Artificial intelligence has arrived in the software business not as a feature but as a force acting on the structure of demand itself. It changes what a product does, and it changes what a company is for. The firms we study most closely are not those announcing AI roadmaps; they are those whose underlying economics — the reasons a customer stays, the work that would have to be redone to leave — are being quietly reshaped by it. That distinction matters, because the same technology that deepens one company’s position dissolves another’s.

We have always underwritten durability rather than novelty. The question we put to any software business is narrow: why will this revenue still be here in five years, and what would it cost a customer to make it stop? AI does not change the question. It changes the answers, and it changes them in both directions at once.

Where AI compounds a moat

Durable advantage emerges where AI meets something that cannot be rebuilt at will. Three sources hold up best under our scrutiny.

  • Proprietary data. Models are widely available; the data used to train them on a specific use case is not. A system that has recorded years of a bank’s transactions, a hospital’s records, or a manufacturer’s operations does not merely store that history — it owns the context required to make a model useful. A general model trained on the open web cannot replicate structured, permissioned data, and a competitor cannot acquire it without first acquiring the relationship that produced it.
  • Deep process integration. Software that embeds AI into a business-critical workflow — rather than bolting it on as a loose feature — becomes part of the customer’s infrastructure. The system begins to anticipate, to draft, to flag, and the customer’s own processes reorganise around those outputs: staff are trained on them, downstream controls assume them, auditors come to expect them. What was once a useful tool becomes load-bearing infrastructure.
  • Rising switching costs. When an AI-backed system grows more precise with use, the distance to any alternative widens every month. The value now rests on a feedback loop between the customer’s data and the vendor’s model that a newcomer cannot start halfway through. The customer stays not out of inertia, but because leaving would measurably set them back.

In these cases AI is not a box to tick but a layer that reinforces an advantage that already exists. It is the same mechanism that has always made business-critical software defensible; AI simply tightens it.

A capability anyone can buy is not a moat. The question is never whether a company uses AI, but whether AI makes the customer’s decision to leave more expensive — and what a competitor cannot replicate even with the same model.

Where AI erodes one

The harder observation is that AI lowers the cost of building competent software. Capabilities that once required years of engineering — extraction, classification, summarisation, conversational interfaces — are increasingly available to anyone willing to call a model. For a business whose advantage rested on having built something difficult, rather than on owning something scarce, that is a genuine threat. The feature was the moat, and the feature is now a commodity. Even deeply integrated vendors are not immune: when the underlying capability becomes a commodity, value creation — and with it pricing power — shifts to whoever controls the scarce resource.

This is where measured judgement earns its keep. A company can present an impressive AI narrative while its actual defensibility is thinning, because the thing AI automates is precisely the thing customers used to pay it for. Conversely, an unglamorous business with deep workflow integration and proprietary data may be strengthening even as it says little. The narrative and the economics can point in opposite directions, and our work is to tell which is which. We ask of every company where the willingness to pay actually comes from, what of it a model can reproduce more cheaply and what it cannot — and whether what remains supports a business that deserves the name.

What we conclude

None of this argues for or against AI as a category. Categories do not have moats; companies do. The thesis we hold is simpler and older than the technology: durable demand comes from being genuinely difficult to replace, and recurring revenue is only as safe as the switching cost beneath it. AI raises the bar on both counts. It rewards businesses that compound a real advantage and exposes those that mistook a head start for one.

So we watch this development closely, and we keep underwriting the same thing we always have — quality over category, defensibility over momentum. We continually assess new opportunities through that lens, and we expect the most durable of them to be companies for which AI is not the story, but the reason the existing story has become harder to dislodge.