Complete Guide to Building a Defensible AI Business

In today’s fast-paced tech landscape, artificial intelligence (AI) innovation has become both a goldmine and a battlefield. For engineers, CTOs, and tech professionals aiming to transition into leadership roles, understanding what it takes to build a defensible AI business is crucial - not just for career growth but for creating lasting impact.

Armis Gritsunis, founder of Swirl AI and a seasoned AI professional, joined the Chain of Thought podcast to share his expertise on constructing robust AI solutions, navigating the complex enterprise ecosystem, and fostering skills that advance the future of AI. This article distills key takeaways from the discussion, offering actionable insights for ambitious professionals aspiring to thrive in the evolving AI space.

The Core Challenge of AI Businesses in Enterprise

One of the most striking points emphasized by Armis is the unique challenge of building AI infrastructure for enterprises. Unlike consumer-facing products, enterprise solutions demand a high degree of maturity, security, and reliability. These aren’t environments where "vibe coding" or quick demos can cut it; enterprises expect systems that deliver scalability, stability, and compliance.

Armis notes that enterprise buyers are highly particular, often benchmarking new tools against competitors. This creates a crowded, competitive market where startups must excel on both technical and business fronts. Success is not just about having a good idea - it’s about flawless execution combined with strong distribution capabilities.

Why Enterprises Struggle with Open Source

While open-source tools have made significant contributions to AI development, Armis highlights the challenges enterprises face when adopting them:

• Immature solutions: Open-source tools often lack the polish and enterprise-grade features - such as robust security protocols, role-based access control, and on-premise deployment options - that businesses demand.
• Support limitations: Enterprises require dedicated support and ongoing maintenance, which isn’t always feasible with open-source tools unless paired with a support team.

This is why many open-source startups fail to gain traction in the enterprise space unless they back their products with strong engineering teams and significant funding.

The Rise of Agentic Systems and Context Engineering

Agentic systems - AI systems designed to autonomously perform tasks with minimal human intervention - are becoming a hot topic in AI development. However, both Armis and the host, Connor Bronston, caution that the current hype surrounding agents has led some builders to prioritize flash over function.

Armis emphasizes the importance of context engineering, a critical yet often overlooked component of building effective agentic systems. Here’s why it matters:

• Dealing with multi-turn conversations: Agents can easily overload their context windows when handling complex, multi-step tasks. Without proper context compression and optimization, the system becomes inefficient and prone to failure.
• Storing and retrieving relevant actions: Discarding unnecessary information and focusing only on key actions ensures that agents can perform reliably without wasting computational resources.
• Avoiding overreliance on abstraction layers: Many developers lean heavily on high-level tools and orchestrators, which, while convenient, can obscure core fundamentals and lead to long-term inefficiencies.

As Armis explains, "If you’re building multi-turn agentic systems, you’ve felt the pain of exploding context windows." The solution isn’t just better tools but better engineering practices that prioritize fundamentals over shortcuts.

Observability and Eval-Driven Development

Another key theme of the discussion was the importance of observability and evaluation-driven development in creating reliable AI systems. Observability allows engineers to monitor, analyze, and improve AI performance throughout its lifecycle, while evaluation (eval) mechanisms ensure the system meets its intended goals.

Armis argues that many AI teams still fail to adopt eval-driven approaches, leading to wasted resources and unsatisfactory outcomes. "Too many teams are building agentic systems in isolation for months, only to release something that doesn’t solve a real business problem", he warns.

Practical Tips for Better Eval Practices:

1. Start small and iterate quickly: Don’t wait five months to roll out a system. Launch an MVP early and gather user feedback to refine your approach.
2. Create meaningful eval datasets: Many projects stall because teams underestimate the time and effort required to build effective datasets for evaluation.
3. Focus on core metrics: Instrument your system to track not just accuracy but also reliability, latency, and other critical KPIs.

Opportunities in AI Engineering

Armis’s insights also touched on the broader future of AI engineering, particularly the tension between innovation and practicality. While cutting-edge research drives the industry forward, many fundamental skills - like data engineering and structured system design - remain undervalued.

The Overlooked Role of Data Engineering

Data engineering, often relegated to the background, is the backbone of any successful AI system. Armis notes that while AI engineering focuses on building agentic systems, data engineering ensures the AI has the quality inputs needed to function effectively. Unfortunately, it’s not as "sexy" as AI development and often lacks visibility and investment.

For technical leaders, this presents an opportunity: those who can bridge the gap between data engineering and AI application development stand to deliver massive value to their organizations.

Predictions for the Next Six Months in AI

Given the rapid pace of change, what’s next for AI? Armis offers tempered predictions:

1. Slower innovation in LLMs: While the early days of large language models (LLMs) saw dramatic leaps in capability, improvements are beginning to plateau. Future advancements may require entirely new architectures rather than incremental updates.
2. Emerging dominance of self-improving agents: The concept of agents that can autonomously rewrite their own code is gaining traction, though practical adoption remains years away.
3. Continued challenges with distributed agents: Despite the excitement around multi-agent systems, enterprises are unlikely to implement them at scale in the near term due to the complexity of observability and evaluation.

As the industry matures, the focus will increasingly shift from "cool demos" to solving real business problems - and that’s where the best engineers and leaders will shine.

Key Takeaways

• Enterprise AI demands maturity: Security, scalability, and reliability are non-negotiable for enterprise adoption, making it crucial to build robust systems from the ground up.
• Open source has limits: While valuable, open-source solutions often require significant engineering and support to meet enterprise needs.
• Context engineering is essential: Building effective agentic systems isn’t just about prompts - it’s about managing and optimizing the entire context of interactions.
• Observability fuels improvement: Instrument your systems from day one and adopt evaluation-driven practices to ensure long-term success.
• Pivoting is key to success: Founders and engineers must be prepared to adapt quickly to changing market conditions and user feedback.
• Data engineering deserves more attention: High-quality AI starts with high-quality data pipelines, an often-underappreciated aspect of system design.
• Agent-based systems are still early-stage: While promising, distributed multi-agent systems face significant challenges and will require time to mature.
• Define defensibility carefully: In an era where models and tools are easier to replicate, focus on creating clear value propositions and strong distribution strategies.

Final Thoughts

For technical professionals transitioning into leadership roles, the path to success lies in balancing innovation with fundamentals. As Armis puts it, "Fundamentals are important. The industry may be moving fast, but a strong foundation is what ensures long-term success." Whether you’re building your first agentic system or strategizing for enterprise-scale AI, focusing on fundamentals like context engineering, observability, and data pipelines will set you apart in a rapidly evolving field.

The next decade of AI will be as much about execution as it is about invention. For those ready to lead, the opportunity is immense.

Source: "From Demo to Defensibility: How to Build an AI Business that Lasts | Aurimas Griciūnas" - Galileo, YouTube, Aug 27, 2025 - https://www.youtube.com/watch?v=lhkXjNGTkQA

Use: Embedded for reference. Brief quotes used for commentary/review.