Zoho’s AI Blueprint: Balance Model Size for Maximum Affordability

Last week, Zoho extended a warm invitation to analysts for its Zoho Day 24 event, held in McAllen, Texas. The choice of McAllen as the venue was a deliberate one, emblematic of Zoho’s commitment to a strategy it has coined “Transnational Localism.” This approach involves expanding the company’s footprint across a diverse range of geographic locations, with a marked preference for smaller towns or cities that are grappling with brain drain due to a scarcity of promising, future-oriented employment opportunities.

McAllen represents one of Zoho’s latest expansions, and the city’s response, as articulated by the mayor and the director of economic development, was one of genuine enthusiasm and anticipation for Zoho’s integration into their community.

Amidst a series of engaging activities that celebrated local cuisine and talent, Zoho unveiled insights into its business philosophy and software solutions. One of our key takeaways from the event was a deeper appreciation for Zoho’s foundational principles: simplicity and affordability. These principles underscore Zoho’s dedication to providing accessible, user-friendly technology solutions that address the complex needs of modern businesses at an affordable price.

Simplicity and Affordability: Navigating the Selection of Optimal Language Models

Zoho’s philosophy towards technology utilization, particularly in harnessing transformer-based language models across its suite of applications, is guided by a strategic focus on simplicity and affordability.

The team at Zoho is acutely aware of the investment required—not only the initial development costs of language models but also the ongoing operational expenses associated with runtime inference. This awareness is crucial for ensuring that Zoho’s applications remain both powerful and accessible to users as they execute various functions.

In conversations with Ramprakash Ramamoorthy, Director of AI Research at Zoho Corp, and through demonstrations at the analyst event, a clear strategy emerged: prioritize the “smallest best model” approach. This strategy led to the identification of four distinct categories of AI models, each suited to different tasks and requirements:

  • Narrow Models: Specialized in executing singular, well-defined tasks with precision. Examples include grammar checking, event prediction, and dataset categorization.
  • Small Language Models: Positioned within the 3 – 7 billion parameter range, these transformer-based models are compact enough to be operated on CPUs. Their size facilitates easier fine-tuning compared to larger counterparts. Use cases encompass translation, advanced noise cancellation, and transcript generation.
  • Medium Language Models: With 20 – 50 billion parameters, these models offer the ability to be fine-tuned and display more “emergent” capabilities than their smaller counterparts. Applications include direct inquiries about document contents, marking documents as anomalous for compliance, and video transcript analysis.
  • Large Language Models: These behemoths exceed 50 billion parameters, necessitating GPUs for inference and presenting challenges in fine-tuning. Despite higher costs, their “emergent” behaviors unlock advanced capabilities. Their functionality ranges from generating content to identifying and suggesting corrections for anomalous legal documents.

This tiered approach enables Zoho to judiciously apply the most fitting language model for a given task, balancing performance with cost-effectiveness to uphold its commitment to simplicity and affordability.

Combining Language Models for Affordable Business Optimization and Automation

Zoho recognizes the transformative potential of language models, aiming to select the most efficient and cost-effective option for each specific task. The company embraces the notion that “AI models will become commoditized,” with the real value for customers stemming from the strategic integration of models to deliver powerful, yet economical solutions. This approach not only maximizes business value but also ensures that cutting-edge solutions remain accessible to businesses of all sizes.

A compelling illustration of this strategy is the integration of an OCR (optical character recognition) model within the Zoho Expenses app. This model adeptly scans images of receipts—say, from a dining experience—and automatically populates the expense form with precise details like the restaurant’s name and the total charge, thereby bypassing the need for laborious manual data entry. Here, a narrow model excels in identifying the restaurant name, while a small language model proves more adept in certain scenarios, such as discerning the correct meal cost in instances where a receipt reflects a significant cashback amount. In such cases, the narrow model might mistakenly inflate the meal’s cost by including the cashback figure, whereas the small model accurately calculates the genuine expense associated with the meal for entry into the form.

Another example showcases the synergy of different language models within a document workflow process, outlined as follows:

  • Initial Step: Upon sending a document for signatures, a narrow model is employed for phishing detection, ensuring the document’s security.
  • Translation: A small model then takes over, translating the document to the required language, facilitating global business operations.
  • Summarization: Subsequently, a medium model provides a concise summary of the document’s content, enhancing efficiency in document review.
  • Anomaly Detection: The same medium model also undertakes anomaly detection, scrutinizing the document for any discrepancies or irregularities that may affect compliance or accuracy.
  • Enhanced Review: Lastly, a large model is utilized to generate suggestions based on a comprehensive review of the document, offering insights that might have been overlooked.

Through these examples, Zoho demonstrates its adeptness at leveraging a mix of language models to streamline and enhance business processes, thereby delivering substantial value while maintaining affordability. This strategic model combination not only augments operational efficiency but also supports Zoho’s commitment to innovation and accessibility for a diverse clientele.


Categories: Conversational Intelligence, Intelligent Assistants, Articles