Building long-term winners in the AI era

2024 was the year LLMs (large language models) went mainstream. Now in early 2025 and turbo-charged by the recent launch of Deepseek with its performance-on-a-shoestring claims and low cost pricing, there’s a debate about the long-term defensibility and even necessity of these massive and expensive models.

Temps de lecture : 5 minutes

Regardless of where that debate ends up, and I don’t think we’ve seen the last iteration of the debate, 2025 looks to be the year of the apps. We’ll see transformative applications built on AI-enabled technology emerge from both the US and Europe, the kind that redefine industries in the same way Google, Salesforce, and Facebook (now Meta) did in their eras. But while LLMs and AI tech have made it easier than ever to build software applications, not everyone will build such winners—companies with the potential to revolutionise industries and stand the test of time.

With the general availability of large language models (LLMs) and their rapidly increasing capabilities, the rules for building successful software companies are evolving. Traditionally, application workflows were built as deterministic “if-this-then-that” decision trees, where a subject-matter expert designed the logic and a programmer implemented it in code. In the AI era, workflows are increasingly probabilistic, where AI not only determines the process logic but also generates the code to implement it.

This paradigm shift fundamentally changes how companies create long-term defensibility and competitive differentiation in building AI-native software applications. Many founders will indeed capitalise on the AI hype by creating superficial apps that instantly address niche problems. However, they risk quickly fading into obscurity, because they lack true defensibility. Two areas stand out as foundational for sustainable differentiation—and building the true winners of the AI era: unique and proprietary data assets, and agentic workflows.

1. Unique and proprietary data assets

Most LLMs are trained on publicly available data. While LLM providers are beginning to integrate semi-private data—such as copyrighted articles, music, and user-generated content—the models generally lack access to private data. This gap creates a significant opportunity for companies to leverage unique and proprietary data to develop applications that LLMs cannot replicate.

Proprietary data assets can be generated internally as part of a company’s business processes or curated by aggregating multiple non-public data sources, such as paid databases, subscription services, or partnerships. Companies like Kodiak Hub and Funnel exemplify this approach, curating proprietary datasets that serve as competitive moats.

Proprietary data can be leveraged in two key ways:

Fine-tuning models: Companies can use their datasets to train customised AI models tailored to specific use cases.
Building knowledge graphs: These maps of objects and their relationships enhance LLM prompts, producing more detailed, relevant, actionable, and in general better, outputs through a process called Retrieval Augmented Generation, or RAG in short

No matter what the approach, the proprietary and unique data asset is the core for creating sustainable differentiation and therefore long-term customer value.

2. Agentic workflows

Modern software development often involves integrating readily available software components to create specific functionalities. In the AI-native world, a similar approach applies: companies can stitch together workflows using AI agents, each performing specific tasks. For instance, an AI agent could automate customer onboarding by extracting data, verifying documents, and generating personalised recommendations—all seamlessly integrated into one workflow.

This requires in-depth understanding of the agents’ capabilities; however, building agentic workflows isn’t just about understanding AI capabilities—it’s about deeply comprehending the customer’s problem and innovating around the business process itself while also feeding it the right data inputs in an automated fashion. This shift prioritises business process innovation over technical cleverness. So, the advantage here lies with the company that really understands the industry into which it is selling. Put simply, understanding your Ideal Customer Profile (ICP) is as critical as ever.

Market adoption and Ideal Customer Profile (ICP)

While AI will become a ubiquitous component of all software, and AI-native applications will increase their share of overall software applications, adoption rates will vary significantly across industries and organisation types. Early adopters, unsurprisingly, will be smaller companies with fewer procurement barriers and compliance requirements. They’re also likely to be in tech-savvy verticals where people understand both the potential and the limitations of AI. In contrast, heavily regulated sectors like finance and healthcare face higher hurdles due to stringent governance and compliance standards.

If you’re building an AI-native product, targeting tech companies or other agile organisations as your initial ICP can accelerate early traction. Over time though, massive opportunities will emerge to serve more regulated industries by addressing their unique challenges—for example, through federated model training and learning, AI monitoring, compliance, and audit solutions.

Service-as-a-Software

The rise of AI code generators and task-specific AI agents has unlocked massive opportunities for automating manual business processes. This evolution is giving birth to the concept of “Service-as-a-Software.” Unlike “Software-as-a-Service,” this paradigm turns the established model on its head: instead of providing software to facilitate services, it reimagines traditional services as fully automated, scalable software agents. For example, we will likely see a huge impact in the professional services industry and how it has traditionally operated, as some tasks will be handled by software rather than people. If I were to guess, within a few years, that industry will look quite different.

Building winners for the long haul

The AI era presents unprecedented opportunities to build transformative software companies. By focusing on proprietary data assets, agentic workflows, and strategic market adoption, founders can create differentiated, defensible, and impactful businesses. The road to success requires more than technical prowess; it demands a deep understanding of customer problems, industry nuances, and a commitment to building for the long term. In an era of rapid change, these are the enduring principles that will separate winners from the rest.

Mikael Johnsson is the Cofounder & General Partner at Oxx.