Why Your New AI Agents Are Costing More Than Planned

Prashanthi Kolluru is the founder of KloudPortal. Helping global capability centers (GCCs) hire product-ready engineering pods.

​If the past year has taught me anything, it's that building AI agents is no longer the hard part. Operating them is. ​

CIOs and technology leaders I speak with rarely open questions about model selection or prompt engineering—those decisions are largely behind them. What they want to understand now is why AI operating costs keep climbing even after a deployment goes well.

Recent research from McKinsey shows that enterprise AI adoption continues to accelerate, with organizations expanding AI across multiple business functions. But the budgeting conversation hasn't kept pace with that growth: Most companies planned carefully for development, and few anticipated the ongoing effort required to manage dozens, and eventually hundreds, of agents across the enterprise. As adoption grows, many are discovering that operating AI is a far bigger job than deploying it.

Most enterprise software projects follow a predictable cost pattern. The majority of spending happens during development, and once the application goes live, operating costs are generally easier to forecast.

AI agents don't behave that way. Deployment isn't the finish line for the investment; it's the start of a new operating cycle. ​

A production agent is constantly at work: retrieving enterprise data, evaluating prompts, invoking models, connecting to APIs, applying business rules, logging activity for governance and, in many organizations, routing sensitive responses through human review. None of these steps look expensive in isolation. Stacked together, they add up to an operating model that looks very different from traditional enterprise software.

That gap is easy to miss during a pilot. It's much harder to ignore once AI usage spreads across multiple business units.

There's a common misconception that higher AI adoption automatically improves the economics of a deployment. But after working with enterprise clients, in practice, the opposite often happens. ​

As business teams gain confidence, they start asking for more. Customer support wants richer responses. The operations team wants deeper integrations. Sales want agents that can tap into more internal knowledge. Compliance wants additional monitoring. Every one of those requests makes sense on its own, but together, they mean more model calls, larger context windows, more retrieval operations and heavier governance requirements.

The agent is doing more valuable work, but it's also burning through more resources to do it. In my experience, organizations underestimate this trade-off because they're still evaluating AI the way they'd evaluate conventional enterprise software. That comparison doesn't hold: AI runs on a consumption model, where every interaction carries its own operational cost.

Most technology budgets account for infrastructure, licensing and implementation. What they tend to leave out is everything it takes to keep AI agents running responsibly: monitoring output quality, evaluating model performance, updating prompts as business processes shift, maintaining data connections, enforcing governance policies and continuously measuring outcomes. None of that is optional in an enterprise setting.

That shift is still catching many organizations off guard. According to McKinsey's November 2025 State of AI survey (linked above), only 23% of organizations have scaled agentic AI in at least one business function, while 39% are still in the experimentation phase. Adoption remains limited across the enterprise, with fewer than 10% of respondents reporting scaled deployment in any single function. Most enterprises, in other words, are still early in exactly the phase where operating costs start to compound.

That gap shows up in governance data, too. Deloitte's State of AI in the Enterprise research finds that only 1 in 5 companies has a mature model for governing autonomous AI agents, even as governance, risk management and operational readiness remain among the biggest hurdles as organizations try to scale AI beyond pilot projects.

From where I sit, this is the point where AI programs either mature or start to strain under their own weight.

When AI spending runs over budget, the conversation tends to default to model pricing. That's too narrow a lens.

A better question is whether each agent delivers sufficient business value to justify the cost of running it. An agent automating a high-volume customer process can easily outweigh several cheaper agents solving narrow, isolated problems, even if its usage bill is larger.

That calls for a different way of measuring AI. Rather than asking, "How much does this model cost?" leaders get further by asking, "What business outcome does this agent improve, and what does it cost to deliver that outcome reliably?" That question is the one that gets technology, finance and business teams speaking the same language.

Cloud computing gave rise to FinOps because organizations needed real visibility into infrastructure spending. Enterprise AI is creating the same need, just faster. ​

As AI workloads become more dynamic and harder to predict, many organizations are beginning to apply FinOps principles to AI, using the same discipline that brought cost visibility to the cloud to bring accountability to AI operations.

The organizations best positioned for large-scale AI adoption aren't necessarily running the most sophisticated agents. They're the ones building operational discipline around them: reviewing usage patterns, measuring outcomes, retiring redundant agents and treating AI operations as an ongoing capability rather than a finished project.

Enterprise AI is progressing beyond the experimentation stage and becoming part of day-to-day business operations. That's often when organizations realize their original budgeting assumptions are very different from the costs of implementing it.

Building an AI agent has gotten a lot easier over the past year. Managing a growing portfolio of them with a clear read on operating costs, governance requirements and business impact is becoming a harder leadership challenge.

Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?