Part 2: Picking Your Mountain – Where to Build in the Age of AI
Picking a Corner They Can’t Touch
In Part 1, we talked about how AI has broken the old model of durable software. Code is cheap, disposability is common, and the big players – OpenAI, Google, Anthropic – have the advantage on flat ground.
So what now?
If you’re a founder, operator, or investor, the challenge is to pick a space that AI giants can’t or won’t dominate. Somewhere messy. Somewhere that needs more than a slick demo.
Let’s talk strategy.
Three Types of Defensible Niches
You’re looking for terrain that’s:
Too weird
Too complex
Or too closed-off for the chariots to enter
Here are three categories where that’s possible:
1. Data Moats (they can’t see what you see)
Private or proprietary data gives you an advantage no general model has.
You can fine-tune or prompt smarter than anyone else because you’re working with info nobody else can touch.
Example: Internal process data at a logistics company. Regulatory case notes at a law firm. Patient behaviour data in a specific clinical trial.
Build: domain-specific copilots, internal search, automation tools for heavily documented industries.
2. Distribution Moats (they can’t go where you go)
You already serve customers they don’t.
You have access, reputation, or compliance cover they won’t get.
Example: A local ERP vendor in Morocco. A medical software supplier with FDA clearance. A partner with deep roots in industrial manufacturing.
Build: AI-powered add-ons to your existing stack. Layer AI into workflows that are already entrenched.
3. Complexity Moats (they don’t want the mess)
Some workflows are too chaotic or nuanced for general models to handle out of the box.
That’s good – that friction is defensibility.
Example: Legal case triage. Construction site safety reporting. Cultural nuance in multilingual customer service.
Build: agents that assist humans, not replace them. Tools that operate in grey zones, not clean use cases.
What Not to Build
Here’s a short list of danger zones – categories where OpenAI and friends will flatten you unless you’re already dominant:
AI-powered writing tools (too many, too easy to replicate)
General search UIs (ChatGPT is absorbing them)
Low-code frontends with LLMs under the hood (becoming features, not companies)
Chatbots that aren’t deeply verticalised (unless trained on unique data or supporting mission-critical ops)
Example Plays That Work Right Now
Here are some real-world ideas that are working today by sticking to the strategy:
Cursor – AI-native code editor that goes deep on developer workflows (rumoured OpenAI target)
Windfall – AI for healthcare billing, tightly tied to messy industry-specific systems
Windsurf – niche platform for agencies, with high-friction onboarding that makes it sticky
You.com / Perplexity – trying to carve a lane in search, but already getting squeezed as OpenAI bundles shopping, ads, and search together
All of them either have a unique data position, or they operate in complex enough terrain that the general models haven’t flattened them (yet).
The Big Strategic Shift
In the 2010s:
Move fast
Go horizontal
Raise big
Own the market
In the 2020s:
Move deliberately
Go vertical
Get profitable faster
Own the niche
You’re not trying to win the world. You’re trying to find one slice of the world that AI giants won’t bother to touch – and own it.
The Hard Part Isn’t Building. It’s Choosing What to Build.
AI didn’t make software easier. It made the illusion of progress easier.
The hard part now is strategic:
Picking what to build
Picking where to build it
Picking why it’s going to matter later
If you get that right, the tools will take care of themselves.
