AI on Bad Data Is a Faster Way Off a Cliff, Warn Tech CEOs at Industry Events

Data Quality Remains the Primary Barrier to AI Success

AI applications face a critical obstacle that has little to do with model selection or engineering talent. According to Elastic CEO Ashutosh Kulkarni, speaking at ElasticON in Mumbai, the foundation for reliable AI applications lies in data quality and proper governance. "Taking actions and making decisions on bad data is basically going off a cliff faster than you ever could," Kulkarni told ET AI1

1

. The risk intensifies because AI systems can use flawed information to make decisions and take action autonomously, amplifying errors at machine speed.

Kulkarni emphasized that in nine out of ten cases, AI failures trace back to data issues rather than model choice or implementation teams. "If you don't figure out that data problem correctly, it's garbage in, garbage out," he said1

1

. Gartner estimates that unstructured information, including documents and multimedia files, makes up 70% to 90% of organizational data, creating significant challenges for deploying AI effectively1

1

.

Why AI Systems Struggle Beyond Demonstrations

AI applications often perform impressively in controlled demos but fail in real-world business environments due to missing data context. Kulkarni illustrated this with an e-commerce customer support agent scenario. While a human agent investigating a faulty product complaint might check for product recalls, an AI agent needs structured access to inventory systems, customer service procedures, product issue databases, and historical support tickets to reach similar conclusions1

1

.

Business context sits scattered across documents, customer support tickets, system records, and security alerts, with frequent updates and varying access permissions. Companies must connect these sources and prepare relevant context so AI agents can retrieve information quickly. Testing typically exposes missing connections and gaps in the business knowledge available to AI systems1

1

.

Fresh Data and Clear Access Rules Define Operational Success

At Inc42's CTO Summit 2026 in Bengaluru, industry leaders reinforced that AI applications require up-to-date information and explicit rules governing data access to work reliably at scale. PhonePe's head of engineering for merchant payments, Kisalay Ranjan, explained that the company tracks data provenance and sets usage rules at the individual data element level across transactional databases and data warehouses. "The power is first ensuring that whatever data we are consuming is of high quality," Ranjan said2

2

.

PhonePe uses shared platforms to maintain a single authoritative source of sensitive customer data instead of duplicating it across systems. Teams must understand where information is stored, whether it contains sensitive elements, whether its use meets compliance requirements, and which AI applications can access it2

2

.

Unified Data Layers Enable Real-Time AI Performance

OceanBase GM for India business, Bhanu Jamwal, highlighted that organizational data spreads across separate systems for transaction processing, analytics, AI searches, and data lakes. Companies deploying AI agents must connect these sources so applications can retrieve current information quickly and accurately. "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," Jamwal said2

2

. A unified data layer makes information available to AI agents and other applications with minimal delay, addressing the accuracy requirements for autonomous agents2

2

.

Generative AI Expands Credit Models Through Document Processing

At IPO-bound lending tech startup Fibe, CTO Anil Sinha described using generative AI to extract signals from documents that vary widely in format, including records from hospitals, clinics, and educational institutions. Rather than asking GenAI to make lending decisions directly, Fibe feeds extracted signals into its existing credit model. This approach has expanded the model from 8,000 to 23,000 variables as the company incorporates more information sources2

2

.

Balancing AI Model Performance Against Cost and Speed

Beyond data governance, companies must balance AI model performance against cost and response times. Ranjan explained that smaller models running within PhonePe's own infrastructure suit high-volume tasks requiring quick responses. More advanced models may be appropriate for tasks requiring stronger reasoning capabilities, with choices depending on use case requirements, model availability, and cost considerations2

2

.

Kulkarni noted that companies have moved from initial AI adoption rush enthusiasm to questioning whether resulting expenses make business sense. He compared the current situation to the cloud migration wave, where lift-and-shift approaches sometimes increased costs because companies used flexible technology inflexibly. "It is the same with AI. It is a very powerful technology, but a fool with a tool is still a fool," he said1

1

. Organizations now focus on finding the right information for AI systems, avoiding redundant model work, and monitoring usage and costs more closely while ensuring AI spending improves service or reduces costs elsewhere1

1

.