In a significant revelation for the corporate world's rush to adopt artificial intelligence, a top executive from global IT giant Cognizant has sounded a cautionary note. Babak Hodjat, the Chief Technology Officer for AI at Nasdaq-listed Cognizant, has pointed out a fundamental and unresolved challenge: the inability of today's advanced AI systems to be fully trusted for complex, large-scale reasoning in critical business operations.
The Core Challenge: AI's "Catastrophic Breakdowns" in Reasoning
Speaking at the company's AI Lab in Bengaluru, Hodjat addressed the enterprise race to embed Large Language Models (LLMs) deeper into workflows. While acknowledging the power of current LLMs, he stressed that the industry is "far from having a single panacea" to determine if an AI's output can be trusted. This trust gap becomes especially pronounced as companies move towards deploying autonomous, multi-agent AI systems.
Hodjat explained that despite their sophistication, LLMs tend to break down when pushed into longer or more intricate chains of logical reasoning. Research indicates that even the most advanced models suffer what he termed as "catastrophic breakdowns" when required to execute extended sequences of reasoning steps. This flaw isn't just theoretical; it poses tangible risks for real-world business applications.
Real-World Risks for Telecom, Finance, and Supply Chains
To illustrate the practical implications of this limitation, Hodjat cited the classic Tower of Hanoi puzzle. This logically straightforward task sees LLMs beginning to make errors after just a few hundred reasoning steps. When this limitation is extrapolated to enterprise environments, the stakes are immensely higher.
This reasoning fragility introduces significant risk for sectors deploying AI across intricate, multi-step workflows. In domains like telecom network management, global supply chain logistics, or complex financial systems, decisions are rarely isolated. They often compound over thousands or even millions of sequential steps. An error in early reasoning can cascade, leading to flawed outcomes with potentially severe operational or financial consequences.
Cognizant's Solution: Embedding Human Oversight
In response to this critical challenge, Cognizant is not waiting for a perfect AI model. Instead, the company is proactively designing safeguards into its AI systems. The cornerstone of their approach is the integration of multiple human-in-the-loop mechanisms to catch and correct AI reasoning failures before they cause harm.
One practical method involves setting up triggers for human intervention. When the AI system's own confidence score in its reasoning or output drops below a predefined threshold, the process is automatically flagged for human review and oversight. This creates a vital safety net, ensuring that human expertise remains integral to high-stakes decision-making processes, even in highly automated environments.
Hodjat's insights underscore a pivotal moment in enterprise AI adoption. The message is clear: while the technology's capabilities are expanding rapidly, blind reliance on AI for complex reasoning is premature and risky. For Indian and global businesses investing heavily in AI transformation, the path forward requires a balanced, hybrid approach—leveraging AI's power while firmly anchoring it with human judgment and robust evaluation frameworks to ensure reliability and trust.
