The trouble with companies using AI for the sake of it

It has become almost fashionable for companies to claim they are using AI. Every boardroom, every quarterly report, every strategic offsite seems to revolve around the same language of transformation. But what is quietly emerging beneath all the noise is a pattern of misplaced ambition with businesses applying advanced technology to areas that barely need it, while neglecting the parts of their work that could be fundamentally reimagined. 

A recent McKinsey & Company article on the first year of agentic AI argues that this technology, defined by its ability to act autonomously rather than just predict or generate, exemplifies the above misalignment perfectly. According to the report, many organisations are building agents not to solve the right problems, but to showcase that they can. So, the more telling question is not what AI can do, but where it should do it.

This tension is particularly visible in Bangladesh and similar Global South contexts, where firms are experimenting with AI without fully understanding the nature of their own work processes. The temptation to adopt the latest technology is strong. But enthusiasm alone cannot make AI effective.Success with agentic AI requires an understanding of the work itself, while understanding its patterns, variability, and dependence on human judgment.

Not every job benefits from autonomy. Low-variance, high-standardisation tasks like data reconciliation or payroll processing do not need a cognitive engine. They need efficiency, not intelligence. High-variance, low-standardisation work, such as strategy, creativity, and negotiation, remains distinctly human. It is the messy space of judgment, instinct, and error, where the cost of misinterpretation outweighs the benefit of speed.

The true potential of agentic AI lies in the middle ground: tasks that have sufficient structure to define workflows but enough variability to require decision-making, judgment, or coordination. These are workflows that repeat often enough to define, but vary just enough to require decisions along the way.

Examples include supply chain management processes that require adaptive routing, customer service workflows that escalate in complexity, or financial operations where data is standardised but decisions are contingent on market conditions. In these contexts, AI can act as a collaborator rather than a substitute, handling execution while humans focus on oversight, interpretation, and higher-order judgment. In most use cases, this middle ground is often neglected.

It needs to be kept in mind that agentic systems are not plug-ins; they are co-workers. They demand clarity of process, auditability of action, and a willingness to redefine what roles mean. The companies that fail often do so not because the technology underperforms, but because they refuse to redraw the boundaries of their work. The real shift will come when organisations begin redesigning workflows around the agent itself. The most effective AI deployments so far have been those that respect the tacit knowledge humans bring. A system can decide, but only if someone first defines what good decisions look like.

For many economies still defining their digital maturity, this reflection feels urgent. In places where processes are semi-structured and still heavily dependent on human oversight, the temptation will be to leapfrog directly into agentic AI. But skipping the slow work of standardisation first is a mistake. Before we can build agents that act, we must build workflows worth acting within.

In the end, the real transformation would not come from what AI can do, but from what companies learn to stop doing. The compulsion to automate everything, to equate progress with replacement, blinds us to the nuance of good design. Not all inefficiency is waste; not all human friction is bad. Some parts of work are supposed to stay unpredictable. And agentic AI, if deployed thoughtfully, may help us find balance again.