Voice AI Sales for the Financial Services Industry

Below is my take on exploring possible avenues for GenAI voice assistants in the financial services industry (primarily for sales).

From my (little) experience in insurance and now in lending, one thing I have heard from veterans and come to see is that products which require relationship-building will be difficult to replace by voice bots (at least till the time that we have built agentic bots to such an extent that they can almost mimic a human).

I think the fundamental difference between a human agent selling and an AI selling is that AI runs on heuristics and rules. What makes us fundamentally human is that we are irrational. If we were rational, then every well-off Indian would have had adequate health insurance. However, most Indians are underinsured or uninsured. Given that premise, most companies and human insurance agents use selling by fear — wherein you induce fear to make the customer realise not through stats but via emotions as to why they need insurance and all that could go wrong if they did not take insurance. That remains the norm, not just in India but across the world.

In the context of lending — there are 2 scenarios. Low ticket-sized loans (think consumer durable loans) don’t require relationship building. These are one-off transactions. Share basic details about yourself (made easy thanks to India Stack) and you have the loan disbursed to your account. This needs to be repaid in 2–3 months so isn’t stressful for the customer.

However, the picture becomes complicated when it comes to high ticket-sized loans or SME loans because even today we don’t have completely automated underwriting for high ticket sizes. There is an element of human intervention left in terms of deciding the final eligibility. The human sales agent is the customer’s best bet at making a case with the underwriters to get the maximum loan amount. The fact that underwriting involves human intervention is also true for complex insurance products.

Hence digital penetration of health insurance and high ticket-sized loans is very low as compared to motor insurance and consumer durable loans. (the latter don’t need a high degree of human touch or human intervention to complete the process). Adoption of complete end-to-end digital processing of health insurance is yet to be cracked by any company.

For the above scenarios, current models might not be able to sell to the customer (or at least companies will not be able to see the value of the same at the moment) but can add value in terms of extracting nuance from call centre or agent transcripts.

Said that, we could define financial service products in terms of relationship building and non-relationship building or transactional products:

Relationship building products:

1. Wealth Management and Private Banking: — Personalized investment advice, Portfolio management, Tax optimization
2. Insurance (Complex Products): Life and health insurance with complex underwriting, marine insurance, customized insurance for SMEs
3. Retail Banking (Complex Products): mortgage lending, financial planning for retirement advice

Transactional products:

1. Insurance (Standard Products): Travel insurance, motor insurance, pet insurance, pocket insurance
2. Small ticket-sized personal loans and credit cards
3. Basic investment products (e.g., mutual funds)
4. Digital and Online Banking: Online account opening, robo advisory, customer service
5. Digital Payments: wire transfers, remittances, raising fraud issues
6. Value-added services: Personalised finance manager, Free credit health score report

Most transactional products are seeing a high growth in digital adoption. For the immediate future, I believe, for selling via voice — the best bet would be to focus on products which can be processed end-to-end digitally i.e. transactional products. Since voice offers an easier medium of filling forms and communicating (more so if the language is vernacular), it can open up the market to hitherto underserved or unserved customer segments.

By embedding voice in transactional products, we can ramp up learning (both in terms of understanding user behaviour and model improvements) via faster GTM to build better models for the long term, which will eventually allow us to use GenAI voice-bots for relationship-building products.

Thanks for reading all the way through. Would be happy to hear your thoughts.