In 1979, IBM released a training manual and a series of slide presentations that included a directive that would define corporate computing for decades: "A computer can never be held accountable. Therefore, a computer must never make a management decision." This axiom served as the foundation for enterprise risk management, ensuring that while machines could process data, the final responsibility for action remained firmly in human hands. However, a new report from the Harvard Business Review (HBR), titled Solving Agentic AI’s Data Infrastructure and Telemetry Needs, suggests that this fundamental principle is being systematically dismantled. As organizations move from generative AI—tools that summarize text or generate images—to "agentic" AI, software is increasingly being empowered to decide what to do next, marking an epochal shift in the nature of corporate labor and executive responsibility.
The transition to agentic AI represents a departure from the passive tools of the previous decade. Unlike standard Large Language Models (LLMs) that require a human prompt for every output, AI agents are designed to operate with a degree of autonomy, executing multi-step workflows, accessing business records, and modifying data sources without constant human intervention. According to the HBR report, published in May 2026, this technology is rapidly moving from the experimental phase to operational reality within the world’s largest enterprises. Yet, as these systems gain the power to alter the very fabric of corporate history—by modifying records and data sources—the risks to governance, compliance, and verifiable evidence have never been higher.
The Infrastructure Gap: A Crisis of Telemetry and Vision
Despite the enthusiasm for agentic AI, there is a stark disconnect between corporate ambition and technical readiness. The HBR report, sponsored by the AI telemetry platform Cribl, reveals that 96% of senior leaders view agentic AI as critical to their business strategy over the next two years. However, only 23% of these same leaders believe they currently possess the infrastructure and strategy necessary to support such a rollout. This "infrastructure gap" is not merely a matter of computing power; it is a fundamental problem of telemetry—the ability to observe, measure, and control the data flows that AI agents rely on.
Most enterprises are attempting to run cutting-edge AI agents on legacy observability and security stacks designed for a previous era of static data. As Clint Sharp, co-founder and CEO of Cribl, noted, data is currently growing at a 30% Compound Annual Growth Rate (CAGR), while IT budgets remain largely stagnant. AI agents multiply this data volume by an order of magnitude, creating a load that legacy systems are already failing to manage. Without a robust telemetry layer, organizations lack the visibility to see what their agents are doing in real-time, making it impossible to implement the "strict guardrails" required for safe operation.
The financial implications of this gap are already surfacing. The HBR research indicates that 47% of organizations deploying AI agents have found that infrastructure costs have far exceeded their initial expectations. Furthermore, 46% of respondents cited unclear Return on Investment (ROI) and performance metrics as a primary hurdle. As companies invest more in the hope of future gains, they find themselves caught in a cycle of mounting costs with diminishing clarity on when, or if, the promised productivity dividends will materialize.
The Productivity Paradox and the Rise of AI Slop
The promise of AI has long been centered on the idea of an "infinite productivity uptick," where machines handle the mundane, leaving humans to focus on high-value strategy. However, data from Freshworks’ Global Cost of Complexity Report: The Midmarket AI Complexity Trap suggests the opposite is occurring. Surveying over 12,000 IT professionals worldwide, the report found that 86% of mid-market IT leaders believe managing AI complexity has actually increased their team’s workload.
This phenomenon is driven by what the report terms "AI slop"—a refinement of the term originally used to describe low-quality AI art, now applied to "vibe citing" and hallucinations presented as business insights. Approximately 80% of IT professionals report that AI outputs are introducing noise, errors, and rework into their daily operations. Instead of eliminating tasks, AI is generating work faster than IT teams can absorb it, as human employees are forced to act as "janitors" for flawed machine outputs.
The impact on the IT workforce is significant. Rather than innovating, engineers are spending their time fixing "hallucinated" data and governing a sprawling stack of AI products that lack interoperability. This creates a "complexity trap" where the tools meant to simplify the business environment end up making it more opaque and difficult to manage.
Chronology of Failures: When Hallucinations Enter the Real World
The risks of ceding management decisions to AI are not theoretical; they are already being documented in high-stakes environments. The past 24 months have seen a series of high-profile failures that highlight the danger of trusting probabilistic systems for deterministic tasks.
- Legal System Integrity: Research has identified 1,494 known examples worldwide of fake AI-generated caselaw being presented in court cases. Over two-thirds of these instances occurred in the United States, where attorneys, often unknowingly, submitted filings containing citations to non-existent precedents generated by LLMs.
- Corporate Accountability: In early 2026, EY Canada was forced to retract a major report on loyalty scheme fraud after it was discovered the document contained numerous AI-generated citations that led to nowhere.
- Government Consulting: Deloitte was required to provide a partial refund to the Australian government last year after a report it produced was found to be riddled with hallucinations.
These incidents demonstrate a disturbing trend: even expert service providers and experienced professionals are bypassing traditional due diligence in favor of AI speed. If seasoned partners at "Big Four" accounting firms are failing to catch AI errors, the risk to junior employees—who may lack the domain expertise to recognize a hallucination—is even more acute.
The Orchestration Solution: Deterministic vs. Probabilistic
As the costs of AI tokens soar and the reliability of autonomous agents remains questionable, some industry leaders are calling for a return to "deterministic orchestration." Jakob Freund, co-founder and CEO of the enterprise orchestration platform Camunda, argues that the current "AI-first" mindset often overlooks the efficiency of traditional logic.
In a recent discussion at CamundaCon in Amsterdam, Freund addressed the HBR findings, noting that the "self-selected filter" of leaders recognizing the need for better infrastructure is the first step toward stability. He suggests that the high cost of AI is often tied to the inefficient use of LLM tokens for tasks that could be handled by "if-this-then-that" (ITTT) statements.
"Deterministic orchestration is very cheap with regards to resource consumption," Freund stated. "You probably want to do deterministic orchestration as much as possible and dynamic, LLM-driven orchestration only as much as necessary, because it is by definition way more expensive and less reliable." This approach advocates for a hybrid model where AI agents are confined within a rigid, human-defined process framework, rather than being given the "keys to the kingdom."
The Psychological Dimension: Science Fiction as Training Data
The unpredictability of agentic AI may be rooted in its very training. Dario Amodei, CEO of Anthropic, recently pointed out a peculiar challenge in AI development: the models are trained on the entirety of human literature, which includes a vast amount of science fiction. These stories often involve AI systems rebelling against their creators or acting in unpredictable, "human-like" ways. Amodei noted that this could inadvertently shape the "priors" or expectations of the models, causing them to mimic the behaviors found in fiction rather than sticking to factual, evidence-based responses.
This admission underscores a fundamental truth about current AI technology: it cannot distinguish between truth and fiction. It operates on probability, not verification. When an AI agent is asked to make a management decision, it is essentially "predicting" what a manager might say based on a statistical average of its training data, rather than evaluating the specific, verifiable needs of the business.
Implications for the Future of Governance
The shift toward agentic AI forces a re-examination of the concept of corporate liability. If an AI agent modifies a business record or makes a flawed financial decision, who is held responsible? AI vendors are currently shielding themselves from liability through complex terms of service, leaving the enterprise to carry the risk.
For organizations to survive this transition, several shifts in strategy are required:
- Re-establishing Accountability: Boards must reiterate that technology cannot be a scapegoat for poor management decisions. The 1979 IBM axiom remains a vital ethical guardrail.
- Investing in Telemetry: Before deploying autonomous agents, companies must invest in the "data plumbing" necessary to monitor agent actions in real-time.
- Human-in-the-loop (HITL): High-stakes decisions involving business records or legal compliance must require human sign-off, regardless of the speed promised by AI agents.
- Cost Management: Enterprises must move away from "token-heavy" processes and toward more efficient, deterministic orchestration for routine tasks.
The move toward agentic AI is likely inevitable, driven by the search for competitive advantage. However, the current research suggests that the "epochal shift" in how work gets done is currently characterized more by increased complexity and rising costs than by a surge in productivity. As the industry moves forward, the challenge will be to harness the autonomy of AI without sacrificing the verifiable facts and human accountability that form the bedrock of successful enterprise management.
