Not Every AI Job Needs a Language Model
Not every AI task needs a language model: specialised AI can handle classification, scoring and fixed rules faster, cheaper and with more control.
- Large Language Models (LLMs) are ideal for language-based tasks like summarizing text or drafting replies, but can be slow, expensive, and inconsistent for jobs that don't require generating words.
- Many business processes require non-language answers like a classification (choosing from a list), a score, or a simple true/false decision, for which smaller, specialized AI models are more efficient.
- Effective AI systems often use a "diversified stack," combining LLMs for language tasks with other specialized models for classification, scoring, and applying fixed rules.
Last month, a San Francisco startup called TypeSafe AI released Jev, and it spread quickly on social media among people building AI systems. The interesting part is that this AI does not write sentences. It gives a decision: true or false, one choice from a list, or a ranked score.
That sounds less impressive than a chatbot until you look at how companies have used AI over the past year. Many teams put a chatbot-style AI system into a real workflow and hit the same wall. It could be slow, expensive to run, and inconsistent.
What Language Models Do Well
A language model, the kind of AI behind ChatGPT, is excellent at language work. It can summarize a report, draft an email, explain a policy, or reply to a customer in natural language. It works by predicting the next word, one word at a time.
That design is powerful when the answer should be words. But many business steps need a different kind of answer: yes or no, one category from a list, or a risk score. A language model can do these jobs, but it was built for language, so it can take longer and cost more than a smaller tool designed for the exact decision.
Example of a Real AI Agent Use Case
Take an example of customer support for banking. A customer types: “Why was my card charged twice?” The first step is to understand the message. Here, a language model is useful. The AI has to understand that the customer is asking about a duplicate card charge.
The second step is to sort the issue into one of the issue types, such as duplicate transaction, failed transfer, card limit, or fraud concern. This is a job for a sorting AI, also called a classification model, which means a tool that chooses the right box from a fixed list. This is the kind of task Jev is designed for.
The third step is to check risk. The system may need to ask: does this transaction look normal for this customer, or does it deserve extra caution? The answer could be a number, produced by a scoring model, which is an AI system that turns many signals into a score.
The fourth step is to apply the refund policy. If the rule says a confirmed duplicate charge under a certain amount can be refunded immediately, the system should follow that rule. A fixed rule gives the same answer every time, which matters in banking.
The fifth step is to reply to the customer. Here, a language model becomes useful again. It can explain the answer politely, in the customer’s language, and in a tone that feels human.
Something also has to decide which tool handles each step. That tool is often called a router, which simply means the part of the system that sends each task to the right place. In this five-step flow, only two steps truly need language. The others need a choice, a score, or a rule.
Diversified AI Stack: Words, Choices, Scores, Rules
Put together, AI starts to look like a toolbox. Some tools write. Some tools sort. Some tools score. Some tools enforce a policy exactly as written.
Smaller specialized AI systems are getting smarter and cheaper to run, and they are becoming easier to connect together. This pattern is already showing up beyond banking. In SCB 10X’s AI-Volution series, I interviewed a healthcare startup that uses separate checks inside its system: patient-care policy, standard medical guidance, and when the case should be handed to a human.
Even the human brain is not one lump doing every job. It has separate systems for movement, emotion, planning, and many other functions. Good AI systems are starting to look similar: several parts, each built for a different job, working together toward one outcome.
Before choosing AI for a task, ask what kind of answer the task really needs. If the answer is words, use a language model. If the answer is a choice, a score, or a fixed policy decision, a different tool may be faster, cheaper, and easier to control.
At SCB 10X, we see these specialized AI tools rising quickly. They are getting smarter, cheaper to run, and easier to plug together. AI becomes an architecture of different systems working together, and the advantage goes to organizations that can design that architecture well.
Oravee Smithiphol is Tech Intelligence & Insights Manager at SCB 10X, where she tracks emerging AI developments across startups, enterprises, and financial services organizations worldwide.
