7 Startup Ideas the Big AI Companies Won't Kill

7 Startup Ideas the Big AI Companies Won't Kill
Founders used to ask me "what if Google builds this?" Now it's "what if OpenAI builds this?" It's a fair question, and I think it has a better answer than most people expect.

When I say AI in this post I mean the big AI companies. OpenAI, Anthropic, Google and a few others. I’m not asking whether AI as a technology will change your industry. It will. I’m asking whether one of those companies will release a feature that makes your startup pointless.

They’ve told us they will, for some startups. In April 2024, Sam Altman said on the 20VC podcast that if you build a business around patching the gaps in the current model, “we’re going to steamroll you”. That’s about as clear as a warning gets.

It isn’t only tiny startups either. Anthropic released a set of plug-ins for its Claude Cowork agent on 30 January 2026, covering legal, sales, marketing and data analysis work. A few days later LegalZoom’s share price fell 19.7% in a day and Thomson Reuters fell nearly 18%, its biggest daily loss on record, according to Reuters.

Share prices aren’t the same as lost customers, and some analysts called it a sentiment problem more than a real one. But it shows how seriously the market takes the threat.

How I chose these

I’m not going to claim anything is AI-proof. Nothing is. What I looked for is a reason the big AI companies can’t or won’t build the thing.

“Can’t” usually means they’re missing something a better model doesn’t give them. A human, a dataset, a licence, a warehouse.

“Won’t” usually comes down to size. More on that in number 3.

A quick test you can apply to any idea: does your product get better or worse when the next model comes out? If a smarter model makes your product better, you’re building in the right place. If it makes your product unnecessary, you’re in the way.

I’ve written before about making a startup defensible. This is that question again, with a much bigger competitor.

1. Products where the human is the point

I’ll start with one we built. Joycards is a group video greeting card. Friends, family or workmates each record a short video message and the recipient gets them all in one card. We recently rebuilt the product.

AI could write a nicer birthday message than most of us. It could now generate a video of your mum saying it too. And that’s exactly why it’s worthless here. The value of the card is that real people stopped what they were doing and recorded something for you.

There’s research behind this. An Ohio State University study of 208 adults found that when people learned a friend had used AI to help write a message to them, they felt the friend had put in less effort and were less satisfied with the friendship. It was a hypothetical scenario and a small sample, so don’t over-read it. But it matches what most of us feel.

A product like Joycards can still use AI to improve the experience around the edges. Reminders, captions, tidying up the audio. It just can’t replace the people in the videos without destroying the product.

The honest risk for a product like Joycards isn’t OpenAI. It’s that building software is cheaper now, so more competitors will turn up. That’s ordinary competition, and you win it with a better product and a better brand.

The same logic applies to businesses built on connecting the right people. Hampton is a private network for founders. Members are vetted, then placed in a small group of other founders in their city that meets in person every month with a trained moderator.

You can ask AI for business advice at 2am and the answer will be pretty good. What it can’t be is another founder who has been through the same thing, knows your numbers, and asks you next month whether you did what you said you would. The product is the other members. A model can’t vet them, and it can’t be one of them.

That kind of business is slow to build. It’s worth nothing until the right people are in it. But that’s also why it’s hard to copy.

Why it survives: the more AI-generated content and advice there is, the more people value what comes from a real person.

2. Businesses sitting on data nobody else can get

The big models were trained mostly on what’s publicly available, and the AI companies will pay to get more of it.

Google reportedly agreed to pay Reddit around US$60 million a year for access to its content. OpenAI signed its own deal with Reddit a few months later. And in June 2025 Reddit sued Anthropic, alleging it had scraped the site without permission. Anthropic disputes that. Reddit’s content is public. Anyone can read it in a browser. If the AI companies will pay for that, and end up in court over it, data they can’t see at all is worth a lot more.

You don’t need to be Reddit. The best startup version of this is data that isn’t public and only exists because you do the work. Sensor readings from equipment you service. Outcomes from thousands of insurance claims you’ve processed. What a job really cost compared to what was quoted.

The founders I’d back here are people who’ve spent years inside an industry and know which data matters and where it lives.

Why it survives: a better model doesn’t help if it has never seen the data.

3. Niches too small for a trillion dollar company

This is the “won’t” category, and I think it’s the most underrated one.

Look at the numbers. Sam Altman has said OpenAI expected to finish 2025 on about US$20 billion in annualised revenue, with roughly US$1.4 trillion in data centre commitments over eight years, and that it plans to grow to hundreds of billions in revenue by 2030.

A company with those bills can’t spend time on a market worth $5 million a year. It isn’t that they couldn’t build it. It’s that a win there doesn’t move anything they care about.

For a founder, a $5 million a year business is life changing.

This has worked at scale. Canada’s Constellation Software has built a business with US$11.6 billion in revenue in 2025 almost entirely by buying small software companies that each serve one specific industry. Software for things most people never think about.

Think crane hire scheduling, funeral home management, or quoting for theatrical set builders. If your first reaction is “that’s tiny”, you’ve got the point.

I should be upfront that investors are nervous about this model right now. Constellation’s shares are down by about a third from their peak, largely on fears that AI makes niche software cheap to replace. Its revenue is still growing.

That fear points at the real risk here. It isn’t OpenAI building crane hire software. It’s your customer deciding a general AI assistant is good enough, or building something themselves. So a niche product has to hold things a chat window doesn’t: the customer’s records, connections to the other systems they run, and a detailed understanding of how that industry really works.

Why it survives: they need markets worth billions. You don’t.

4. Businesses that hold the licence and the liability

A model can explain a loan contract. It can’t hold a credit licence, and it can’t be sued in your place.

The AI companies are careful about this themselves. OpenAI’s usage policies prohibit providing tailored advice that requires a licence, such as legal or medical advice, without appropriate involvement from a licensed professional.

So in lending, insurance, financial advice, health and law, someone still has to be the accountable party. That someone can be a startup. Use AI for the heavy lifting inside the business and keep licensed people making the decisions that need a licence.

Why it survives: regulators want a name on the paperwork, and it has to be someone who holds the licence.

5. Marketplaces with real supply

An AI model can recommend a hotel. It doesn’t have any rooms.

When OpenAI launched apps inside ChatGPT in October 2025, the launch partners included Booking.com, Expedia, Zillow and Spotify. OpenAI didn’t try to rebuild their inventory. It plugged them in.

OpenAI then went a step further. In September 2025 it let people buy products inside ChatGPT. By March 2026 it had pulled that feature back, leaving the purchase with the merchant.

That tells you where the value sits. The assistant can be the front door, but it still needs someone who has the supply. If you’ve done the slow work of signing up one side of a market, whether that’s tutors, tradies, venues or specialist suppliers, you have something a model can’t generate.

There’s a catch, and it’s a real one. A startup isn’t Booking.com and won’t have that kind of bargaining power. The AI assistant may end up being where the customer starts, which makes you a supplier to it. And if your supply is listed on every other site too, you don’t have much. So own the supply side properly. Exclusive relationships, verified quality, things that are hard to copy.

Why it survives: supply has to be signed up one relationship at a time, and no model does that for you.

6. Anything that ends in the physical world

Software is the easiest thing for an AI company to copy. A van, a warehouse and a crew are the hardest.

I need to be careful here, because the big AI companies are moving into hardware. OpenAI has said devices and robotics are part of its plans, and Google’s parent company owns Waymo. So I’m not claiming they’ll stay out of the physical world.

Y Combinator’s latest Requests for Startups list shows where investors think the opportunity is. Several of its 11 ideas are physical: robotics supply chains, data centre equipment, and financing for all of it.

What the AI companies aren’t going to do is run a local service business. That’s the opening. You don’t need to build robots. Combine software with physical delivery. Installation, inspection, maintenance, logistics, anything where a person or a machine has to turn up. AI will make the scheduling, quoting and paperwork far better. Someone still has to turn up.

Why it survives: someone has to turn up, and it won’t be an AI company.

7. Checking the AI’s work

Nobody should mark their own homework, and that includes AI companies.

There’s a growing market in this. The UK government estimates its AI assurance sector, meaning the companies that test and verify AI systems, was worth about £1.01 billion in 2024 and could reach £18.8 billion by 2035. That’s a government projection, so I’d treat it as optimistic. Its report also makes the point that who does the checking matters as much as how it’s done.

I don’t think a startup wins the general version of this. I’d expect the big accounting and consulting firms to take it. The opening is to go narrow. Independent testing of AI systems in one industry you know well, or audit trails for automated decisions in one regulated field.

Why it survives: independence is the product, and the AI companies can’t sell that about themselves.

What these have in common

Go back through the list and the same few things keep appearing:

  • Something a better model doesn't provide. A human, private data, a licence, supply, or physical delivery.
  • A market the big companies won't bother with. Too small or too specialised to be worth their time.
  • A reason the customer needs someone independent of the AI companies. They can't be the ones checking their own work.

Nearly all of these ideas still use AI, and most use a lot of it. That isn’t a contradiction. Being hard to replace doesn’t mean avoiding AI. It means AI makes your product better instead of being your whole product.

What I’d stay away from

It’s worth being clear about the other side.

  • A prompt with a nice interface. If your product is a clever instruction to someone else's model, they can add it as a feature.
  • Fixing a current weakness in the models. This is exactly what Altman warned about. The weakness gets fixed and your reason to exist goes with it.
  • General tools for general work. Writing, summarising, meeting notes, basic research. This is home ground for the AI companies.
  • Broad professional work with no specialist angle. The legal plug-in I mentioned at the start is the warning sign. "AI for lawyers" is exposed. AI for one narrow, regulated corner of legal work, with your own data, is a different business.

A caveat

I could be wrong about some of these, and this post might date badly. Things move quickly and the AI companies have surprised everyone more than once.

None of these ideas is easy either. Hard to copy usually means hard to build. Signing up supply is slow. Licences take time and money. Niche markets need someone who knows the niche.

But I’d much prefer to be working on something hard that’s still worth owning in five years than something easy that’s a free feature by next quarter.

Got an idea you want to pressure test?

We’ve built 50+ startups with founders across fintech, health, marketplaces and B2C, including Joycards. If you’re weighing up an idea and want an experienced view on whether it’s defensible and what it would take to build, our startup development and AI development team in Sydney and Canberra can help. Get in touch and we’ll talk it through.

Let's chat

Get an obligation free web design or web development quote from an experienced local Australian development team.