Bring Back Purpose-Built Machine Learning, Says Zoho’s AI Research Lead

A new disease of abundance is spreading across the AI industry, encapsulated by a term called tokenmaxing.

As companies rush to embed AI into products and workflows, concerns are emerging about runaway token consumption and rising inference costs. Recent reports have highlighted that firms such as Uber are questioning whether increased AI use is translating into better outcomes.

Companies, including Salesforce, have also reportedly tightened controls around AI spending as enterprises grapple with the cost of large-scale model usage.

As concerns around AI costs grow, enterprises are being forced to rethink how they deploy AI models, and whether this situation could even offer a glimmer of hope for humans, who may prove more cost-effective than burning tokens.

In an interaction with AIM, Ramprakash Ramamoorthy, Director of AI Research at Zoho, says enterprises need to move away from the assumption that every problem requires a large language model.

Instead, he argued for a return to the discipline that defined earlier machine learning systems, where models were built for specific tasks and deployed with clear boundaries.

AI Everywhere Does Not Mean AI for Everything

“You don’t always need to use an LLM to find an anomaly or to forecast your outcomes,” says Ramamoorthy.

Ramamoorthy explained that enterprises should rely on purpose-built machine learning systems wherever possible and reserve LLMs for tasks that genuinely require language understanding.

“Machine learning models were purpose-built. I will build a machine learning model for anomaly in finance, then a different model for anomaly in my operational data and so on,” he says. “I think we need to bring that back.”

For instance, he says Zoho’s AI stack combines more than 80 machine learning algorithms that work alongside LLMs rather than replacing them.

He also argued that judgement will become an increasingly important skill for engineers, particularly when deciding which use cases should be routed through which models.

“For something like an email summary, you wouldn’t need a Claude scale or a ChatGPT scale model,” Ramamoorthy says. “At Zoho, we are a strong advocate of right-sizing your LLMs.” Apparently, this was one of the reasons Zoho launched Zia LLM, a small, in-house 8-billion-parameter model for just the specific tasks on proprietary data. 

Ramamoorthy also says that restricting the scope of a model can not only reduce costs and token consumption, but also minimise hallucinations and improve domain-specific accuracy through targeted fine-tuning.

When it comes to embedding language models inside customer-facing products, Ramamoorthy advocates a more pragmatic approach. He says that in enterprise software, everything is often very predictable and repetitive. 

Ramamoorthy pushes back against enterprises that label their products as ‘AI-enabled’ by just bundling in LLMs. 

Another aspect of the tokenmaxing debate is whether junior engineers will continue to have a place in organisations if entry-level software engineering tasks can be performed more cost-effectively by humans rather than being routed to LLMs.

However, Ramamoorthy suggested that it is a short-sighted way of looking at talent. He argued that when junior engineers spend years within an organisation, they accumulate company-specific knowledge and domain expertise that compounds over time and becomes difficult to replace.

“The most powerful executives are those who have spent a lot of time at a company and who are not headhunted,” says Ramamoorthy.

Hiring in the Age of AI

Beyond the debate over AI costs, another question is emerging across the industry: should engineers now be expected to know how to combine AI tools, models, and workflows to solve problems efficiently and reduce token burn?

Ramamoorthy says the rise of AI-assisted development has changed what companies should look for when hiring engineers. Traditional coding tests focused on algorithmic exercises are becoming less relevant when developers increasingly write software alongside AI agents embedded within development environments.

“It doesn't make sense to whiteboard a bubble sort today,” he says.

He says familiarity with modern AI tools is becoming important, and candidates who have never worked with them could be at a disadvantage. 

At the same time, he argued that tool usage alone should not become the primary measure of technical ability.

Ramamoorthy also pointed to a growing category of professionals who specialise in combining multiple AI tools to complete tasks.

Zoho has created an Applied AI team focused on evaluating new AI products and identifying practical uses for them across internal operations. The team's role is to experiment widely, assess emerging tools and identify applications that can create value for the business.

According to Ramamoorthy, not every experiment needs to succeed if the overall process helps uncover useful outcomes. “Do ten experiments. I’m sure one experiment will be really productive for the company. It’s okay if you waste those nine experiment tokens,” he says, advocating for long-term gains over short-term cost savings. 

Despite the rapid pace of change, he says hiring decisions still come down to fundamentals. He seeks evidence of past accomplishments, adaptability to future tech, and independent problem-solving, not just technical skills.