Industry executives and experts share their predictions for 2026. Read them in this 18th annual VMblog.com series exclusive.
By Jeff Kuo, CEO, Ragic
By 2026, the industry will finally recognize that bigger isn’t always better when it comes to AI. For years, the trajectory of artificial intelligence has been simple: bigger is better. Large language models (LLMs) have grown exponentially, boasting billions of parameters and delivering impressive leaps in reasoning, summarization, and code generation. But as AI moves from research labs to real-world business applications, the limits of “bigger” are becoming increasingly clear.
Latency, cost, and lack of control are proving that massive models are not always the right tools for structured enterprise problems. In 2026, the next frontier in AI won’t be about scaling size-it will be about scaling efficiency through smaller, faster, and more specialized models.
When Smaller is Smarter
Take AI-powered customer support for e-commerce, one of the most common business use cases for generative AI. At first glance, it seems ideal to deploy a large model-say GPT-5 Thinking or Sonnet 4.5-to handle every part of an interaction: reading customer messages, interpreting tone, checking inventory, processing refunds, and generating empathetic responses.
But once deployed, three critical issues emerge. First, response time: large models can be several times slower than smaller ones, which is unacceptable in a real-time chat environment. Second, cost: running a large model can be ten times more expensive than a lightweight one on the same query. Third, consistency: big models may be correct 90% of the time, but because they’re so complex, improving that final 10% can be nearly impossible.
Using a massive LLM for a structured workflow is like using a sledgehammer to crack a nut. What businesses need instead are nimble, modular systems made up of specialized models-each doing one job extremely well.
The Bigger Brain Fallacy
Over the past five years, AI development has focused on building general-purpose “thinking” models designed for open-ended reasoning. These systems excel at creative writing, idea generation, and freeform conversation. Yet, the majority of business use cases-like CRM, ERP, and e-commerce workflows-don’t require open-ended reasoning. Their problem spaces are clearly defined.
In most enterprise settings, AI doesn’t need to “think” creatively; it needs to execute reliably. For example, a customer support workflow is not about discovery-it’s about applying known rules to predictable situations. What businesses value most is predictability, transparency, and control-not an opaque genius that might hallucinate under pressure.
Modular AI: Divide and Conquer
A more efficient approach is to decompose a complex problem into a series of smaller, well-defined tasks, each handled by a lightweight model. These can be combined like building blocks into a composable system that’s faster, cheaper, and easier to maintain.
Consider the same e-commerce support example:
- Intent classification: A small model identifies what the customer wants-refund, order status, or product info.
- Policy enforcement: Another model checks store policies and decides whether to approve, reject, or escalate the request.
- Data interaction: A third model updates or retrieves order data from the database.
- Response generation: Finally, a lightweight text model drafts the response-or even uses a simple template.
Each model is small, fast, and inexpensive. Together, they can reduce processing time by up to 70% and cut costs by half. Because each component is modular, developers can track performance, identify weak points, and replace underperforming models without retraining an entire system.
A Return to Classic Engineering
This modular approach feels familiar-it’s a return to classic software principles. Each AI component functions like a microservice, complete with measurable metrics such as latency, cost per token, and accuracy. Models can be monitored, versioned, and swapped out independently, aligning AI development with modern DevOps and continuous integration pipelines.
Companies with well-defined AI workloads-such as Duolingo, which uses AI for language tutoring-are already moving toward this model. Their systems rely more on structured, repeatable tasks than on open-ended creativity, making them ideal candidates for modular AI architectures.
The Future: Networks, Not Monoliths
As AI enters the DevOps era, the focus will shift from model size to system design. The most advanced AI products of 2026 won’t be defined by the number of parameters they use but by how intelligently they orchestrate many smaller, purpose-built models to achieve business outcomes.
The future of AI won’t be powered by a single giant brain-it will be built on networks of distributed micro-intelligences working together at machine speed.
##
ABOUT THE AUTHOR
Jeff Kuo is the CEO of Ragic and has been working in the tech industry since 2003. From 2003 to 2008, they worked as a Developer for Springsoft, where they were responsible for the implementation and maintenance of the Oracle ERP system, as well as the design and development of web applications such as Quotation System, Bug Tracking System, Employee Portal, Customer Support System, and License Management System. In 2008, they founded Ragic.
Jeff Kuo attended National Taiwan University from 1997 to 2001, where they earned a Bachelor’s degree in Information Management. Jeff then attended National Chiao Tung University from 2001 to 2003, where they earned a Master’s degree in Information Management.
