The Late-Mover Advantage in AI Applications

Knowledge work is changing faster than most people realize. Software development is the clearest example. The best engineers are moving from writing code themselves to managing AI-driven workflows: thinking in systems, diagnosing agent failures, and coordinating many parallel processes.

But this shift is not limited to coding. It is happening across knowledge work. In many forward-looking industries, workers are already spending more time in ChatGPT or Claude than in dashboards, CRMs, or ticketing systems.

So what does this mean for application software?

A basic assumption before we answer it: successful application software, AI-native or not, still needs to be designed around the user. The challenge is that the user is changing quickly. In some cases, the number of users may also shrink. However, domains with strong human liability and accountability systems will continue to require human workers for at least the next decade.

Some examples are obvious: doctors, lawyers, CPAs, and other licensed professionals. But accountability systems also exist in sales, consulting, HR, and similar functions. If something goes wrong, a company still needs to know who is responsible. These users will need software that helps them get work done while staying accountable.

The sales executive of tomorrow will build GTM agents, review their outputs, and intervene when the logic breaks down.

The lawyer will spend less time manually synthesizing case law and more time negotiating, advising, and assessing what agents have surfaced.

The accountant will move from executing reconciliations to understanding where agents make mistakes and how to design safeguards against fraud.

Across these examples, the shift is the same: AI absorbs routine execution, while judgment, communication, accountability, and deep contextual understanding become more important.

Three human skills will matter more in this world:

• The ability to launch and manage agent workflows
• Strong relationship-building for high-stakes decisions
• Deep domain expertise to debug and validate agent outputs

Deep domain expertise to debug and validate agent outputs

Everything else is at risk of compression.

The biggest opportunity in application software is to build for analytical humans working with large numbers of agents in their respective domains.

Anthropic, Cursor, and OpenAI will likely continue building this kind of tooling for software developers. But there is a large opportunity for application startups in other domains.

Frontier labs will try to compete across categories with significant resources. But their advantage is strongest in fully verifiable domains like coding. In less verifiable domains, the opportunity for application software startups is clearer to perceive.

DNA of the eventual application winners

The winning applications of the future will help users launch, manage, and customize agent workflows in their own domains.

I expect the eventual winners in enterprise AI applications to share a few characteristics:

• App builders, not static apps: Winning companies will look less like traditional SaaS products and more like platforms that let users build workflows and mini-applications suited to their needs.
• Generative UI: Interfaces will be generated dynamically and tailored to each user with minimal human oversight.
• Integrated systems of record: Core databases must be built into the product. Future systems of record may be simple, low-cost databases with customizable input and output layers.
• Self-evolving systems of action: Every agent action will be recorded for auditability and learning. Over time, this creates a customer-specific system of action that becomes difficult to replace.
• Foundation models as commodity infrastructure: Applications will support plug-and-play model selection based on accuracy, cost, and latency. They will also tune smaller models where useful, reducing dependence on any single model provider.
  A common mistake is asking, “How do I build my own domain-specific model to compete with frontier labs?” But there does not need to be one end-to-end model for every domain. Some tasks need domain-specific tuning; many do not. The real edge lies in knowing which tasks deserve that investment.
• Domain-specific retrieval logic: Domain-specific agents will know which sources matter most and how to retrieve, store, and process the most relevant datasets efficiently.
• Reliability through in-house evaluations: Winning companies will use domain experts to help customers evaluate performance and, in some cases, offer bespoke evaluation services.

‍

Most popular AI application companies today do not satisfy all these criteria. That is likely because the shift from functional expert to agent manager is still early.
As a result, many first-wave AI application startups may not have the winning products of the future. Some will relaunch. Others may struggle against a second generation of applications built for the new user.

Moats for AI-native application software

Traditional software moats are becoming weaker.

• Workflow depth: New enterprise users will want custom workflows with simple UIs, while most tasks run in the background.
• Integrations: Integration code is being automated. Computer-use agents also reduce the defensibility of even complex integrations.
• Customer data: Customer data in most incumbent SaaS platforms is relatively easy to export and difficult to lock in.

Winning application companies will need different moats.

Proprietary industry data

Data can be a moat, but only certain kinds create real differentiation.

Customer-specific data is often weak as a broad moat because it may not be allowed for training, especially in enterprise contexts. It is also usually relevant only to the company that produced it.

Two types of data can create stronger advantages:

• Industry-wide data collection: Examples include Crunchbase for private financings, Apollo for buyer lead datasets, or AlphaSense for company filings. Every dollar spent collecting this data serves many customers. Over time, that can become a moat, especially when agents are built on top of it.
• Data with opt-in network effects: This is more likely in categories with a large SMB long tail, where customers may be more willing to share data with a broader network. One example could be an AI-native communication product for procurement and logistics, where every new customer or supplier adds more data for better matching.

Domain-specific models

In non-verifiable domains, the “best” model depends not only on the amount of data but also on its quality. Deep-domain startups may find better post-training approaches than general-purpose AI labs.

A strong example may eventually appear in coding itself. Watch whether Cursor can develop a coding model comparable to Anthropic’s. If Cursor can reach even 90% of that quality, the opportunity becomes more credible in other domains too.

At the early stage, we should assume a few things:

• Domain-specific post-training will be necessary only for a small set of use cases.
• Application companies can still get started without it.
• For many use cases, strong application companies will use cheap-to-run open-source models.
• Model switching will be abstracted at the infrastructure layer, and users should not need to think about it. The same way users do not care which route Zoom uses to transfer audio and video packets, they should not care which model is being used behind the scenes.

Proprietary systems of action

Software companies can still hand over customer data when required, such as when a customer wants to switch vendors. But if a product captures action trails that make its agents more effective, those trails may not be easily reusable by a competitor.

Traditional SaaS exposed much of its complexity in the UI. New AI-native SaaS will hide much of its complexity in the underlying action and embedding layer.

—--------------------

For an early-stage founder building a B2B software application company today, many categories may look overfunded and hard to enter from the outside. But the rules have changed.

The user persona is new. The workflows are new. The strongest moats in an AI-abundant world are also different.

On closer inspection, the opportunity is still open. It’s time to capitalize on the late-mover advantage.

LinkedIn DMs open for thoughts and feedback. Write to me at [email protected].

‍