At the inaugural CTO Summit 2026 in Bengaluru, senior engineers from PhonePe, OceanBase and fintech start-up Fibe debated the plumbing that keeps artificial intelligence reliable when it moves from prototype to production. Their message was clear: without up-to-date data and precise access policies, even the smartest model can produce misleading answers or leak confidential information.
Across three continents, a parallel experiment in the United Kingdom tested how digital inclusion can lower loneliness among disabled adults. The pilot combined one-to-one coaching, assistive-technology support and community-driven drop-ins to see whether better online confidence translates into real-world connection.
Fresh data rules: the backbone of scalable AI
PhonePe’s engineering lead, Kisalay Ranjan, explained that the company now tags every data element at ingestion, indicating source, sensitivity and compliance flags. By maintaining a single, authoritative repository for customer information, the firm avoids the chaos of duplicated records and retroactive permission checks.
OceanBase’s India GM, Bhanu Jamwal, stressed the need for a unified data layer that can serve AI agents with “fresh data at low latency.” He argued that scattered warehouses, transactional tables and data lakes create bottlenecks that undermine both accuracy and speed, especially when agents must query the latest transaction status in milliseconds.
Fibe’s CTO, Anil Sinha, illustrated a concrete use-case: generative AI scans heterogeneous medical and educational documents, pulls out credit-relevant signals, and feeds them into an existing loan-scoring engine. The model’s variable count grew from 8,000 to 23,000 as more document types were integrated, showing how clean, well-governed data can expand model richness without sacrificing stability.
Digital inclusion pilot: more than a device
The UK pilot, built on lived-experience research, treated digital inclusion as a bundle of access, confidence, skill and usability interventions. Participants received tailored help from trusted charities, ranging from basic device setup to safe-online-behaviour workshops. Researchers found that when people could use a platform that respected their assistive-technology needs, they were more likely to stay in touch with family, apply for benefits and join local activities.
Nevertheless, the study highlighted that online contact cannot fully replace face-to-face interaction. The most profound impact was observed in participants’ willingness to engage with services, not in a measurable drop in loneliness over the short pilot window. The authors recommend longer-term tracking to capture any downstream social benefits.
Responsible AI: standards, tools and business impact
Across both continents, the discussion converged on the rising relevance of responsible AI frameworks such as the OECD Principles (2024) and the NIST AI Risk Management Model. These guidelines stress fairness, transparency, accountability, privacy, security and human oversight as the pillars for trustworthy systems.
Bias detection tools, red-team exercises and ongoing model-monitoring were cited as practical ways to keep AI behavior in check. For high-risk domains—recruiting, lending, healthcare—the EU AI Act already mandates human-in-the-loop review, a requirement that aligns with the summit’s call for “smaller, in-house models for high-volume tasks and larger, supervised models for reasoning-heavy jobs.”
Finally, the panel warned that engineering fundamentals cannot be ignored. As Razorpay’s SVP of engineering noted, weak monitoring or logging infrastructure will cripple any responsible-AI effort, no matter how sophisticated the governance paperwork.
