Agentic AI: Why business hopes are rarely fulfilled

Contents

When AI agents are discussed outside the IT department, the conversation revolves around lower costs, greater efficiency, and staff reductions. Prominent examples include headlines: Salesforce was able to reduce its customer service team from 9,000 to 5,000 employees with the help of AI agents. In the meantime, their digital agents have handled over 1.5 million customer conversations. This creates the impression that AI agents primarily replace employees – making their deployment amortize in a very short time.

Eine Analyse von Harald Weiss

Harald Weiss is a freelance journalist and consultant in the fields of IT and telecommunications, specializing in industrial and business applications and software development.

Consequently, Agentic AI has already become a top priority for many companies to achieve maximum business benefits. Cost is hardly a factor: According to Deloitte, 67 percent of companies are increasing their GenAI investments, and 78 percent are planning further AI investments in the coming fiscal year. The expectations for these investments are clear: if a system independently closes tickets, adjusts schedules, or orchestrates workflows, it must be reflected in lower costs. However, this does not seem to be the case. In a recent PwC survey, 56 percent of CEOs say that their company has so far realized neither revenue nor cost benefits from AI; only 12 percent report that both have been achieved.

Agentic AI is not classic automation

The reasons for the weak business success are systemic. The assumption of automatic cost savings is based on the experience of classic workflow automation, which follows a linear pattern: a stable process is digitized, and a manual step is eliminated. The effect is directly measurable; effort and benefit can be clearly assigned.

Agentic AI works differently. Here, decision logic is delegated to a probabilistic system that combines tool use, context states, and dynamic data sources. The effort does not scale with the number of tasks, but with the complexity of the decision space. Each additional tool integration increases the number of possible system states. With each level of autonomy, the decision space grows combinatorially – and thus the testing and validation effort. Agentic AI does not behave like classic process automation, but like a continuously operating distributed system with a non-deterministic core.

The numbers reflect this. According to Deloitte, AI projects often only amortize after two to four years; only six percent report a Return on Investment (ROI) within one year. For agent-based systems, only ten percent see a significant ROI.

The underestimated expenses

When planning the productive use of AI agents, the model is the primary focus: a powerful LLM, better reasoning capabilities, fewer human interventions. However, this is the smallest problem, because a productive AI agent is not a model, but an orchestration stack of LLM, tool use, state management, memory, policy layer, logging, monitoring, and much more. Since an agent not only reads but also "writes" – i.e., closes tickets, triggers orders, or changes system parameters – transaction security becomes a critical factor.

It is particularly difficult to attribute the many additional expenses to an AI agent. This includes, for example, data integration: agents require consistent, context-rich information. In practice, this is distributed across ERP, CRM, MES, or proprietary systems. Semantic inconsistencies, missing identifiers, and different time references make "Context Stitching" a standalone project. Retrieval pipelines, embeddings, access controls, and caching must be built and maintained.

The inference costs are often the smallest part. Another item is the interface logic: tool calls expand the error surface. Idempotency, rollback strategies, timeouts, and race conditions become real operational risks. Each additional integration increases the state complexity of the overall system, which must be adequately tested.

And finally, there are not only personnel savings because, in autonomous systems, part of the work shifts: the agents must be monitored, tool calls must be logged, decisions must be traceable, and edge cases must be analyzed. Observability is not an option here, but a mandatory requirement. Drift, regressions, and version conflicts between model, prompt, and tooling are not exceptions.

Cost block: Security effort

All of this means, that with increasing autonomy, the security effort increases: guardrails, policy engines, access restrictions, and fallback mechanisms must be implemented and maintained. Human-in-the-loop is often replaced by Human-on-Call – people no longer intervene regularly but must be available when the system goes off the rails. While token costs can be calculated well, the ongoing costs for monitoring, incident handling, compliance checks, and system maintenance are regularly underestimated in business cases.

Agentic AI has reached a considerable level of maturity: the models work, tool use and orchestration are solvable. However, economic viability is not trivial. The more agentic a system becomes – meaning the more it independently decides and executes – the more it behaves like a complex distributed system with a non-deterministic core. Integration, security, and operation then dominate the cost structure. The ROI therefore rarely fails due to inference performance, but rather due to the assumption that the autonomy of an AI agent is equivalent to classic workflow automation.

(mack)