The cloud era didn’t just change how we deliver software – it rewrote the rules of how we build, sell, and price software. Now, AI is doing the same, but faster. At Gradient, we’re seeing a tidal wave of startups racing to bring AI to the enterprise. This three-part series shares some of my perspectives on building, selling, and scaling AI applications for the enterprise, drawn from our portfolio companies and market observations.

Part I: The Build Edition

The Complexity Trap

In selling to enterprises, complexity isn’t a badge of honor - it’s a trap. The allure of autonomous agents is undeniable, but not every problem requires such technical complexity. The key is to build the right system that delivers immediate value. Agents can be used for open-ended problems where it’s difficult to hardcode a fixed path (think scientific research). However, the autonomous nature of agents means higher costs, and the potential for compounding errors. For well-defined tasks involving many steps, especially in industries with high transaction volume and low-risk margin (think banking), RPA or LLM workflows may offer the required predictability and consistency. Even as autonomous agents mature, I suspect some use cases will continue to be addressed with contextualized LLMs (think therapy). Good agile teams are mastering the “complexity paradox” by delivering real value without over-engineering.

Your Customer Needs Your Help

In 2025, more likely than not, your enterprise customer will need your help translating abstract use cases into production-grade applications. The AI race has shortened the tech cycles, and many enterprises lack the in-house expertise to support the transition to AI workflows. Accenture booked about as much AI revenue as OpenAI in 2024 for a reason. Recognizing this dynamic, OpenAI began hiring Forward Deployed Software Engineers (FDSEs) — technical experts embedded within customer operations to facilitate seamless integration and optimize solutions in real-time. We see similar practices at portfolio companies like Writer, LastMile, and Firstwork. There will be a broader industry movement toward more hands-on, customer-centric approaches.

In the height of the SaaS era, building custom software or providing intensive service work for individual clients was taboo because of scalability and ROI concerns. We’re in the early innings of a new cycle governed by new rules. Embracing a more tailored approach not only helps AI companies close deals faster, but getting deeply entrenched in the custom workflows also mitigates the risk of customers eventually churning to build their own solutions. As AI models become increasingly proficient at coding, FDSEs, and similar roles should gain more leverage and deliver higher-value outcomes with less effort. The true opportunity with the FDSEs, however, lies in helping customers identify and implement use cases that drive truly transformative impact. For instance, in HR, while traditional RPA might handle routine tasks like updating a home address, a well-designed agentic AI can revolutionize the function by acting as an 'HR Operations' agent that orchestrates truly company-specific end-to-end workflows, delivering configurable outcomes, and engaging in human-like interactions, unlocking entirely new possibilities for enterprise efficiency and automation.

Quality is Queen

When you’re building at the application layer in 2025, you're relying on model infrastructure that is improving at an unpredictable and unprecedented rate, with major releases happening several times a year. If you're not careful, you might spend weeks on a feature, only to find that the next AI model release automates it. And because everyone has access to great APIs and frontier large language models, your incredible product idea can be built by anyone. So how do you build a moat?

Being the first to go to market is not an AI moat. If it only took you a month to do, it will take even less time to undo. Offering a cheaper solution isn’t either. Let it be DeepSeek or something else, infrastructure costs are taking a dive every other month, barriers to building software are coming down and enterprises are nowhere near tapped out of their budgets.

The best strategy appears to be having a razor focus on quality – even if it means you need to do unscalable things at the beginning. One well-covered, VC favorite strategy is to go really narrow. Pick a very specific use case and become exceptional at it. Think Harvey, Abridge. A more controversial one is to stay on the bleeding-edge of functionality. Granola founder Chris Pedregal thinks startups should give a small number of users a “Ferrari-level product experience” by using the most expensive, cutting-edge models. It might be expensive on a per-user basis, but you probably won’t have many users at first and the cost of AI inference is decreasing exponentially – today’s cutting-edge models will be affordable commodities in a year or two. Either way, as always but more so today, speed of iteration is the best unfair advantage most startups have in providing a high-quality service. I’m a fan of Fal.ai’s fast shipping culture which allows them to offer support for most new models on the day they’re released publicly.

Error Prone Experimental Technology vs. Reliable E2E Solution

Software is evolving from deterministic to stochastic. Traditional software operates deterministically, producing identical outputs for identical inputs, whereas agents are non-deterministic, with outputs varying even for similar inputs, especially given the unpredictable nature of human language in prompts and conversations. Even the most popular copilots in the market today (think Sierra, Decagon) fail to solve customer requests 25-50% of the time. However, business logic remains structured and enterprises demand reliability. Your internal champion is not going to back out of a long term vendor relationship and upend complex corporate workflows for a service that performs only occasionally. Avoid getting stuck in the POC zone and fail forward by knowing your limitations and pulling together an end-to-end solution that delivers on your customer’s performance requirements. For example, when Cascade’s employee assistant fails to answer a question, it seamlessly escalates the query to a human expert via its own specialized ticketing system to make sure all questions are answered correctly and quickly. We’ve seen this improve employee experience by creating a more personalized space. Capturing all context-relevant communication inside the platform also builds a natural feedback loop, where a regular feed of human-labeled training data helps monitor performance, gather user feedback and improve future responses.

The Last Mile in the Enterprise is the Longest Mile

There is so much to get right besides choosing what model to use. In fact, the exceptional experiences delivered by your app will likely have little to do with AI as everyone has access to the same models. The difference between a delightful experience and a great demo that’s disappointing to use is your software craftsmanship. Depending on the customer’s needs, you may need to build specialized workflows, navigate on-prem deployments, ensure data security, etc. Success hinges on (as always but especially now) starting with a well-defined use case, delivering an absolutely delightful user experience, hitting enterprise requirements, establishing trust, and strategically positioning for upsells by delivering value. Deploying AI solutions successfully requires a deep partnership with the customers. Product depth and customer focus build defensibility. That typically comes from capturing customers’ business processes, domain knowledge, and brand voice really well.

At Gradient, we’re excited to back founders who are redefining what’s possible in enterprise AI. If you’re building in this space, we’d love to hear from you.