The four moats I check on every AI deal

The barrier to launching a startup has never been lower. It's cheaper and faster to build than at any point in history, which means more competition, at a global scale, than founders have ever faced. So, the question I keep coming back to on every AI deal is simple: if someone launches the same thing in three months, why do the customers stay?

The answer comes down to four moats. Here's what each one is, and a company from our portfolio that shows it working.

Nobody wants another tab. We all have too many apps open already. The strongest AI products embed so deeply into how someone already works that removing them would break the day.

Maeve AI does this well. It uses retailers' returns data to fix fit, and over 50% of fashion items get returned, mostly because of fit, so the problem is enormous. Maeve doesn't ask retailers to adopt a new system. It overlays on the stack they already run and reaches all the way back to design and merchandising. When we backed it, it was the largest pre-seed round for a solo female founder in the UK at the time, and nearly 3x oversubscribed.

The question here is whether the company creates data that doesn't exist anywhere in the public domain, and that gets more valuable the more the product is used.

Sanius Health built the UK's largest proprietary dataset across rare disease, chronic disease, and oncology. It pulls together health records, testing data, and wearable data into one complete medical picture of a patient, the kind of data that has sat fragmented for decades. That dataset can help discover drugs faster. A competitor can't replicate it, because the data itself is the asset.

Feedback comes in from customers, product ships improvements, customers respond, and round it goes. In AI, that loop has to run fast, and there's a harder layer to it now. As AI evolves, customer expectations shift too, so a signal that mattered three months ago may be irrelevant today. The best teams take the feedback, work out which parts are still relevant, and get only those into the product.

This is why we look so hard at how a team turns real signal into shipped product, and why it runs through the future-of-work companies we've backed, Capably AI, Cold, and Glow. None of them asks you to change how you work. They embed into your existing stack and make the work better from inside it.

If a cheaper competitor launches next quarter, what keeps your customers? That comes down to older rules than AI: brand, and a reason to stay.

The clearest version is when a startup builds something close to a fan base, the way a consumer brand does. Retention climbs, switching drops, and because customers stay, every new feature becomes a chance to sell more to people who were never going to leave. Research Grid, the first AI company we backed, in November 2024, built this through the quality of its clients. It automates clinical trials and diversifies patient recruitment, and it brought on Tier 1 pharmaceutical companies as customers early. That calibre of client, won through the strength of the product, is its own moat.

Next time an AI deal comes across your desk, run it against the four. Does it embed in the workflow, or is it another tab? Does it create data nobody else has? Can the team turn real signal into product fast? And if a cheaper copy launched tomorrow, why would the customers stay?

Very few companies have all four. The strongest have two or three working together, and that's usually enough to hold off the competition that lower barriers now guarantee. If a deal has none of them, no amount of impressive demo will save it.

I go through all four moats, and more of our portfolio, in this week's video.

Helping investors sharpen that judgement is something we’re passionate about at Arāya. It’s why we’re delighted to introduce the new Arāya Substack.

A view on where the world is heading. A challenge to the consensus. Insights from the team backing exceptional founders across the UK, Europe, the GCC and Asia.

New issues every Wednesday.

Warmly,
Rupa

P.s. When you're ready, here are 3 ways I can help: