SIAgentNews.com
← All news

SquadStack.ai launches Maya to lead sales-specific AI conversations

The new model claims to outperform major tech giants in handling the nuances of real-time sales calls, particularly in the Indian market.

· First reported by The Economic Times

Risograph poster: picture for the story 'SquadStack.ai launches Maya to lead sales-specific AI conversations'
AI-made risograph poster for this story. No real people or products are shown.
40 crorereal sales interactions used for training

The short version

  • Maya scored 96.7% on a custom benchmark, beating models from OpenAI, Google, DeepSeek, Alibaba, and NVIDIA.
  • The model was trained on six years of data from over 40 crore real sales interactions.
  • It is specifically designed to handle Hinglish and various Indian languages with nuances like hesitation.
  • The system is hosted in India to comply with local data protection laws like the Digital Personal Data Protection Act.
  • Maya maintains conversational context on 96% of turns, outperforming its closest rival's 90%.

Why it matters

For businesses, this represents a shift toward 'vertical' AI—models built for one specific job rather than general tasks. It also highlights how localized data and hosting can help companies meet strict regional privacy regulations.

A specialized sales model

Indian AI startup SquadStack.ai recently introduced Maya, which the company describes as the world's first Sales Large Language Model (LLM). Unlike general-purpose AI, Maya is built specifically to manage the complexities of live sales conversations. The company claims it outperforms models from major tech giants like OpenAI, Google, DeepSeek, Alibaba, and NVIDIA in a specialized benchmark.

To test these capabilities, SquadStack.ai developed a benchmark using 9,962 conversational moments from 471 real sales calls. They assessed eleven different models on their ability to generate the next logical response in a conversation. Maya achieved a score of 96.7% on this in-house test, notably ahead of Google's Gemma 4 31B and OpenAI's GPT-6 Luna.

The company used Anthropic's Claude Haiku 5.5 as a neutral judge to evaluate the responses. The judging process looked at seven specific parameters, including the model's ability to use tools correctly and maintain the primary objective of the conversation. Maya maintained conversational context on 96% of turns, while the next-best model reached only 90%.

Training on real-world data

The performance of Maya comes from a massive dataset of real-world interactions. SquadStack.ai post-trained the model from an open-weight foundation model using more than 40 crore real sales interactions. This data spans six years of running sales calls, with each interaction tagged by its outcome, such as a sale, a callback, or a drop-off.

By using this specific data, the company aimed to bake sales behavior directly into the model's weights. Vikas Gulati, co-founder and CTO of SquadStack.ai, explained that they built Maya for this specific job because live calls are difficult for LLMs. He noted that thousands of calls can start simultaneously and customers can hear every slow turn of the AI.

This approach allows the model to handle the messy reality of human speech. Maya is designed to understand informal responses, customer hesitation, and frequent interruptions. It currently manages more than 50 lakh conversations every day across various Indian languages and Hinglish, which is a mix of Hindi and English.

Localized hosting and privacy

A significant part of Maya's value proposition is its geographical footprint. The model is built and hosted entirely within India, which ensures that all customer conversations remain within the country's borders. This infrastructure choice is specifically intended to help organizations meet strict regulatory standards.

For industries like banking, lending, and insurance, this localization is a key factor for compliance. It helps these companies address data protection requirements under India's Digital Personal Data Protection Act. By keeping data local, these firms can potentially use AI for sales without violating privacy mandates.

The model is the second project from the company's in-house model factory. It follows their previous speech recognition model, known as Arth V2. By focusing on a domestic infrastructure, SquadStack.ai is positioning itself as a primary partner for Indian enterprises looking to integrate AI into their customer-facing operations.

Future technical iterations

SquadStack.ai is not stopping at the current capabilities of Maya. The company is already developing new systems to improve the model using insights gathered from live calls. These future updates aim to make the AI even more responsive to the subtle cues of human emotion and behavior.

The company plans to incorporate more sophisticated cues into the model's logic. This includes identifying pauses in a conversation, detecting changes in a customer's tone, and referencing a customer's previous interactions. These features are intended to make the AI feel less like a script and more like a natural conversation.

By refining these inputs, the company hopes to increase the accuracy of the model's responses. The goal is to ensure that the AI can navigate the high-pressure environment of a sales call where every second counts. As the model evolves, it will continue to be tested against the benchmark to ensure it maintains its lead over general models.

What to watch

  • How Maya's performance compares to newer updates from OpenAI and Google.
  • Whether more Indian startups will follow the 'vertical AI' path for specific regional languages.

Try it

Look up India's Digital Personal Data Protection Act to see how local data hosting requirements affect AI adoption in the finance sector.

Sources and links

First reported by The Economic Times. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

More AI news