A hallucination is a false statement generated by an AI model that is not grounded in any training data or documented facts. It is essentially an arbitrary-fact failure where the system makes things up when it doesn’t know the actual answer, producing a response that looks convincing but is entirely fabricated.
This happens because the model isn't "aware" of its own mistakes; it is simply following its training to guess the next word without any built-in mechanism to say "I don’t know."
Hallucinations are a major issue in agentic technology because it makes AI agents unreliable when they are plugged into business processes. It affects everything from legal research to math computations, where a single fabricated fact can lead to a chain of wrong decisions. Because these errors are baked into the model's internal parameters, developers have to spend extra time and resources building secondary workflows and "evals" just to catch the error before it causes real-world damage.
Some scholars argue that these errors aren't actually a bug, but a fundamental "feature" of how the technology works. The criticism is that since models are probabilistic by nature, expecting them to be perfectly factual is a misunderstanding of the tool itself. Others suggest that the problem is overblown because we can eventually train models to reward "honesty," while some believe that the same creative "guessing" that leads to hallucinations is exactly what makes AI powerful and flexible in the first place.
In work automation, this creates a unique bottleneck where AI agents are expected to act autonomously. When an agent starts making decisions in a business process, a single hallucination can trigger a chain of incorrect actions without any human ever seeing the error.
