AI on bad data is a faster way off a cliff, warns Elastic CEO Ash Kulkarni

Bad data and missing business context are the biggest obstacles to taking artificial intelligence (AI) from demos into everyday business use, according to Elastic chief executive officer (CEO) Ashutosh Kulkarni, who warned that automation can make companies act on wrong information faster.

“The hardest part of getting AI to be valuable for you is making sure that you get the data right,” Kulkarni told ET AI at ElasticON in Mumbai.

Elastic makes search and analytics software that helps businesses find information across their systems and provide data to AI applications.


Kulkarni said companies faced the same problem when they began digitising their operations more than three decades ago. They had plenty of data but often did not know which information to trust or how to use it. The risk is greater with AI because it can use that data to make decisions and take action.

“Taking actions and making decisions on bad data is basically going off a cliff faster than you ever could,” he said.

Why AI struggles beyond demos

Gartner estimates that unstructured information, including documents and multimedia files, makes up 70% to 90% of organisational data. Business context sits across documents, customer support tickets, system records and security alerts, with frequent updates and different rules governing who can access it.

“Unstructured data governance transforms messy, hard-to-use information into a reliable, accessible resource,” according to Mark Beyer, analyst and research VP at Gartner.

Kulkarni said AI tools can perform well in demos but struggle in everyday business use because they often don't have access to the right information.

He gave the example of an e-commerce company building a customer support agent. A human handling a complaint about a faulty product might investigate whether the company has recalled it. An AI agent needs access to the information that would help it reach the right conclusion, he said.

That requires more than a record of the customer’s latest complaint. The system needs information about inventory, customer service procedures, product issues and earlier support tickets, Kulkarni said. Companies must connect those sources and prepare the relevant context so the agent can use it quickly.

"Almost always, in nine out of ten cases, it's not the model choice or anything else. It's not the people building the system. It is almost always the data. If you don't figure out that data problem correctly, it's garbage in, garbage out," Kulkarni told ET AI.

Testing often exposes missing connections and gaps in the business knowledge available to an AI agent, he added.

The cost of getting AI wrong

Getting that context right also affects costs. Kulkarni said companies moved quickly from the initial enthusiasm for using AI wherever possible to asking whether the resulting bills made business sense.

Kulkarni gave the example of the earlier rush to adopt 'lift and shift' (also known as rehosting), which involved moving existing applications to the cloud. Companies expected savings but sometimes ended up paying more because they used flexible technology in an inflexible way, he said.

“It is the same with AI. It is a very powerful technology, but a fool with a tool is still a fool. If you use a very powerful technology incorrectly, you can still get horrible outcomes,” he said.

Companies now want to find the right information for AI, avoid making models repeat the same work, and keep a closer watch on usage and costs, Kulkarni said. They also want AI spending to improve service or reduce costs elsewhere.

India is a talent centre, not a low-cost centre

As Elastic expands its work with enterprises, Kulkarni said it is also growing its India team. The company has about 300 employees in India across engineering, support, finance, marketing and sales, up from around 50 a few years ago, he said.

“Even for engineering, we have always had the notion of a globally distributed model, in the sense that we don't have captive labs or captive locations where we go for low costs. We have never had that model,” Kulkarni said.

“When we hire in India, we don't view it as a low-cost centre. We view it as a talent centre. We pay people here world-class wages, but we are looking for the best people. It's not about numbers for us,” he said.

Elastic hires across Indian cities as a remote-first company, Kulkarni said. “India is definitely a centre that you should be hiring in,” he said, adding that he expected the team to continue growing.

OpenAI and the need for context

He stressed the need for that information when asked whether OpenAI’s latest announcements to host agents, protect data and make its models cheaper to use threatened Elastic’s role. Models still need the right information about a company to do useful work, Kulkarni said.

“Nothing changes,” he stressed. “The same problems that they have about getting to the right context still exist,” he added.

The search AI company expanded its collaboration with OpenAI in July to help companies build AI applications and agents using the latter's models with Elasticsearch, Elastic’s search technology.

Kulkarni said a model could answer questions about a film but could not determine how best to serve a customer without knowing their purchases, previous problems and preferences.

OpenAI will always be a very strong partner because they are bringing the intelligence. We will always be a strong partner to them because we are bringing the data and context. One without the other is useless," Kulkarni said.