Not All Software is Equal

“We tend to overestimate the effect of a technology in the short run and underestimate the effect in the long run.”¹

Amara’s Law, as it has become known, feels particularly relevant again some fifty years since its origin. AI capabilities are improving with startling speed. Tasks that recently appeared specialist, cumbersome or expensive can now be performed in seconds by systems available to almost anyone with an internet connection.

But capability is not the same thing as economic impact. Markets tend to deal in narratives of the extremes, and in recent months software has been cast in precisely this way: if a process can be defined, it can be handed over to AI; and if it can be handed over to AI, the software around it must be at risk. This is something my colleague Quentin Macfarlane explored in his recent newsletter (It’s a Kind of Magic).

In this newsletter, we explore the main ways in which AI could put pressure on software companies and the structural defences that may separate the more vulnerable from the better protected. At Seilern, our research process is focused on finding companies with durable competitive advantages. In software, that means looking beyond the interface to understand where the value really sits and how hard it is to replace. Contrary to the current narrative, we think that in some cases, rather than bypassing the incumbent, AI could deepen and extend the value they deliver within their customers’ workflows.

Where the pressure may come from

Part of the difficulty in discussing AI risk is that the market often treats both “software” and “AI disruption” as though they were single, clearly defined things. Software covers a wide variety of businesses serving different industries, customers and end markets from Microsoft’s general-purpose productivity tools to Veeva’s software supporting clinical trials, to Autodesk helping architects and engineers design buildings. These are multifaceted businesses, and AI poses different risks to different parts of each model.

At Seilern, we have condensed the potential risks into seven main vectors, which together help us assess where a company may be genuinely vulnerable and where the market may be overreacting. In this newsletter, we focus on two that seem particularly relevant to the recent sell-off: the risk that the product becomes less visible, and the risk that parts of the product become easier to copy.

One of the clearest risks is what we might call the invisible product problem. Part of a software product’s value has traditionally come from being the place employees log into every day. But if AI assistants begin carrying out more tasks in the background, users may spend less time in the software itself. Instead of clicking through menus, they simply ask the assistant to do the work.

If that happens, the advantage of owning the daily user interface starts to weaken. The software may still be used, but it becomes less visible. Over time, a product that users touch less often will find it harder to reinforce its value, harder to cross-sell additional features and harder to maintain the same direct relationship with the people using it. In the most exposed cases, the software risks becoming infrastructure in the background while another layer controls the customer interaction.

Another risk is that some features may become easier to copy. This is particularly relevant where a product’s higher-value functionality depends on generating content, classifying data or producing “good enough” outputs. If AI makes those capabilities cheaper and more widely available, then features that once felt differentiated may become easier for competitors to replicate. What was previously a premium function can begin to look more like a commodity.

This does not mean all software features lose their value. The key distinction is whether the feature sits on top of the workflow or is embedded within something harder to replace. A text-generation tool, a simple assistant or a generic summarisation feature may be easier to copy. A product that combines those capabilities with proprietary data, specialised rules, trusted outputs or a deeply embedded workflow is much better protected.

Where the defences are strongest

The strongest version of the bear case is that several forces hit a business at once. For the most exposed software companies this could be terminal. But not all software is equally exposed, and the most durable businesses share several characteristics. Some software is responsible for orchestrating a mission-critical business process, rather than merely sitting on top of it. Here the elegance of the interface is much less important than how deeply embedded the software is within the workflow.

SAP is a good example. When a pharmaceutical company buys ingredients from an overseas supplier, the transaction may look simple from the outside. In reality, the system must check sanctions lists, confirm budgets, apply import duties, arrange logistics tracking, coordinate quality inspections and update production schedules, often immediately and behind the scenes. If any part fails, the process can grind to a halt. This workflow integration will remain paramount regardless of whether it is a human or an agent carrying out the task.

A second source of protection is specialised knowledge. General-purpose AI can be impressive, but many workflows depend on accumulated expertise built around very specific use cases.

Cadence’s semiconductor design software is shaped by decades of engineering and design experience. Its tools contain thousands of technical rules, checks and exceptions. Furthermore, the cost of getting a chip design wrong is extremely high: failed production runs, expensive redesigns and significant delays. It is hard to see a new AI company replicating the deep knowledge of Cadence. And even if they can, the risk of making a mistake is so high that customers are unlikely to shift from the incumbents, especially if those incumbents are incorporating AI tools themselves.

This same dynamic applies more broadly. The better software businesses are not defended by a single moat, but by several overlapping ones. AI may alter how customers interact with these products, and in some cases, it may change how value is captured. But where a company already owns an important place in the customer’s workflow, AI is often more likely to be incorporated into the existing system than to replace it outright.

From defence to opportunity

Those same characteristics can also become a route to monetising AI. An incumbent with trusted data, established distribution, a track record of delivering results and an embedded customer relationship does not need to persuade customers to adopt an entirely new system. It can build AI into products they already use, where the value of greater speed, accuracy or productivity is immediately apparent.

Microsoft is the clearest example. It already sells to almost every large enterprise, and its products sit inside the daily routines of knowledge workers. Rather than being bypassed by AI, Microsoft has embedded it into Word, Excel, Teams, Outlook, GitHub and the wider enterprise workflow, charging premiums for new Copilot tiers and monetising through a mix of per-user and usage-based pricing. Microsoft’s advantage is being able to place AI capabilities where customers already spend their time, turning productivity gains into incremental revenue across a very large installed base.

RELX illustrates a different but related opportunity: using AI to increase the value of trusted, proprietary content. Its LexisNexis platform already sits within a high-stakes legal workflow where customers need answers they can verify and defend. AI can help lawyers search, summarise, compare and reason through material more efficiently, but only if the output remains grounded in reliable, traceable sources. In this setting, RELX’s advantage is not just the ability to generate an answer, but to connect that answer back to authoritative legal content.

For the most-protected incumbents, integrating AI can therefore be more than a defensive necessity. It can make existing products more useful, automate low-value tasks, improve customer productivity and deepen the company’s role in the workflow. The same characteristics that protect a business from disruption can also give it a privileged route to monetise the technology.

What this means for our funds

AI will create winners and losers in software, and the risks should not be dismissed. Some companies may face real pressure as products become less visible, revenue models are challenged and competitive barriers weaken. But in our view, the majority of the software companies we own are well protected by the depth of their workflows, the trust they have built with customers, and the specialised knowledge embedded in their products. For these businesses, AI is more likely to be absorbed into the existing value proposition than to represent an existential threat.

This brings us back to Amara’s Law. The market has, at times, treated the sector as though the risks were uniform, perhaps overestimating the immediacy of AI’s impact while still being right that the long-term implications may be significant. We do not think those implications will be felt evenly. In some cases, AI may pressure economics. In others, it may strengthen the incumbent’s product, improve customer outcomes and create new avenues for growth. That distinction is central to how we are assessing our software holdings today.

¹ Attributed to Roy Amara, an American researcher and former president of the Institute for the Future (IFTF). Though not formally published by him, the adage has been widely attributed to him since the mid-1970s and has been popularised through second-hand attribution since then.