In the rapidly evolving world of enterprise technology, the shift from traditional software-as-a-service to AI-native applications is more than just a trend—it is a fundamental restructuring of how businesses operate. Vijay Raina, a seasoned specialist in enterprise SaaS technology and software architecture, has spent years advising founders on how to navigate the complex intersection of code and commerce. As the industry moves away from tools that simply organize data toward systems that autonomously execute tasks, Vijay provides a crucial perspective on the architectural and economic hurdles that define this new era. Our conversation explores the massive wave of investment hitting the sector, the narrowing profit margins of AI-first companies, and the vital importance of data context in building a defensible product. We dive into why the old playbook of seat-based pricing is crumbling and what founders must do to ensure their automated workflows actually solve real-world problems without compromising trust.
The scale of US investment in AI has reached staggering heights recently. How do you interpret the fact that nearly $286 billion is flowing into this sector, and what does it say about the competitive landscape for new founders?
The sheer volume of capital entering the US market is breathtaking, with private AI investment hitting $285.9 billion in 2025. When you realize that there are 1,953 newly funded AI companies in the US—which is more than ten times the count of the next closest country—it becomes clear that the gravity of innovation has shifted entirely toward these native intelligence models. This isn’t just speculative money; it’s a massive bet on a new architectural standard. For a founder, this means the barrier to entry has changed from “can you build a platform” to “can you build a platform that completes the work.” You aren’t just competing with other startups; you are competing in an ecosystem where buyers are pouring $37 billion into generative AI spending, specifically looking for applications that move the needle on productivity.
We are seeing a clear pivot where enterprise buyers prefer products that complete work rather than just helping users manage it. How does this change the way you approach software architecture?
In the old world of traditional SaaS, we built digital filing cabinets—platforms that moved data out of messy spreadsheets and into organized dashboards. A sales tool would show you a lead, but you still had to write the email; a support tool would show you a ticket, but you still had to resolve it. Architecture now has to account for the “next step” by integrating decision-making logic directly into the workflow. If an AI-native finance tool identifies an unusual transaction, it shouldn’t just flag it; it should explain why and suggest a remediation path. This requires a much deeper level of integration between the data layer and the model, ensuring the software isn’t just a passive observer but an active participant in the business process.
There is a notable financial trade-off here, with fast-growing AI companies reporting 25% gross margins compared to the 60% we see in traditional SaaS. How should founders manage this economic pressure?
This margin gap is the “hidden tax” of the AI revolution, and it’s something every architect and founder needs to look at closely. In traditional SaaS, once you build the feature, the cost to deliver it to the next thousand users is negligible, but with AI-native products, every single model call and data retrieval carries a direct inference cost. When you see margins sitting at 25%, it’s often because companies are using expensive, over-powered models for routine tasks that don’t require that much “brainpower.” To survive, founders have to get surgical about their model usage—knowing when to use a lightweight model for simple summaries and when to call in the heavy-duty infrastructure for complex reasoning. If you don’t map your model calls to the actual value delivered, your growth will eventually cannibalize your profits.
With over half of generative AI spending now moving toward applications rather than just infrastructure, what does this tell us about where the real value is being created?
It tells us that the “shovels” phase of the gold rush is starting to balance out, and people are finally building the actual mines. According to Menlo Ventures, about $19 billion of that $37 billion spend is going straight into applications, which means enterprise buyers are tired of just playing with chatbots and want tools that solve specific business problems. This is a huge signal that the market is maturing beyond the “AI-enabled” gimmick—where you just slap a chatbot on a dashboard—and toward “AI-native” solutions where the AI is the core product. For me, as a software expert, this is the most exciting part because it rewards deep industry knowledge and complex workflow integration over generic technical capability.
Speed to market seems to be the primary obsession for founders right now, with two-thirds of them prioritizing it during testing. What are the risks of moving that fast?
It’s true that 66.7% of founders are racing to get their features live, and while AI-powered code generation certainly speeds up prototyping, it also creates a false sense of security. A prototype can look brilliant in a controlled environment but fail miserably the moment it encounters “dirty” real-world data or complex security requirements. The danger is that fast development often skips over the rigorous validation needed to ensure the AI doesn’t give a “confident but wrong” answer. You can build a tool that handles routine support requests in a weekend, but if it hallucinates a refund policy that doesn’t exist, you’ve created a massive liability. Speed is a competitive advantage only if you have a plan for what the product does when it reaches its limits or returns a low-confidence result.
You’ve mentioned that “context” is the new moat for software companies. Why is that more important than the underlying AI model itself?
Almost every founder has access to the same foundational models, so the model itself isn’t a long-term advantage; it’s a commodity. The real defensibility comes from what Bessemer calls “context”—the deep workflow history, the proprietary customer data, and the specific industry integrations that a generic model simply doesn’t have. If your software understands the unique legal nuances of a specific jurisdiction or the historical transaction patterns of a particular niche, it becomes incredibly difficult to replace. The model provides the capability to think, but the context provides the “memory” and the specialized knowledge that makes the output actually useful. That’s how you build a product that stays embedded in a customer’s daily work for years.
Since AI-native products have such different cost structures, is the traditional seat-based pricing model finally dead?
Seat-based pricing is definitely under fire because it assumes that every user costs the same to support, which is no longer true. If one customer uses your AI occasionally while another runs ten thousand automated tasks a day, the cost of serving them is radically different. We are moving toward models based on usage, completed tasks, or even specific business outcomes. The goal is to align your pricing with the value the customer receives—if the tool is doing the work of a full-time employee, the price should reflect that output, not just the fact that someone logged in. This shift is essential to protecting those 25% margins and ensuring the business remains sustainable as it scales.
For sectors like healthcare or law, where the stakes are incredibly high, how do you balance automation with the need for human control?
In high-stakes environments, you can’t just let the AI run wild; you need a “human-in-the-loop” architecture that provides a safety net. This means the product should be designed to flag high-uncertainty results for manual review before they ever trigger a final decision. Founders need to define what an “acceptable result” looks like and test their products against edge cases and misleading data. If the information the AI needs is missing or outdated, the system shouldn’t guess—it should admit it doesn’t know and hand the task off to a person. Trust is the hardest thing to earn in this industry and the easiest thing to lose with one bad hallucination.
What is your forecast for the future of the SaaS development market in the US?
I believe we are entering a period where the distinction between “software” and “services” will continue to blur until they are essentially the same thing. In the next few years, I expect we will see a massive consolidation where traditional SaaS platforms that refuse to evolve into work-completion engines will be replaced by leaner, AI-native alternatives. We will also see a shift in how development services operate; it won’t be enough to just provide engineering capacity. The most successful partners will be those who can map out AI opportunities, validate data readiness, and architect solutions that manage both quality and inference costs simultaneously. The winners won’t be the ones with the most features, but the ones who can prove they deliver reliable, useful outcomes at scale.
