Companies must adapt to rising AI costs, tech CEO says

For the past three years, the most costly technology experiment has been funded by others, according to Kristy Brown, chief executive of Fusion5. A white paper from the tech services provider highlights the need for a managed approach to AI agents to prevent excessive costs, while adhering to new regulations and business risks. Brown explained that many AI agents now operate on a pay-per-token basis rather than a flat user license or a hybrid model.

This includes popular coding and customer service agents like Anthropic's Claude, Google's Gemini, Microsoft's CoPilot, and others.

Initially, the focus was on gaining trust and encouraging AI usage. However, the current challenge is to use AI more efficiently through smart processes that enhance search accuracy and deliver quick results. With usage-based pricing, every prompt, reasoning step, and autonomous agent comes with a real, metered cost, and someone must bear the expense.

Autonomous agents perform tasks such as spending, reconciling transactions, and making decisions that impact the organization. They also consume significantly more resources than traditional software, often entering loops that generate substantial costs before anyone realizes the issue.

Monitoring and controls are crucial to prevent businesses from facing large bills and unexpected governance problems. Companies that build discipline in operating AI, not just deploying it, will be the ones to succeed when the free money and support cease. To improve performance, businesses should collect key business processes into an ontology, a structured description of their business model, including relationships, company details, executive information, financial results, and minor details like font and color preferences.

With this information, AI workloads can efficiently reach outcomes without the need for extensive reasoning steps.

Incomplete data or digitally digitized institutional knowledge forces businesses to ask AI agents to find unavailable information, leading to wasted compute resources and mistakes. According to Brown, businesses pay twice: once for the extra compute and again for errors. The transformation is already underway, with tech giants like Tesla, Uber, Meta, and Amazon implementing employee restrictions on AI agent usage as businesses shift to the traditional cost-measurement model.

The four largest US tech firms are projected to spend nearly US$700 billion on AI infrastructure this year alone, with OpenAI spending about US$60 billion annually on compute, generating roughly US$13 billion in revenue.

Microsoft is leading the pack in generating revenue through its AI Agent Copilot, integrated into its Office tools suite, according to its latest fourth-quarter earnings report. This year, Azure revenue surpassed US$100 billion for the first time, and Microsoft 365 Copilot reached over 30 million paid seats, reflecting customer confidence in the company's AI capabilities.

Kristy Brown proposes a dedicated function to monitor every agent in real-time, track costs with thresholds and alerts, and promptly alert human personnel when an agent deviates from its guidelines. This concept is the reason behind Fusion5's investment in building an Agent Operations Centre, which will be launched in the fourth quarter of the year, targeting small and mid-sized businesses that lack the expertise to manage AI agents independently.