From data to decisions: How Piramal Finance built an AI-native knowledge engine

As organisations turn artificial intelligence (AI)-native, knowledge work will shift from discrete specialised skills to real-time, embedded intelligence. In future, employees will collaborate with AI agents that provide contextually relevant information, externalise tacit judgement, and ensure that humans remain “in-the-loop.”

In an interaction with ET.com, Debashish Nagendra Barai, EVP of Technology at Piramal Finance, a listed non-banking finance company (NBFC) with assets under management (AUM) of over Rs 1 lakh crore, throws light on the efforts towards the same. Edited excerpts:

In this age of the AI-native organisation, how is Piramal Finance redefining the organisation’s knowledge, and what capabilities are you focusing on to capture this knowledge?

As a fairly young NBFC, data use is one of Piramal Finance’s key differentiators. As we mature into an AI-native organisation, our main focus would be to formalise and distribute the knowledge embedded across the organisation. Essentially, from being data-driven to knowledge-driven.

Already operating from a cloud-native, digital-first architecture, the amount of data being collected across the ecosystem is huge. We are taking it to the next level by adding context.

Earlier, in a traditional data-driven set-up, we would provide different customer inputs to sales representatives before they interacted with customers. However, the customer’s latent needs often emerge during the interaction itself. That behavioural context is now being fed back into the system as a feedback loop, helping us institutionalise knowledge rather than just data.

We are not trying to replace humans. Rather, we are making every function AI-native —customer-facing teams and internal operations alike. Reports and dashboards that were consumed intermittently are now delivered in real time. Recommendation engines for underwriting, risk, and profiling also operate continuously.

In short, data is now a built-in, operational layer of the business, not a separate, post-hoc analytical capability.

Which kind of knowledge inside Piramal Finance is the hardest to convert into data or AI models?
The most difficult challenge for any organisation is digitising the knowledge that resides within individuals. For example, a relationship manager understands borrower intent beyond submitted documents. A credit officer interprets local market conditions before making lending decisions. Collection teams understand regional factors before deciding on collection strategies. These are highly human-driven judgement calls.

The challenge is figuring out how to convert those judgements into code. As our AI and data platforms evolve, and as contextual intelligence improves, we believe we will reach a much better state over time.

So how has adding context to data impacted the business?
Traditionally, our business focused on home loans, real estate loans, and similar products, which are slower-moving. But today, we are pivoting towards products like personal loans, where disbursement can happen within two to three hours, depending on the scenario.

What’s important is that this shift did not require us to rebuild our technology infrastructure or create entirely new systems. The building blocks were already in place. So, while sales acceleration is one outcome, there is also faster time-to-market for new businesses.

For example, today we are launching products like gold loans. If you look at the overall product launch cycle, you will see how flexible the environment has become in enabling us to quickly launch new businesses and serve customers directly. The ROI (return on investment) from Snowflake’s AI Data Cloud has been significant across operational efficiency, business agility, compliance, and innovation.

Around 40% of customer service interactions are now happening completely in self-service mode. The average call handling time for the remaining interactions has been reduced to two-three minutes. The time from loan sanction to disbursement has been shortened to just a few hours across different product portfolios. Even the time from logging to sanction approval has significantly reduced.

So, the combination of data, AI, and our technology stack has improved every stage of the journey—customer experience, sales, operations, and marketing.

With agentic AI embedded into underwriting and operations, how do you balance model-driven insights with human intelligence, especially when there’s conflict between the two? What governance or escalation mechanisms do you use?

We focus on three areas to manage this balance.

Explainability: We require clear reasons for any recommendation an AI agent makes, so business users can assess it. Sometimes the model spots factors users may miss; at other times, humans see nuances that the model doesn’t.

Escalation governance: We have a clearly defined escalation and governance framework set up, both in terms of function and impact level, to settle any dispute between model recommendation and human judgement.

Feedback loops: All human overrides and their precursor signals are logged, and our system is able to learn from human judgements to make future model predictions.

Now that enterprise AI assistants are starting to communicate knowledge in natural language, how do you look at the evolution of skills and team structures across products, sales, and risk?
Today, teams need a combination of subject matter expertise, analytical thinking, and the capability of interacting with AI.

We have introduced several initiatives to support this transformation. For example, we created platforms like ‘build your own report’, ‘build your own tech’, and ‘chat with data.’

These allow teams to build their own lightweight applications and automate repetitive tasks without waiting for the technology team.

Today, even operations teams are developing their own applications independently. Alongside that, we conduct organisation-wide AI training programs every quarter to improve AI literacy. We also analyse how employees interact with these systems, so that we can continuously refine adoption and usability.

What will be lenders’ greatest advantage in 3-5 years? Is it proprietary data, model IP, agent orchestration, or talent?
I think it will be a blend of everything. Competitive advantage will be generated through the intersection of data, model IP, orchestration, domain expertise, organisational capacity, and learning in a loop with human judgement.

From a Piramal Finance perspective, our strategic focus areas over the next few years include:

• Building a more mature unified enterprise knowledge and data fabric
• Expanding AI-assisted decision systems, increasing AI-supported decision-making from around 40% today to nearly 80%
• Enabling true real-time operational intelligence with minimal latency
• Strengthening responsible AI governance frameworks
• Building scalable human-plus-AI interfaces

These are the key areas we believe will define the next phase of transformation.

How do you see the future of knowledge work evolving over the next three to five years?
Businesses and employees will increasingly need to work alongside AI systems. Today, prompt engineering itself has become an important skill. Over time, many of these capabilities will become embedded into systems, but employees will still need to understand how to collaborate effectively with AI.

Thanks to the data layer we have already built, teams don’t have to worry separately about accessing data or insights because those are seamlessly integrated into business workflows.

For example, earlier teams depended on data teams to create graphs and reports for presentations. Today, using tools like Cortex AI and Snowflake CoCo, employees can generate visualisations and insights directly through prompts. This fundamentally changes how business functions interact with technology.