If the problem can be solved by an if-check, don’t ask AI to do it: Sumanta Ghosh, CTO, Bandhan Life

For the past two years, enterprise technology has been consumed by one question. How quickly can organisations adopt AI?

The question is now changing. As enterprises move beyond pilots and proof-of-concepts, another, far more difficult question is emerging. How much AI actually makes business sense?

For Sumanta Ghosh, CTO of Bandhan Life, this is no longer a technology debate. It is an economic one.

A year ago, enterprise conversations revolved around generative AI. Today, every boardroom wants to discuss agentic AI, autonomous systems capable of making decisions and taking actions. Yet beneath the excitement lies a growing reality. AI may be becoming more capable, but it is also becoming more expensive. Infrastructure costs, token consumption and increasingly complex deployment models are forcing organisations to rethink the relationship between innovation and return on investment.

“The expertise lies in figuring out the balance so that you can take the good part of AI while putting proper guardrails around the bad part,” Ghosh says.

His philosophy stands in contrast to the prevailing narrative that AI should be embedded into every business process. Instead, he argues that technology leaders must first understand the nature of the business problems they are trying to solve.

Why insurance demands a different AI strategy 

Insurance, after all, is built on predictability. “Organisations mostly run on determinism,” Ghosh says. “For an insurance company, I would say nearly ninety percent of the work is deterministic.”

That distinction fundamentally shapes Bandhan Life’s AI strategy. Traditional enterprise systems are designed to produce the same outcome every time. AI, by design, does not. Ask the same model the same question twice and there is no guarantee that the answer will be identical. For consumer applications, that variability may be acceptable. In a regulated industry handling financial transactions and customer data, it is considerably harder to justify.

Instead of replacing deterministic systems, Bandhan Life is positioning AI as an intelligent layer around them. “We are taking a copilot approach,” Ghosh explains. AI can recommend an action, analyse information or generate insights, but a human expert remains responsible for approving the final decision before it reaches production or the customer.

That philosophy extends across the organisation. The company is evaluating AI-driven outbound renewal reminder calls and automated quality analysis of customer conversations. Each initiative follows the same principle. AI accelerates work, while humans remain accountable for decisions. The approach is not driven by caution alone. It is driven by economics.

One of the biggest misconceptions surrounding enterprise AI, according to Ghosh, is the assumption that every process benefits from AI. In reality, applying expensive AI models to problems that conventional software already solves efficiently creates cost without proportionate value.

The hidden cost of intelligence 

“If the problem can be solved by a simple parameter or an if-check, just do that. Don’t ask AI to do it,” Ghosh notes. It is an observation that resonates far beyond insurance.

Across industries, organisations are discovering that AI operating costs are often higher than anticipated. Large language models introduce recurring inference costs, token consumption grows invisibly with usage and experimentation scales faster than budgets. Several enterprises have already begun tightening AI governance after discovering that unrestricted adoption quickly translates into unexpected expenditure.

Ghosh believes the industry is now entering a correction phase similar to what cloud computing experienced more than a decade ago. “When cloud started, everybody simply lifted and shifted workloads,” he recalls. “Then costs started shooting up.”

Only later did organisations realise that cloud was not a destination but an architectural discipline. Success depended less on migration than on deciding which workloads belonged in the cloud and which did not.

AI, he believes, is following exactly the same trajectory. “They are the how part. They are not the what part.” In other words, AI is another enterprise tool. It should not become the strategy itself.

The comparison goes even further back. Having witnessed multiple technology cycles throughout his career, Ghosh sees familiar patterns repeating themselves. The dot-com boom promised to transform every business overnight before market realities intervened. Cloud computing followed a similar path. AI, too, has experienced its own wave of inflated expectations.

Every disruptive technology, he argues, begins by presenting itself as the answer to every business problem. Eventually, organisations learn where it genuinely creates value and where conventional approaches remain superior. Today’s debate around token costs illustrates that shift.

Rather than treating token expenditure as a temporary inconvenience, Ghosh sees it as a signal that enterprises are beginning to mature in their AI thinking. Leaders are no longer measuring success by the number of AI deployments. They are evaluating the business case behind every deployment.

“The people who create these models don’t know your context,” he says. “You know your context.” That single word, context, becomes central to his philosophy.

Technology vendors build increasingly powerful generic platforms. Enterprise leaders must determine where those platforms fit within their own operational realities, regulatory obligations and financial constraints. The competitive advantage no longer comes from adopting every new capability. It comes from orchestrating them intelligently. Architecture therefore becomes as important as AI itself.

Good architecture, according to Ghosh, also determines how effectively organisations respond to evolving regulations. Whether it is data privacy requirements or future compliance mandates, companies with clearly separated systems and well-defined data boundaries are inherently more adaptable. Businesses that prioritise architectural discipline early often discover that regulatory compliance becomes significantly easier later.

The same principle applies to AI adoption. Organisations require frameworks that distinguish between deterministic workflows, where traditional software remains sufficient, and situations involving ambiguity, judgement or contextual reasoning, where AI can genuinely contribute.

Looking ahead, Ghosh remains optimistic despite the current scrutiny surrounding AI economics. History suggests that the cost of transformative technologies inevitably declines. Cloud infrastructure became cheaper as adoption increased and providers improved efficiency. AI, he believes, will undergo a similar evolution. Models will become more efficient, infrastructure will mature and sustainability challenges surrounding compute-intensive workloads will drive another wave of innovation.

The economics, however, will only improve for organisations that maintain discipline during the transition. “The entry barrier will continue to reduce,” he says. “But you have to know when to use it, how to use it and where to use it.”

In many ways, that observation captures the next phase of enterprise AI. The first wave rewarded experimentation. The second rewarded adoption. The third is likely to reward judgement.

For technology leaders, the challenge is no longer proving that AI works. That debate has largely been settled. The more important challenge is determining where AI creates measurable business value without introducing disproportionate cost, complexity or risk.

In the race to build AI-first enterprises, the organisations that ultimately succeed may not be those deploying the most AI. They may simply be the ones disciplined enough to deploy it only where the return justifies the investment.