the IDE layer

You can still win in these, but only with real distribution or a specific wedge nobody else has.

Where the gaps actually are

The interesting work is where the moat comes from something other than the model.

Regulated and high-trust verticals

The further from "general consumer" you go, the fewer competitors and the higher the willingness to pay. Healthcare documentation that survives an audit. Legal workflows with proper citation and privilege handling. Tax and accounting tools that integrate with the actual filing systems. Clinical trials. Pharma regulatory submissions. These are slow, gated, and require domain experts on the team — which is exactly why they're underbuilt.

Boring industries with offline workflows

Construction takeoffs. HVAC quoting. Dental practice management. Auto body shop estimates. Trucking dispatch. Marine surveying. The workflow is still PDFs, phone calls, and a 20-year-old Windows app. AI that ingests the PDF, calls the supplier API, and spits out a quote in 30 seconds is a real product. The hard part isn't the AI — it's understanding the workflow well enough to replace it.

Local-first and on-device AI

Llama 3, Phi-3, Qwen, and Gemma now run usefully on a laptop. Categories worth exploring:

• Privacy-sensitive document processing that never hits a cloud
• Field tools for areas with no connectivity
• Anything where API per-token costs would kill unit economics at scale
• Enterprise deployments where the security review is the actual moat

Apple's pushing this hard with on-device models. Most SaaS founders are ignoring it because their instinct is "ship a web app".

Voice that doesn't suck

Real-time voice is finally good enough — sub-500ms latency, decent interruption handling. The applications are mostly untouched outside of customer support: language tutors that actually converse, voice-driven CRMs for field sales, hands-free interfaces for trades, dictation that understands jargon. Most "AI voice agent" startups are still chasing call centres. Everything else is wide open.

Eval, observability, and infra for AI products

If you've built one production LLM feature, you know the tooling is bad. Eval is hand-rolled. Prompt versioning is a Google doc. Cost monitoring is a Grafana dashboard somebody built on a Friday. Braintrust, Langfuse, Helicone are early movers but the category is far from settled. Sell picks and shovels to people who are panic-buying picks and shovels.

Synthetic data and fine-tuning pipelines

Foundation models are commoditising. Custom models trained on proprietary data are where defensibility lives. The tooling to generate, clean, and curate fine-tuning datasets is still mostly Python notebooks. There's a real product in "give us your messy data, get back a fine-tuned model and an eval harness".

Agent infrastructure for narrow domains

"General-purpose agents" are a research problem. Narrow agents that book travel for one airline, reconcile invoices for one ERP, or manage inventory for one POS — those ship. The opportunity is in the integration layer: auth, retries, idempotency, human-in-the-loop, audit trails. Boring backend work that nobody wants to do.

Hardware-adjacent AI

Robotics is finally pulling LLMs into the control loop. Drones. Inspection. Agriculture. Warehouse picking. Vision models that run on a Jetson and do one thing well. The barrier to entry is higher (you need to touch atoms) which is exactly why it's less crowded.

Creator tools for specific mediums

Generic "AI video editor" is saturated. AI for podcast post-production with proper multitrack handling? AI for tabletop RPG GMs? AI for architectural visualisation? AI for tattoo flash? Pick a craft, learn it deeply, build for the 50,000 people who do it professionally. Small TAM, low competition, loyal users.

How to pick

A rough filter when you're evaluating an idea:

1. Can a solo founder clone this in a weekend? If yes, walk away.
2. Does the moat come from data, distribution, integrations, or domain expertise? At least one needs to be true.
3. Would the workflow exist without AI? If yes, AI is making it better. If no, you're inventing demand — much harder.
4. Will OpenAI ship this in a keynote? Anything that looks like a general-purpose feature is in their roadmap.
5. Can you charge enterprise prices? $20/mo consumer SaaS doesn't survive paid acquisition in 2025.

The unglamorous truth

The next wave of durable AI companies won't be founded by people who saw a Twitter thread and spun up a wrapper. They'll be founded by people who spent five years in an industry, know exactly where the 40-hour-per-week busywork lives, and build the thing that eats it. The AI part will be the least interesting slide in the deck.

Pick a niche you actually understand. Build something nobody can copy in a weekend. Charge enough to survive.

Everything else is noise.

@misc{ctrlsze-the-ai-saas-glut-and-where-the-gaps-actually-are-2026,
  author = {Sze},
  title  = {The AI SaaS glut and where the gaps actually are},
  year   = {2026},
  month  = {June},
  url    = {https://ctrlsze.studio/blog/the-ai-saas-glut-and-where-the-gaps-actually-are},
  note   = {CTRLSZE}
}