There's an asymmetric advantage in the AI era, and it favors software companies.
Software companies can access foundation models via API. Every model that Anthropic or OpenAI builds, software companies can buy with an API call. But the reverse isn't true—foundation model companies can't easily replicate the complex software systems that companies like Salesforce, ServiceNow, or Rippling have built.
Riding the R&D Wave
Software companies get to ride the wave of foundation model R&D investment. When Claude gets better, or GPT-5 launches, software companies immediately have access to those improvements. They don't have to train foundation models to compete—they always have the best available models at their disposal.
But foundation model companies would have to build payroll software. They'd have to get licensing in place. Master end-of-year filings. Handle money movement. Deal with insurance companies and enrollment issues. Help with escalations when things go wrong with tax filing and payments. Make sure they're calculating overtime, wage splits, and everything else correctly. Build a PEO.
Why Deep Access Matters
AI agents need deep access to systems to be effective. The deeper the access, the better they are.
When someone has a question about how their payroll was calculated, an AI assistant can peer into the payroll codebase and return answers based on the abstract syntax trees there. When you ask about how a certain value came to be in the system, it can query history APIs that aren't publicly accessible to understand where it came from.
You can't get this access via public APIs. You can't get it via the user's auth token. Controlling their browser won't get it for you. If you're not full-stack, you're going to be stuck drinking data from a straw into your AI-native thing. No one else is going to spend time building the complete, maximally-expressive read-write API for you to do stuff. Why would they?
The Rippling Example
At Rippling, if we were to integrate an AI assistant, it would need deep access to the system to be effective. When someone asks about payroll calculations, it would need to understand not just the public API, but how the calculations work internally—the logic that determines how values are computed, the rules that govern different scenarios, and the context that shapes each result.
This level of access isn't something you can build as an external agent. It requires being part of the system itself. Software companies control that access, and that creates a durable advantage that foundation model companies cannot match.
The Salesforce Example
The same dynamic applies to other enterprise software companies. Salesforce has built an incredibly complex CRM system over decades. Their AI assistant, Einstein, can access the full depth of that system—the custom objects, the workflows, the data models, the integration points.
An external agent built by Anthropic or OpenAI could access Salesforce via their public API, but it would be limited. It couldn't see the internal logic, the custom configurations, the complex relationships between objects. It would be working with a fraction of the information that Salesforce's own AI assistant has access to.
This isn't just about data access—it's about understanding the system at its deepest level. That understanding only comes from being part of the system.
The Palantir Example
Palantir demonstrates this principle in a different way. When Palantir gets inside a company, they deploy their Foundry platform deep into the organization's systems. They integrate with internal databases, legacy systems, and proprietary workflows. Once they're embedded, they can deploy AI models on top of that platform with full access to the company's data and systems.
An external foundation model company couldn't replicate this. They'd be working through APIs, limited by what the company exposes publicly. But Palantir, once inside, has the deep access needed to make AI models truly effective. They control the platform, they control the access, and they can deploy models that understand the full context of the organization.
This is the same advantage that software companies have—deep access to systems that external agents cannot match. The platform becomes the moat, and the models deployed on top of it are far more powerful than anything that could be built from the outside.
The Moat That Can't Be Crossed
This dynamic is already evident in the market. Foundation models aren't withholding capabilities from software companies—they're selling access via API. But software companies can and are withholding capabilities from foundation models by controlling access to their systems.
If this dynamic breaks down—if foundation model companies start refusing to sell access to their models to Salesforce, SAP, ServiceNow, Rippling, and others—it might change how this market evolves. But as long as there are at least three companies building foundation models that are reasonably neck-and-neck, it will be impossible for the labs to cut off access, even if they wanted to.
The foundation model companies are in a death match with each other to eke out six months of lead on underlying model performance. They're not focused on building payroll software or CRM systems. They're focused on the model.
The Window of Opportunity
All of this requires companies like Rippling, Salesforce, and ServiceNow to do the work to incorporate these capabilities and transform their products to make use of them. This is hard work, and it should be the highest priority initiative within these companies.
But there is a window for software companies to do this—arguably, a much longer window of opportunity than most people realize. The complexity of these underlying systems, their interface with the real world, and the need for deterministic correctness when dealing with payroll, insurance, and other critical functions all create barriers that foundation models cannot easily cross.
Software companies that lock down this access will have durable advantages over foundation model companies that foundation model companies cannot match. The question isn't whether foundation models will get better—they will. The question is whether software companies will build the deep integration needed to make those models truly effective within their systems.
The Asymmetric Advantage
The advantage is asymmetric. Software companies can access foundation models, but foundation models cannot access the deep internals of software systems. This creates a moat that foundation model companies cannot build a durable R&D advantage against.
Software companies that understand this and execute on it will win. They'll have the best models available, integrated deeply into systems that foundation model companies cannot replicate. That's where the advantage lies.