SquadStack.ai launches Maya for sales calls, claiming an edge in internal tests
The company’s evaluation measures replies within sales conversations. Its reported lead over Google and OpenAI models has not been independently validated.
Loading page…
The company’s evaluation measures replies within sales conversations. Its reported lead over Google and OpenAI models has not been independently validated.
Listen to this story
SquadStack.ai launched Maya on Nov. 16, 2026, positioning it as a sales-call model tuned for Indian conversations rather than a general assistant; the company says it handles more than five million conversations daily. Maya scored 96.7% on a company-built next-response benchmark, ahead of several named models, but another model judged the answers and the result was not independently validated or evidence of higher sales conversion. India hosting is intended to support regulated customers, while plans to incorporate vocal cues and past-call history remain in development.
Maya’s post-training draws on more than 400 million real sales interactions collected over six years and labeled by outcomes such as sales, callbacks, and drop-offs.
The benchmark tested 9,962 conversational moments from 471 calls; Claude Haiku 5.5 judged 11 models across seven criteria.
SquadStack reports 96% context retention and 95% appropriate responses, versus 90% and 87% for the closest rival; these are judged behaviors, not sales outcomes.
Indian sales conversations are the target of SquadStack.ai’s new AI model, Maya, which the company says already handles more than five million conversations daily. Launched on November 16, 2026, it arrives with a claimed lead over Google and OpenAI models on an internal sales benchmark—not an independently validated finding.
Maya builds on an open-weight foundation model, rather than starting from scratch. SquadStack.ai says it further trained that base on more than 400 million real sales interactions gathered over six years. Each interaction was tagged with its outcome: a sale, a callback or a customer dropping off.
That additional training, known as post-training, is intended to give Maya a feel for Indian sales conversations. The company says its design accounts for informal answers, hesitation and interruptions. Its reported daily conversation volume spans Hinglish and Indian languages, placing the model’s focus on those exchanges rather than general-purpose assistance.
Co-founder and chief technology officer Vikas Gulati framed the engineering problem around simultaneous calls and audible delays. He said the company wanted sales behavior embedded in the model itself, rather than supplied through a long prompt each time it runs.
A live sales call is the hardest place to run an LLM. Thousands of calls can start in the same second, and the customer hears every slow turn. So we built Maya for this one job, with the sales behaviour in its weights rather than in a long prompt
Vikas Gulati, co-founder and CTO of SquadStack.ai
SquadStack.ai designed its Telesales LLM Benchmark around 9,962 conversational moments drawn from 471 real sales calls. Eleven models were asked to generate the next response. The assignment therefore tests a particular decision within a conversation: what the AI should say next, given the exchange up to that point.
The company used Anthropic’s Claude Haiku 5.5 to judge the answers across seven criteria. Those included keeping the conversation directed toward its objective and using tools correctly. Although another company’s model supplied the judgments, this remained SquadStack.ai’s own evaluation, not an independent assessment of Maya’s performance.
On that test, SquadStack.ai reported a 96.7% score for Maya. It put Google’s Gemma 4 31B at 92.5%, DeepSeek V4.1 Flash at 92.2% and OpenAI’s GPT-6 Luna at 89%. These are scores for the sales-response task the company constructed, rather than a general ranking of the models’ abilities.
The company also supplied narrower measures. It said Maya preserved conversational context on 96% of turns, versus 90% for the next-best model. It selected an appropriate response 95% of the time, compared with 87% for its closest rival. Those figures describe judged conversational behavior, not completed-sale rates.
SquadStack.ai says Maya is built and hosted in India, keeping customer conversations within the country. The company presents that setup as a way to help banks, lenders and insurers address requirements under India’s Digital Personal Data Protection Act. That is a claimed compliance benefit, not a guarantee of compliance.
Maya is the second model from the company’s in-house development operation, following its Arth V2 speech-recognition model. The next work on Maya is aimed at drawing more information from live calls, beyond the words exchanged in the immediate conversation.
SquadStack.ai is developing systems to feed insights from those calls into improvements. It plans to incorporate pauses, changes in tone and a customer’s earlier interactions. Those inputs remain development plans, separate from the capabilities and benchmark results claimed at launch.
Loading discussion...
Join the conversation
Explain where specialization would help—or get in the way.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.