A friend of mine runs a company that builds software for other companies. He has the same problem in most sales meetings. He shows what the system can do, the prospect leans forward and asks how it works, and my friend, who is not daft, goes slightly vague. If he explains it properly, they might go away and build it themselves.

Some of them go away and build it anyway. About a year later, a fair number rang back and asked for help.

So there is still money in knowing how to make something work. What AI has done is make the gap between a clever demo and a defensible company rather uncomfortable to look at. The tools that let your team build sophisticated software cheaply let a competitor rebuild it cheaply too, and your customers can have a go over a long weekend. The development bill has come down for everyone. That includes the people trying to put you out of business.

Your supplier has a product team too

In my previous article, I asked what would happen to a startup if its model disappeared tomorrow. There is a quieter version of that dependency. The supplier carries on answering your API calls and starts selling what you sell.

On 24 February, Anthropic released finance plugins for Claude covering equity research, investment banking, private equity and wealth management, with workflows that run across Excel and PowerPoint. If you were selling an AI equity research assistant on 23 February, you had a new slide to write for your next board meeting.

Figma had it worse. On 17 February, it announced an integration with Claude Code, letting interfaces built in Claude drop straight onto the Figma canvas. On 17 April, Anthropic launched Claude Design, which produces interface mockups and interactive prototypes. Three days earlier, Anthropic’s chief product officer had stepped down from Figma’s board, which, in hindsight, was something of a hint. Figma’s shares fell 7% on launch day. Presumably the next partnership meeting had a slightly different atmosphere.

You may also be using that provider’s tools to build the product it is about to replace. I wouldn’t hold that against anyone. I would just keep it in mind.

Somebody else’s high street

In February 2025, Chegg, which sells homework help to students, reported that its non-subscriber traffic for January was down 49%, against a decline of 8% in the second quarter of 2024. It blamed Google’s AI Overviews for answering the questions that used to send students to its site, and sued Google that same week. That is Chegg’s account, and Google disputes it. The shape of the problem reaches well beyond homework.

The platform that sends customers to your door can start answering their questions on the doorstep. You still have a website. Fewer people have a reason to visit it.

Before anyone talks about a moat, it is worth checking how much of the surrounding land you actually own.

Six months used to be a moat

A feature that once took six months of engineering may now take a fraction of that. Marvellous, until you realise those six months were most of what kept the competition out.

Customers do the same sum. Some will decide their own developers can rebuild enough of your product in-house. They will underestimate the maintenance and the awkward exceptions, as my friend’s returning clients did. Their attempt still costs him a year of sales, and he doesn’t get the year back when they ring.

So where is the moat?

It starts with knowing how to get a better result. It deepens with the things you learn that a competitor can’t pick up by watching your demo.

Take my own industry. Getting an AI to produce a plausible trade recommendation is easy to show on stage. Getting software to behave inside a broker’s real trading infrastructure is another matter. Permissions have to hold when an instruction is ambiguous; someone has to reconcile the books when a transaction fails halfway through. None of this appears in the sales presentation. It turns up in the third week of production, usually on a Monday.

A competitor can copy your screen without knowing how you handle any of that. A customer can build a prototype without ever finding out why your production version took so long.

Proprietary data helps, provided you actually have the rights to use it. A pile of public documents anyone can download offers rather less protection. Calling it a knowledge base does not make it yours.

The test is simple enough. If someone offered your visible features at half the price, would your customers stay? “We shipped first” is not an answer.

Make sure you can change the engine

Whatever the advantage is, it has to survive a change of model. 

Founders should know what they would run instead of their current provider, and whether it is good enough on their customers’ actual work. Benchmarks are not customers. For some applications, an open-weight model hosted in-house is the sensible answer.

Buying a server does not give you a moat. It does stop a supplier’s product roadmap from closing your company.

If switching providers would strip out everything your customers value, then the valuable part was never yours. You were renting it.

Whose moat does regulation dig?

This is also why I am skeptical of calls for an AI slowdown. A voluntary pause is hard to hold when competitors keep building and customers can leave. The data centres still need paying for.

Regulation can do something a voluntary pause can’t. It can make getting in more expensive.

A large provider spreads the cost of compliance across a business that already exists. A startup has to find that money before it has much revenue. Rules that make independent models prohibitively expensive would leave smaller companies even more dependent on the suppliers already eating into their margins.

There are good reasons to regulate AI. But when the biggest providers ask for more rules, I’d like to know who can afford to follow them. A requirement their legal department absorbs before lunch may be enough to stop a smaller competitor getting funded at all.

Europe’s AI Act makes some allowances for this, including exemptions for certain open-source models. Much depends on the detail. Writing rules that only the largest companies can afford to follow is an excellent way of keeping them the largest companies.

What they were paying for

Founders should assume cheaper development means more competition. The defence has to be something that grows every time you serve a customer and is still standing after the next model release.

Clients came back because copying the idea was easier than delivering the result. AI made their first attempt cheaper. It did not give them the vendor’s experience. They saved on development and spent a year finding out what they had been paying for.

The pattern is familiar. An internal team looks at a vendor’s product and concludes it could be rebuilt in a couple of sprints. Occasionally, it can. More often, the organisation pays for the lesson twice: once in the build, and again in the operational gaps that only appear in production.

The post OPINION: Your product is no longer a competitive advantage appeared first on EU-Startups.