Arrowhead AI Wants to Turn Voice Agents Into the New BFSI Workforce

For decades, India’s financial services industry has relied on large human teams to handle customer calls, sales, collections, and support. Bengaluru-based startup Arrowhead AI believes that model is beginning to change—and it wants its voice agents to be at the centre of that transition.

The startup started deploying its inbound customer support voice AI deployment with digital lending platform Kissht earlier this year after raising $3 million in seed funding with participation from Kissht executives. Within two months of launch, Arrowhead’s AI voice agents began handling 40% of all inbound customer support calls received through Kissht’s Ring app. 

Across the selected support categories, around 88% of calls are being resolved without human intervention, while average handle time is approximately 30% lower than comparable calls handled by human specialists, the company claims.

The deployment is significant because Arrowhead and Kissht introduced the technology directly into inbound customer support, rather than beginning with outbound sales or engagement.

The AI currently handles loan application rejection queries and requests related to mandate or No Objection Certificate cancellations—areas that can involve frustrated customers, sensitive financial information, and regulatory concerns.

For Arrowhead, however, the Kissht deployment is only an early step towards a much larger proposition of AI agents that can eventually manage customer relationships across multiple channels and stages.

“We are already seeing entire teams of human agents and traditional call-centre operations being significantly scaled down because of AI,” Devyani Gupta, Co-founder and CEO of Arrowhead AI, tells AIM. “The longer-term vision is that an AI agent can potentially manage the entire customer relationship end-to-end.”

From Analysing Human Calls to Replacing Them

Arrowhead did not begin as a voice AI company. The startup initially built technology to analyse human-led sales conversations. One of its early large customers was upGrad, which used Arrowhead’s models to identify performance gaps and instances of potential misselling.

The pivot came when voice AI began becoming commercially viable.

“When voice AI started coming into the market, that's when we decided to pivot,” Gupta recalls. “We realised that the very activity we were analysing—human-led sales conversations—could eventually be replaced by voice AI, and we needed to keep up with where the technology was heading.”

Arrowhead has since onboarded more than 50 large companies.

AI Workforce

The Kissht deployment illustrates the economics Arrowhead is targeting. AI agents complete calls about 30% faster than human specialists handling comparable queries. That could translate into approximately 30% savings on fully automated calls, Gupta estimates.

But the more important number may be the 88% resolution rate.

A call is counted as resolved only when the AI answers the customer’s query end-to-end without requiring a transfer to a human agent. Arrowhead uses a separate AI funnel to evaluate whether interactions were actually resolved, while the remaining 12% are transferred to human specialists when intervention is required.

Kissht asserts that the efficiency gains have not come at the expense of customer experience. The agents are also integrated with Kissht’s dialler and CRM system, allowing customer information and application status to be retrieved before the conversation begins so they don’t have to repeat information. 

AI Knows When to Step Back

The high automation rate does not mean Arrowhead is giving the AI unrestricted control over customer conversations.

The initial deployment has been deliberately constrained to predefined categories. If a customer refers to the Reserve Bank of India during a conversation, for example, the call is immediately transferred to a human specialist rather than allowing the AI to continue responding.

The approach reflects a broader challenge for AI agents in BFSI. The system must know not only how to answer, but also when not to answer.

That becomes even more important as Arrowhead moves into financially consequential interactions.

Following the initial deployment, Arrowhead and Kissht are now developing an AI agent for EMI and repayment-related queries. This is a substantially more sensitive use case than answering a general service query. 

The new agent is therefore being developed with stricter controls. One of its core instructions is described as “money facts, absolute”, meaning the system must not independently calculate, estimate or round up financial information.

Why Arrowhead Built Its Own AI Stack

Unlike startups that primarily build application layers on top of frontier models, Arrowhead has developed its core voice AI stack in-house, including an orchestrator, speech-to-text, a small language model, and text-to-speech.

According to the company, its speech recognition system has achieved a 10% lower word error rate than leading models, particularly for regional languages.

Arrowhead has also developed a small language model specifically for BFSI use cases. The specialised SLM has helped Arrowhead cut end-to-end conversational latency to about half.

“When we started, latency was around one second. We built our own small language model and brought that down to around 500 milliseconds, which is among the lowest in the industry,” she says.

For voice AI, that delay matters. Long pauses can make an interaction feel artificial, create interruptions, and make customers less comfortable during sensitive conversations.

Arrowhead has similarly developed its own text-to-speech technology, trained on thousands of hours of voice data from voice artists, with an emphasis on making conversations more colloquial and appropriate for Indian customers.

India’s linguistic diversity is another area where Arrowhead is trying to differentiate itself.

The startup says its speech recognition models are trained on regional-language data covering colloquial expressions, accents and dialects across tier-1 and tier-2 cities.

Guardrails are Critical

As AI agents enter financial services, accuracy becomes critical. An incorrect response about a financial obligation, payment, or regulatory issue can have significantly larger consequences.

Arrowhead has therefore built multiple layers of protection into its systems.

“We have very strict guardrails in place to prevent the system from providing incorrect or inappropriate information. These safeguards are built into our SLM as well as the prompts provided to the AI agent,” Gupta explains.

But she acknowledges that probabilistic models inevitably create edge cases. “We have built our own infrastructure that can flag calls in real time where potentially incorrect information may have been provided,” she states.

The company has also developed what it calls a “circuit breaker”—a red-flag mechanism that alerts its team when it detects a problematic interaction.

This observability layer could become increasingly important as Arrowhead moves from contained customer-service tasks towards more complex financial workflows.

A Startup With Global Ambitions

Founded in 2022, Arrowhead currently has a team of more than 40 people across Bengaluru, Mumbai, Delhi, and Chennai.

Gupta comes with a background spanning BFSI, consulting, and digital transformation. She studied at the Wharton School of the University of Pennsylvania and worked at Boston Consulting Group across the US, Australia and Mumbai.\

Her co-founder and CTO, Vengadanathan Srinivasan, has more than 12 years of technology experience, including stints at Airbnb, Uber, Rippling, and AWS. He was among the first five employees of AWS’s Elasticsearch team and helped scale it to hundreds of people.

The company has raised approximately $3.2 million in its latest seed round and plans to raise a Series A in early 2027, according to Gupta.

While Arrowhead does not disclose revenue, Gupta says the business has doubled every month over the past three months.

The company operates in an increasingly competitive market that includes Sarvam AI, ElevenLabs, Greylabs, and Gnani.

The Hard Problem

As the number of conversations increases, Arrowhead faces a challenge that is less visible to customers but potentially more important to its ability to scale—observability.

“I think the biggest challenge as we scale is going to be scaling observability along with our volumes,” Gupta states.

At smaller volumes, human teams can manually review calls and identify failures. At scale, that approach becomes impossible. Arrowhead will need systems capable of automatically detecting model failures, service-provider problems, and unusual changes in behaviour across potentially millions of conversations.

The company’s ability to monitor those interactions may ultimately determine how far it can push its agents into regulated financial workflows.

“Our goal is to become the largest conversational AI agent company in the world,” Gupta states.