In a recent LinkedIn post, Clare Kitching discusses the often-overlooked complexities and variability in the cost of AI agents, urging business leaders to adopt a more nuanced understanding beyond simple software expense models. Kitching highlights that the cost of running an AI agent can fluctuate dramatically, sometimes by as much as 30 times for the same task performed by the same agent.
This variability, she explains, stems from the fundamentally different operational nature of AI agents compared to traditional software. A key point Kitching makes is the significant difference in token usage between conversational AI and agentic tasks:
“The other big leap is that agentic tasks can use around 1,000 times more tokens than a chat-style task (McKinsey).”
Kitching emphasizes that this inherent cost variability is frequently underestimated in business cases, potentially leading to unexpected expenses for executives.
Key Drivers of AI Agent Costs
Clare Kitching outlines six primary drivers that contribute to the cost of AI agents, as identified by McKinsey. Understanding these factors is crucial for accurate budgeting and cost optimization.
1. Contextual Processing
Agents frequently re-read instructions, documents, and historical data at each step of their operation. This continuous need to process context significantly adds to the computational load and, consequently, the cost.
2. Refinement and Error Correction
A substantial portion of an agent’s operational expense is dedicated to refining its outputs and correcting errors. Kitching notes, “Checking and repairing an answer costs more than writing it. About 60% of an agentic task’s cost goes to refining answers.” This iterative process of quality control is a major cost factor.
3. Autonomy and Decision-Making
The autonomous nature of AI agents means their path, the tools they select, and their ability to retry tasks can all influence the final bill. These dynamic choices, made by the agent itself, contribute to cost fluctuations from one run to another.
4. Intelligence Level Selection
Kitching advises a strategic approach to model selection, suggesting that the most powerful model is not always the most economical choice. “Pay for deep reasoning only where it changes the answer,” she recommends, advocating for matching the intelligence level to the specific task requirements to manage costs effectively.
5. Orchestration Complexity
The interaction between multiple agents, the number of tool calls made, and the hand-offs between different components of a system all add to the overall cost. Kitching points out that the system’s design alone can dictate cost variations by an order of magnitude.
6. Information Design and Input Handling
A significant portion of the expense is associated with the input side of the operation. Kitching clarifies, “Most of the cost comes from the input side: what you asked the model to read, not what it says back.” This highlights the importance of efficient data preparation and prompt engineering.
Design Decisions as a Cost Control Lever
Looking at these drivers, Kitching observes that five out of the six are directly influenced by design decisions. This suggests that organizations have more control over AI agent costs than they might initially realize.
“This means you have more control over your agent cost than you think.”
She notes that many teams are beginning to discuss their AI expenditures, and some have been caught off guard by unexpected billing. When executives inquire about cost optimization, Kitching’s core advice is to begin with a thorough review of the system’s design.
“So when your CFO asks how do we optimise, start with a design review.”
By focusing on the design elements, businesses can proactively manage and optimize the costs associated with deploying AI agents, ensuring that these powerful tools deliver value without incurring prohibitive expenses.
📝 About This Content
This article is based on insights shared by Clare Kitching on LinkedIn.
📅 Originally posted on September 3, 2026 | View original post on LinkedIn →