Glossary
What is an AI hallucination?
An AI hallucination is an answer from an AI model that sounds confident and plausible but is wrong or made up, such as a product feature that doesn't exist, a price nobody set or a source that was never written.
In plain English.
A large language model writes what is most likely to come next, based on patterns in what it learned. Most of the time, what sounds right is right. When the model doesn't have the information it needs, it still writes something that sounds right, in the same confident tone. That's a hallucination. It isn't lying, and it isn't a fault in one model: it's how generative AI works, and good systems are designed around it.
An example. A trade buyer asks a storefront assistant whether a pump is rated for outdoor use. The datasheet doesn't say. An assistant left to itself might answer "Yes, it's IP65 rated", because many similar pumps are. A well-built one answers from the datasheet, says the rating isn't listed, and offers to pass the question to your team.
Why it matters
In business, a wrong answer can become a promise: a delivery date, a discount, a compatibility claim. That's why the answers that matter need to come from your own information, and why AI that acts on your systems needs checks before anything commits you.
How it's kept in check
- Grounding in your sources. With RAG, the AI answers from your documents and data rather than from memory, and can show where each answer came from.
- Confidence on every output. The system scores how sure it is, and below a threshold you set it shows a person or escalates instead of answering.
- A person when it matters. When it isn't sure, it hands over to your team with the conversation so far. That's human in the loop.
- Checking before it's seen. Replies are checked against your content or the customer's account before they go out.
- Watching it in use. Tracking how often a person overturns an answer, and whether confidence matches how often it's right.
How we do it
The assistants and agents we build are grounded in your own sources, show their confidence on every output, and hand over to a person when they aren't sure. Anything that commits money, stock or a promise is approved by a person first.
Related terms and services.
RAG
The main way to ground AI answers in your own information.
Human in the loop
Where a person checks, approves or takes over from the AI.
AI managed services
Watching quality, confidence and cost once your AI is live.
AI agents and automation
Agents with confidence on every output and limits in code.
Going deeper.
How confidence is scored and used to decide when an agent acts, asks or escalates: Confidence-based escalation in AI agents: the engineering guide, from Enceladus AI, the PureNet Group's AI strategy company.