Artificial Intelligence is often presented as transformative and almost limitless in potential. Yet, like any tool, AI has important limitations. Understanding these boundaries helps us apply it responsibly, avoid unrealistic expectations, and focus on where humans must remain in the loop. Pascal Bornet’s work on Agentic AI emphasizes that recognizing limits is part of designing better agents and systems.
1. Data Dependency
AI systems rely heavily on data — its quality, quantity, and relevance. If the data is biased, incomplete, or outdated, the AI’s outputs will reflect those weaknesses. This makes AI less effective in contexts where reliable data is scarce or expensive to obtain.
2. Narrow Intelligence and Reasoning
Today’s AI excels at narrow, well-defined tasks but lacks true general intelligence. Even advanced models cannot easily transfer knowledge between domains. They are powerful specialists, not universal problem-solvers. They lack common sense reasoning. AI does not reason, feel, or understand context the way humans do. It generates outputs based on patterns, not lived experience or moral frameworks. It doesn’t reason ethicly. This gap is why human oversight remains crucial in judgment-heavy areas.
3. Explainability and Transparency
AI systems often operate as “black boxes.” Their internal decision processes are difficult to interpret, which can undermine trust. This opacity becomes critical in sensitive fields such as finance, law, or healthcare, where accountability is essential.
4. Ethical and Social Constraints
AI can amplify biases, threaten privacy, or be misused in harmful ways. These risks are not technological flaws alone but arise from how people design, deploy, and regulate AI. Limiting misuse requires strong governance and ethical standards. AI can hallucinate, meaning that they generate false information with high confidence. AI agents “fill in gaps” in their knowledge with entirely fabricated knowledge that sound plausible. AI agents don’t understand the deeper moral implications that go beyong the rules.
5. Resource Intensity
Large AI models consume massive amounts of computing power, energy, and specialized hardware. This makes them costly to train and operate, raising sustainability questions that must be balanced against benefits.
Healthcare Example: Avoiding Black-and-White Thinking
Consider healthcare, where AI is increasingly used for diagnostics and patient monitoring. A limitation arises when clinicians or patients expect “perfect” predictions. AI might flag risks with probabilities, but human minds often default to black-and-white thinking: “safe” or “unsafe,” “healthy” or “ill.” Here, a well-designed agent can counter this cognitive distortion by presenting nuanced probabilities and reminding users that outcomes are rarely absolute. This aligns with Cognitive Behavioral Therapy’s insights that rigid thinking distorts reality — and AI can be a tool for reframing, not reinforcing, those distortions.
Working With These Limitations
We sould not completely dismiss AI agents because of their imperfections. We can design systems that can leverage their strengths while providing human oversight. AI agents do well with data intensive tasks and the LLMs do well with writing and summarizing text. Another way around the limitations of AI agents is to use more than one of them. We can set up mult-agent systems.