Enterprise technology has long rested on a foundational assumption so obvious it has never needed stating. When a system receives identical inputs, it yields identical outputs. This is the principle of “determinism”. Banks reconcile millions of monetary movements and telecom operators bill subscribers accurately because the underlying software behaves in a perfectly predictable manner. Determinism is not a minor engineering attribute. It is the operational spine that gives enterprises the stability they depend upon. Indian IT services companies have built entire systems infrastructures for their clients based on this foundational technological contract. But now, large language models (LLMs) have begun to tear that contract up.
LLMs do not emerge from the deterministic lineage of classic enterprise systems. They emerge from probability theory, pattern recognition, and statistical learning. Ask the same model the same question twice and its responses may shift without hallucinating at all. This is not a bug. It is its nature. An LLM generates text one token at a time, where a token is a few letters and not always a full word. A tiny difference in the underlying probability distribution leads the system to choose a different token, which influences the next, and so on. Modern LLMs depend on massive floating-point computations executed in parallel across thousands of graphics processing unit (GPU) cores, each involving multiplication, addition, and rounding. The order in which these operations occur shifts subtly based on how hardware schedules tasks at that moment. In isolation each discrepancy may be microscopic. But cascaded across a response, they can produce something meaningfully different from what the same prompt produced an hour earlier. In an older computer system, the same result for the same instruction was guaranteed, even if given months apart. Achieving true determinism across heterogeneous hardware is unattainable today. No current software abstraction can harmonise the differences between GPU families, compilers, and memory architectures.
This would be a manageable technical problem if AI remained confined to consumer applications. Chatbots and creative assistants lose nothing when output varies. The problem is that AI is moving decisively into enterprise decision flows where it is no longer restricted to casual chatbot use but can have lasting impact on the enterprise. The fundamental agency cost of using an AI “agent” is not that the machine is unreliable but that it has no sense of scope. It solves the prompt given today while unwittingly establishing the conditions for a catastrophic failure months from now. When systems do break, diagnosis is slow and uncertain. Responsibility is diffused across undisciplined ideation, prompts, models, and partial fixes. No one holds a complete map.
The evidence is accumulating in places enterprises least expect. All Nippon Airways (ANA) briefly offered business class fares at a fraction of their actual price, and Cathay Pacific found itself selling flatbed seats from Hong Kong to New York for $675 — these were not straightforward software errors that a senior developer reviewing code would catch. They were outputs of systems behaving probabilistically in edge cases that deterministic logic would have blocked at the gate. Turkish Airlines is currently contesting similarly mispriced fares in 2026. The airlines will likely win in court, but the pattern is established. Pricing logic that was once the most deterministic of enterprise functions has acquired a faint and dangerous randomness.
The cybersecurity implications are grimmer still. When Scattered Spider breached Marks & Spencer last year through social engineering at the IT help desk, the attack itself was human. But the conditions that made M&S so exposed — sprawling digital infrastructure, interdependent systems, and third-party access points multiplied by years of digitisation — were the natural habitat of an organisation that had been building complexity faster than it could govern it. The eventual cost was £300 million in lost operating profit. The entry point was a phone call to a help desk. When the forensic plumbers arrived, they found not a single failure but a landscape of interdependencies that no one had fully mapped.
This is where the Indian IT services industry re-enters the story, though not in the role it would prefer. These firms are themselves accelerating the problem, deploying AI and vibe coding approaches at enterprise scale for their clients, generating volume while quietly accumulating the technical debt that will eventually surface as system failures. The fallout, when it arrives, will require something client firms have been moving away from: real people with deep diagnostic skills who can go in and discover why a pricing engine has begun treating business class as a loss leader, or why a compliance model has started producing different justifications for identical cases. Clients will not pay for more code. But they will be compelled to pay for confidence that their systems behave predictably.
Some argue that the answer is AI checking AI. When a model is used to verify another model’s output, you are not getting a second opinion. You are getting another sample from a related distribution. The errors that are improbable in one model may be equally improbable in the verification model — meaning they both miss the same things systematically. Without humans who understand both the why and the how, the enterprise loses its ability to recover from failure. The insurance industry understands this even if the technology industry does not. Insurers will find ways to cover AI-related failures while reducing premiums for organisations that can demonstrate genuine human oversight. Trust and certainty may become the only billable commodity in technology services. Not time and materials.
That is a huge shift. Can services firms make it?
Disclaimer: The views expressed are the author’s own and do not reflect the official policy or position of Financial Express.