The Prevailing View
For years, the dominant narrative in artificial intelligence has been “scale is all you need.” The race to AI supremacy was measured in one primary metric: parameter count. From GPT-3 to Gemini, the industry has operated under the assumption that larger, more generalist models are the inevitable path to superior performance. This has led many enterprise leaders to believe their only viable strategy is to license expensive, monolithic frontier models. A recent research paper, DeepLens Diagnosis Agent: Agentic Workflow Design Lets a Small Reasoning Model Compete with Frontier LLMs, offers a powerful counterpoint to this brute-force approach, highlighting the rising importance of agentic workflows.
Our Position The race for enterprise AI value is no longer about model size; it’s about workflow intelligence. For specialized, high-stakes tasks, smaller models orchestrated by smart agentic workflows will consistently deliver superior accuracy, cost-efficiency, and control.
What the Data Actually Shows
The DeepLens paper provides the critical evidence. Researchers took a relatively small, 7-billion-parameter medical AI model and embedded it in a structured, five-stage reasoning process. This system achieved diagnostic accuracy on a complex medical benchmark that was competitive with massive, general-purpose frontier models. The key wasn’t the model’s raw power, but the intelligent scaffolding around it—the agentic workflow that broke down a complex problem into manageable, verifiable steps.
This isn’t an isolated finding. It reflects a broader industry trend where smaller, specialized models are proving highly effective for targeted use cases, a shift documented across the AI research community. The data is clear: brute-force scale is a strategy with diminishing returns. Intelligent system design is the new force multiplier.
The Real Implication
This changes the strategic calculus for every CIO and CDO. Relying solely on a few providers of large, black-box models is no longer the only path—and may in fact be a competitive liability. It creates vendor lock-in, exposes organizations to unpredictable costs, and offers little durable advantage, as any competitor can access the same API. The real, defensible moat is not the model itself, but the proprietary, task-specific workflow an organization designs around it. This is where deep domain expertise translates into a market advantage that cannot be easily replicated. Companies that master workflow design will build more accurate, efficient, and auditable AI systems at a fraction of the cost.
What to Do Instead
Instead of asking “Which is the biggest model?”, leaders should ask “What is the smartest workflow for our problem?” This requires a shift in focus and investment. First, prioritize process mapping and task decomposition over simply experimenting with prompts for a generalist model. Second, build internal capabilities in system design and orchestration, not just data science. Finally, re-evaluate build-versus-buy decisions; the feasibility of creating custom, high-performing AI systems with smaller, open-source models has dramatically increased. This is the core of modern enterprise AI—moving from renting generic intelligence to owning the logic that creates value. At Thinkia, our work on the design and deployment of multi-agent AI systems is centered on this principle: building defensible AI assets through intelligent workflow architecture, not just access to the latest model.