By Devyani Gupta, Co-founder and CEO, Arrowhead AI
There is a tendency in the AI industry to measure progress by the size of a model, the number of parameters it has or how well it performs on a benchmark. But businesses do not run on benchmarks. They run on customer conversations, transactions, decisions and workflows that need to work reliably every day.
That is where the conversation around enterprise AI needs to change. Large language models (LLMs) have made AI far more accessible, but the next phase of adoption will depend on how well we adapt these capabilities to specific business problems.
Consider a customer speaking to a bank’s voice assistant. The customer says, “No, not that one.” The system needs to recognise the words, understand what they refer to and respond without an awkward delay. A model that can write an excellent report but cannot reliably interpret this exchange is of limited use in that moment.
The question is no longer just what AI can do. It is whether it can do a particular job well enough to be trusted with it.
The gap between AI adoption and business value
AI adoption is growing rapidly, but turning experimentation into dependable business outcomes remains difficult.
McKinsey’s State of AI in 2026 survey found that nearly nine in ten respondents reported regular AI use in at least one business function, and 44% said AI was scaling across their enterprise. Yet only 37% said AI had contributed to their organisation’s earnings, a figure essentially unchanged from the year before. The gap suggests that access to AI is not the same as turning it into business value.
Part of the challenge is that enterprises have different requirements. A legal team reviewing documents, a bank processing customer requests and a retailer forecasting demand are not solving the same problem. They need different data, different measures of accuracy and different safeguards.
A general-purpose model provides a useful starting point. But expecting one model to handle every task equally well can add unnecessary cost and complexity.
Why specialised models matter
Purpose-built models are designed or adapted for defined tasks. These might include recognising speech, identifying customer intent, extracting information from documents or classifying requests.
The shift is not simply about making models smaller. It is about matching the technology to the work.
Stanford University’s AI Index Report 2026 highlights how quickly smaller AI models are closing the performance gap with larger systems. For instance, Ai2’s Olmo 3.1 Think 32B, with approximately 32 billion parameters, is significantly smaller than xAI’s Grok 4, which has been estimated to have around 3 trillion parameters. Yet, the report indicates that smaller models can deliver comparable results on several benchmarks, highlighting the importance of data quality, training methods and model design alongside size.
Costs have moved even faster. Research published by Epoch AI in September 2026 found that the cost of reaching a given level of AI performance has fallen by about 47% every quarter over the past three years, roughly a 13-fold drop each year. These are benchmark-based comparisons, not guarantees of equivalent performance across business applications. Nevertheless, they show why enterprises have more options than simply deploying the largest available model.
For businesses handling millions of interactions, the economics matter. So do response times, reliability and the cost of correcting errors.
Voice AI makes the case particularly clear
In voice-based applications, the challenge begins before an AI system generates an answer. It must first recognise what a person actually said.
Indian customer conversations make this particularly demanding. People switch between languages, use regional expressions and speak over background noise. They rarely talk in complete, carefully structured sentences.
A short response can also carry considerable weight. In banking or insurance, a misunderstood confirmation or rejection could affect what happens next in a conversation. Speed matters too: on a live call, a pause of even a second or two can feel like the line has dropped, and smaller models typically respond faster.
We see this as a systems problem rather than a question of model size alone. Speech recognition, language understanding and response generation must work together. Each stage needs to be evaluated against the conditions in which it will actually operate. That is why we built our own speech recognition, text-to-speech and small language models, designed for the code-mixed way Indian customers actually speak.
A capable language model can recover from many speech recognition errors. A misspelt name or a slightly garbled word usually leaves enough context to work out what the customer meant. But some errors are harder to recover from. If “don’t” is heard as “do”, or ₹5,000 as ₹50,000, the transcript still reads as perfectly plausible, yet the intended meaning changes completely. A smaller model does not automatically perform better, nor is a larger model necessarily more accurate. The answer depends on how well the model handles the language, accents, background noise and context of real-world conversations.
Nor does a smaller model automatically perform better. The answer depends on the task, the quality of the data and how the complete system is designed and tested.
The future is a combination of models
The industry does not need to choose between large and specialised models. It needs to be used intelligently.
A specialised model can handle a routine, well-defined task, while a larger model can be brought in when a request requires more complex reasoning. Human intervention remains important for sensitive decisions and situations where the system is uncertain.
This approach also demands discipline. The National Institute of Standards and Technology’s AI Risk Management Framework emphasises reliability, accountability, transparency and ongoing risk management. Closer to home, the Reserve Bank of India’s FREE-AI framework for the financial sector, released in 2025, sets out similar principles, including accountability, explainability and resilience.
Enterprises must evaluate systems in real operating conditions, monitor their performance and establish clear ways to handle failures.
Ultimately, the best enterprise AI system will not necessarily be the one with the largest model behind it. It will be the one that understands the task, performs consistently, fits the economics of the business and knows when it needs human support.
That is where the next meaningful gains in enterprise AI are likely to come from: not making every model capable of doing everything, but building systems that do the right things well.