Businesses that deploy artificial intelligence (AI) agents to process payments, approve loans, and handle other financial transactions risk costly errors when the underlying systems run on outdated or fragmented data, according to a senior technology executive.

Predee Kajonpai, senior solution architect at database company TiDB, told attendees at the AI Vision Forum Manila on Wednesday that inaccurate or unsynchronized information can cause autonomous AI systems to repeat transactions or make incorrect financial decisions. The risks, he said, grow more serious as companies move beyond conversational AI tools toward autonomous systems capable of executing commercial transactions on their own — a category often described as agentic AI, in which software plans and carries out multi-step tasks with limited human oversight.

“If you ask a chatbot a question and it gives you a wrong answer, that’s fine. We just repeat it,” Kajonpai said.

“What we want is the AI agent to do business for you. What happens if your AI agent hallucinates or makes a mistake during a loan approval or payment? That’s what we don’t want to happen,” he added.

Data Infrastructure as a Bottleneck

One of the main obstacles to enterprise AI adoption, Kajonpai said, is continued reliance on data infrastructure that does not deliver updated information in real time. When information is stored across separate systems, it may not be synchronized at the moment AI agents execute transactions, increasing the likelihood of errors or duplicated operations.

He illustrated the danger with a hypothetical scenario in which an AI agent processes a customer refund but, after a system interruption, restarts on stale data without recognizing that the initial payment has already been completed.

“If a customer complains that they want a refund for services, the AI refunds them the first time, but if it restarts in the middle on stale data, it acts twice and refunds them again,” Kajonpai said.

He argued that businesses need to resolve these data management issues before deploying AI agents for financial transactions and other operations that require accurate, up-to-date information.

High Failure Rates for AI Initiatives

Kajonpai also cited estimates suggesting that as many as 95% of enterprise AI initiatives fail to reach full production, although the underlying research source was not identified during the presentation.

“Meaning that you put the money in 100%, but only 5% reach full production,” he said.

He additionally pointed to a forecast that 40% of AI projects could be canceled by 2027. Technology research firm Gartner previously projected that more than 40% of agentic AI projects would be canceled by the end of 2027 because of escalating costs, unclear business value, or inadequate risk controls.

According to Kajonpai, outdated data and fragmented information systems are among the technical challenges that could affect the deployment of autonomous AI applications. With Gartner's projection running through the end of 2027, the gap between AI pilots and production systems — and the pace of project cancellations — gives the industry a concrete checkpoint for how these technical and financial hurdles play out.

Unified Platforms Proposed

He advocated the use of unified data platforms capable of handling transactions and analytics in real time, reducing the need to copy information across separate systems. Such platforms, he said, could help organizations maintain consistent transaction records and reduce the risk of AI agents repeating completed operations.

“Having AI agents operate on a single platform removes duplicate actions, and we have proved it already,” Kajonpai said.

TiDB provides database technology designed to support transaction processing, real-time analytics, and AI-related applications. The argument effectively casts data architecture as a risk control in its own right, echoing the inadequate risk controls Gartner cited among the reasons agentic AI projects face cancellation.

— Juliana Chloe A. Gonzales

This story originally appeared on BusinessWorld.