Fresh Data, Clear Access Rules: The Foundation For Reliable AI Applications

AI applications need up-to-date information and clear rules on what they can access to work reliably at scale, according to executives speaking at the inaugural edition of Inc42’s ‘The CTO Summit 2026’ in Bengaluru.

The challenge is to make information spread across databases, data warehouses, and documents available to these applications while preserving its quality and controlling its use.

The panel, ‘Building The Data Foundation For AI At Scale’, featured PhonePe’s head of engineering, merchant payments, Kisalay Ranjan; OceanBase’s GM, India business, Bhanu Jamwal; and Fibe’s CTO Anil Sinha. Digio EIR Kshitij Shah moderated the discussion.

Ranjan said PhonePe tracks where data comes from and sets rules for its use at the level of individual data elements, across both transactional databases and data warehouses. Adding these controls later through individual applications becomes increasingly difficult as companies automate more tasks, he said.

“The power is first ensuring that whatever data we are consuming is of high quality,” Ranjan said.

Teams must also know where information is stored, whether it is sensitive, whether its use meets compliance requirements, and which applications can access it. PhonePe uses shared platforms to maintain a single authoritative source of sensitive customer data instead of duplicating it across systems, he added.

Jamwal said organisations’ data is spread across separate systems used to process transactions, run analytics, and support AI searches, alongside data lakes. Companies deploying AI agents need to connect these sources so their applications can retrieve current information quickly and accurately, he said.

“There would be a time when you need high accuracy for your agents,” Jamwal said. “You need one single solution which can give you data sitting at one point, but more importantly, give you that fresh data with low latency.”

The OceanBase executive advocated a unified data layer to make information across these systems available to agents and other AI applications with minimal delay.

At IPO-bound lending tech startup Fibe, the focus includes extracting useful information from documents that vary in format, such as records from hospitals, clinics, and educational institutions.

Sinha said Fibe uses generative AI to extract signals from these documents and feeds the signals into its existing credit model, rather than asking GenAI to make lending decisions. He said the model has expanded from 8,000 to 23,000 variables as the company incorporates more information.

Beyond data quality and access, companies also have to balance AI model performance against cost and response times.

Ranjan said smaller models running within PhonePe’s own technology environment can suit tasks that involve large volumes and require quick responses. More advanced models may be useful for tasks requiring stronger reasoning, with the choice depending on the use case, model availability, and cost.

The emphasis on underlying systems also featured in a separate session at the CTO Summit, where Razorpay’s SVP of engineering Prabhu Ram cautioned that AI cannot compensate for weak engineering foundations, including inadequate systems for monitoring performance and detecting problems.

The post Fresh Data, Clear Access Rules: The Foundation For Reliable AI Applications appeared first on Inc42 Media.