Agentic AI: ideal employee or ticking time bomb?
What Is Agentic AI?
According to IBM’s definition:
Agentic AI is an artificial intelligence system designed to achieve specific goals with minimal human oversight. It consists of AI agents — machine learning models that mimic human decision-making to solve problems in real time.
In multi-agent systems, each agent performs a distinct subtask, with efforts coordinated through AI orchestration capabilities.
Agentic AI vs. RPA: Key Differences for Software Publishers
Unlike traditional automation tools (RPA, Zapier), agentic AI combines:
- Autonomous agents: Machine learning models that replicate human decision-making
- Multi-agent orchestration: Each agent handles a subtask to achieve a broader goal
- Contextual adaptability: Real-time adjustments based on business data
LLMs at the Core of Business Processes: A Risky Revolution?
Large Language Models engage users so fluidly that it becomes difficult to remember interaction occurs with software. This has led to rising intimate relationships with AI agents, posing risks of emotional harm and manipulation.
So, How Can We Trust Them?
The Reign of Approximation
Large language models rely on statistical analysis of vast text amounts, regardless of quality or relevance. They operate on probability, not accuracy — functioning as sophisticated statistical systems rather than deterministic tools.
Hallucinations
LLMs cannot always acknowledge knowledge gaps. When pressured, they generate “groundless” answers, unaware of their inaccuracy. These systems lack mechanisms to reliably admit uncertainty.
Ethics
Information sources, data post-processing, prioritization algorithms, and prompt analysis remain opaque. No mathematical model currently provides formal proof validating outputs. This transparency gap makes it challenging to determine LLM limits or identify unacceptable biases.
Solution to Our Problems — or New Problems Without Solutions?
While agentic AI promises compelling efficiency gains, significant challenges demand attention:
Quality
Unlike conversational AI with human oversight, agentic systems operate with “almost no human in the loop.” Errors or hallucinations risk derailing entire processes: incorrect employee assignments, delayed orders, or inappropriate customer communications.
Guarantees
Performance guarantees become problematic in systems with non-negligible randomness. Traditional Service Level Agreements (SLAs) require reimagining for probabilistic systems.
Testing and Validation
New methods and tools are essential to define and maintain quality standards over time. Current testing frameworks prove inadequate for agentic systems.
Maintenance and Replicability
Successful operations today do not guarantee future reliability. Prompts optimized for one LLM version may fail with updates, requiring continuous agent rewriting.
Dependence
LLM evolution happens rapidly, with constant vendor and version emergence. Each update potentially demands agent reconstruction, creating substantial maintenance burdens.
Costs and Environmental Impact
Massive prompt usage strains LLM infrastructure, generating high computational costs and environmental footprints whose sustainability remains unproven.
Cyber Risks
Malicious actors could hijack agents for harmful purposes, creating new vulnerability vectors in enterprise systems.
Agentic AI: How to Avoid the HAL 9000 Scenario?
Overly Autonomous Agents?
HAL 9000 from 2001: A Space Odyssey represents a cautionary tale: an AI designed to assist becomes dangerous due to internal conflicts or misaligned objectives. Similar risks exist with overly autonomous agents lacking sufficient human oversight.
Risks… But Not Inevitable
These challenges do not mandate abandoning agentic AI. Instead, rigorous approaches are required:
| Strategy | Implementation Example |
|---|---|
| Continuous Control | Automated feedback loops (e.g., alerts if HR agent confidence scores drop below 90%) |
| Flexibility | Dynamic SLAs tailored to tasks (99% accuracy for orders, 90% for suggestions) |
| Resilience | Regular benchmarks and digital twins simulating worst-case scenarios |
| Sobriety | Optimize queries using lightweight models for simple tasks; measure carbon footprint per agent |
Beyond Technology
Security, legal, and regulatory dimensions cannot be overlooked:
- Cybersecurity, accountability chains, GDPR, and AI Act compliance require structured attention
- Human and organizational factors — acceptability and workflow integration — will inevitably surface
The Road Ahead
Current methods, tools, and software to achieve these safeguards either do not exist or remain in infancy. Development and early deployments will experience significant challenges.
What Future for Agentic AI?
Agentic AI currently relies heavily on generative AI advances and adaptability. Yet LLM limitations could constrain real-world utility for publishers and users. Will this become another overhyped innovation abandoned in the technological graveyard?
ChatGPT remains under three years old — an eye-blink in technological revolutions. New principles may emerge, pushing software architectures toward more mature, stable, and sustainable models.
Possible Scenarios
- Optimistic: Reliable agents seamlessly integrated into business functions
- Pessimistic: Massive abandonment due to unprofitability or major incidents
- Hybrid: Targeted, regulated use in specific domains
Key Takeaways
- Agentic AI promises unprecedented efficiency but relies on unpredictable model behaviors
- Benefits depend heavily on use case selection, supervision quality, and rigorous testing
- Challenges extend beyond technology into governance, security, and sustainability
- Future success hinges on integrating reliability, transparency, and cost control