Enterprise conversations about Agentic AI are becoming more practical. Early questions focused on whether an agent could resolve a customer issue or automate a back-office task. Today, the discussion extends to autonomy, ownership and the cost of running entire workflows.

That was the central theme of a recent episode of our HCLTech Trends & Insights podcast featuring Padmaja Enjeti, who leads our modern applications development portfolio.

Padma described a shift in how enterprises think about Agentic AI, with organizations increasingly considering how to redesign the way they work around its capabilities.

“Where do we put autonomy? What happens when an agent scales? Who governs these agents? What does it cost to run an end-to-end workflow?” she asked.

The cost of a workflow is harder to predict

Traditional automation typically follows a known path. Enterprises can estimate transaction volumes and the computing resources involved. With Agentic AI, the work itself can take different routes depending on its complexity.

“An agent can reason. It can call a tool. It can retrieve information. It can invoke another agent. It can check its own work,” said Padma.

Each step consumes resources. An agent might retrieve more information, make additional model calls or loop back to check an answer. That changes the economics when a single assistant becomes a suite of agents coordinating a business process.

Padma described the question she hears from finance leaders: “I understand the value proposition, but can I predict the cost of achieving that outcome?”

Although individual token prices have been falling, greater consumption within complex workflows can increase total costs. The distinction between model efficiency and workflow efficiency is easy to overlook. A model might become 50% cheaper, Padma explained, but an agent making three times as many calls could still cost more to complete its work.

The token bill is only one component. Orchestration, APIs, data infrastructure, observability and human intervention also contribute to the overall economics.

Understanding where the consumption goes

According to Padma, problems often begin when teams start with the technology and only later ask what business outcome it is supposed to deliver. Experimentation can generate considerable activity without a clear way to measure its value.

Costs can also be difficult to see. Routing between models, handoffs between agents and repeated attempts to complete a task create consumption that may not be apparent from a dashboard showing only headline usage.

“You don't want to find out the cost after you get the invoice,” said Padma.

She argued for monitoring both what agents consume and what they produce. That means understanding model calls, workflow behavior and business key performance indicators rather than treating the number of agents deployed as evidence of success. Even where exact costs are uncertain, organizations need a reasonable view of the likely range before usage expands.

Learning from cloud FinOps

Padma sees a useful parallel with cloud financial operations (FinOps). Cloud cost management evolved beyond shopping for cheaper infrastructure toward understanding architecture, consumption patterns and who owns the expense.

For AI, the same questions apply, with an additional complication: Agentic AI consumption is probabilistic. The number of calls or retrievals required to deliver an outcome cannot always be known in advance.

“What are we consuming? Who's consuming it? What outcome is it producing? And could we have achieved the same outcome with less intelligence?” she asked.

That final question points toward model routing. Rather than sending every task to the same model, enterprises could direct work to the level of intelligence appropriate to its complexity and cost.

Padma described this as “intelligently routing work to the right level of intelligence for the right level of cost.”

She believes tokenomics needs to become an engineering and business discipline, with decisions about architecture and consumption directly connected to the value a workflow creates.

What leaders should ask before scaling

For Padma, the starting point is the measurable outcome an enterprise expects to buy. Leaders should then understand the cost of completing one transaction and ask what happens when volume grows from 1,000 to 10,000 or even a million.

They also need visibility into agent behavior: Which models are being called, which tools are used, where retries occur and when people intervene. Failure scenarios deserve particular scrutiny when an agent is embedded in a core business workflow.

“Governance is not something we can bolt on after deployment,” said Padma.

Bounded autonomy, escalation, auditability and clear decision rights need to be considered from the beginning. As agentic workflows become part of core business processes, those controls become integral to how organizations manage their operations.

Her final test for any initiative is simple: “Can we scale this economically and not just technically?”

As Agentic AI moves deeper into enterprise workflows, the answer depends on more than model capability. Padma's emphasis is on making consumption understandable, assigning ownership and showing the business value produced by the intelligence being used.

In the second part of this HCLTech Trends & Insights podcast series, Venkatraman Natarajan, who leads our AI-native Application Development practice, will explore tokenomics through an application modernization lens, examining what the economics of Agentic AI mean for modernizing enterprise applications. through an application modernization lens, examining what the economics of Agentic AI mean for modernizing enterprise applications.