Is the future domain-specific AI?
General-purpose models carry costs in accuracy, spend and compliance that enterprises cannot absorb. The case for language models built for one industry rather than all of them.
AI is finding many use cases in everyday operations like HR, customer management, knowledge management, research, sales & marketing and many other functions. There is now acceptance that using AI has big advantages especially in efficiency and speed. However for many applications, these AI tools are expensive and have limited accuracy. This is because most businesses are using general-purpose AI models, which often miss the mark for the following reasons:
- High opex — cost of training, retraining, and running a generic model is steep
- Accuracy — retrofitting domain logic into a generic model does not deliver the desired result
- Data security — concerns around data loom large, especially in finance, healthcare, and government
- Limited precision — limited reasoning and logic with degraded accuracy is disastrous for mission-critical tasks
- Compliance — inability to adhere to guidelines, lack of industry-specific laws and regulations, especially for evolving regulations
Large language models (LLMs) have billions (sometimes trillions) of parameters, trained on generic internet data. They require massive infrastructure, consume enormous amounts of energy, and still produce limited results. Hallucinations, irrelevant responses, and sky-high compute costs are common. A simple query can trigger all nodes of the model which is like firing a missile at a paper target. It burns tokens, energy, and time, yet doesn't guarantee accuracy.
Can the much-hyped generic AIs truly address the real challenges faced by industry? We believe they offer only limited value and fall far short of the magic many expect. Even OpenAI's CEO has noted, at an MIT event, that the era of giant AI models is over.
The era of giant AI models is over.
What a domain-specific model does differently
Domain-specific LLMs can address many of the above issues. They are purpose-built models for industry sectors, horizontal functions, or geographies, trained from the ground up on a particular subject. These don't just “speak the language” — they understand the logic, constraints, jargon and context of a specific subject. They are leaner, require less compute, and can be deployed with lower latency.
Unlike generic AI, they understand domain logic, workflows, and compliance norms, making them ideal for sectors like finance, healthcare, insurance, legal, and government. They're also region-aware, easily adapted to local languages and cultural nuances, and can be fine-tuned using RAG for faster deployment and continuous learning. As proprietary data accumulates, the model improves, reducing dependence on external sources.
With higher precision, better security and a higher ROI, domain-specific LLMs are expected to be the next leap in enterprise AI. Just as the 1990s saw an explosion of specific web-based services built on the backbone of the internet, domain-specific AI is poised to power the next wave of enterprise applications.
Where they already exist
There are areas where domain-specific LLMs are in the making or in use, although many aren't accessible to individuals. A few examples:
- Medical — Med-PaLM is trained on a vast dataset of medical literature. It can answer medical questions and provide insights into diseases, treatments and procedures, assist clinicians by analysing patient data for potential diagnoses or treatment plans, and analyse medical images. It is currently at over 86% accuracy.
- Finance — Bloomberg GPT can run a wide range of tasks within the financial industry. With 50 billion parameters and trained on over 700 billion tokens, the model achieves state-of-the-art results on various financial tasks.
- Legal — models such as Legal-BERT are trained on large data sets of legal documents, including legislation, court cases and contracts, enabling them to capture the unique characteristics of legal language.
And there are many areas where domain-specific LLMs are being invested in and will be in great demand:
- Customer support — enhanced customer understanding, automated but personalised interactions, multilingual support
- HR — resume and candidate screening, interview scheduling, training and skill-gap assessment, detecting red flags
- Fraud detection — real-time fraud detection, phishing detection, deviations from typical behaviour
As the limitations of general-purpose AI are getting increasingly clearer, the shift toward domain-specific LLMs is becoming a necessity. At AGR, we are not just observing this shift but are enabling it. We are closely collaborating with a leading LLM provider, bringing in domain and functional expertise to help make the models more specialized and industry-relevant.
