The adoption of agentic AI carries hidden costs that could become difficult for CFOs, CIOs and boards to measure and control, warns President Ntuli, MD of Hewlett Packard Enterprise South Africa.
A single prompt or automated workflow can trigger multiple tokens, compute demands, infrastructure pressure and governance risks, potentially turning AI adoption into a new form of shadow IT, he says.
According to Ntuli, business leaders must assess whether they are prepared for the financial and operational impact of scaling AI.
“When people think about the cost of AI, they often focus on the cost of a prompt or a model subscription," he says. "The reality with AI agents is far more complex."
What looks like a single prompt can trigger multiple agent-to-agent interactions, API calls, data retrieval requests, workflow executions and security checks behind the scenes, he explains.
"Each of these actions consumes tokens, compute resources, network capacity, energy and operational oversight. In some cases, agentic workflows can consume orders of magnitude more resources than a traditional single-pass AI interaction.”
Unlike cloud computing, which is billed by the hour, or software as a service, which is typically billed by seat, AI is billed using consumption units known as tokens. Even small changes in phrasing can affect token consumption and, ultimately, cost.
While consumption may appear manageable during pilots, costs can rise rapidly when AI agents operate continuously at production scale. Agents can also create chains in which one agent’s output becomes another’s input, multiplying resource consumption if unmanaged.
Left unmanaged, that consumption becomes an invisible meter.
As agents spread across business functions, AI sprawl can lead to unmonitored spending, governance challenges and operational risks. Organisations need visibility into which agents are running, which models they use, where resources are consumed, what data and systems they access, and the measurable value they deliver.
An AI agent may retrieve information from multiple systems, call external tools, interact with other agents, analyse results and execute a task. Each step creates additional transactions and resource consumption.
“As these interactions increase, organisations require more compute capacity, greater network performance and stronger operational oversight," says Ntuli.
There are also governance implications, he adds, because agents may access sensitive data, interact with business systems or initiate actions on behalf of users.
"This means organisations need clear policies, monitoring and accountability mechanisms to ensure AI remains secure, compliant and aligned with business objectives.”
The cost of AI, therefore, is determined not only by the model being used, but also by how work is orchestrated and data is processed.
Real spend sits underneath
Johan Steyn, AI expert and founder of AIforBusiness.net, says the real spend sits underneath the licence: consumption-based tokens and compute that scale with usage, data-readiness work, integration and maintenance, and ongoing governance.
Steyn suggests these costs are hidden because they are variable and accrue after deployment – agentic AI is metered like electricity, but most business cases are written before anyone models the meter.
“Financially, success is what gets expensive: because cost scales with usage, the better a pilot works and the more it's used, the higher the bill – turning AI into unpredictable opex." he says.
"Operationally, agents act rather than advise, so scaling them is an organisational redesign, not a software roll-out."
Cost is driven by consumption, Steyn explains: the model used, tokens consumed, the number of reasoning steps and tool calls, compute, plus fixed data and integration costs.
"There's no honest single figure: it runs from fractions of a cent per task to major monthly opex at scale, and anyone quoting a firm number without knowing the workload is guessing,” he says.
Share
