What is a hallucination?
A confident-sounding statement a model made up, not something it flagged as a guess.
Want to see it first? Book a demo
Hallucination — A hallucination is a confident, plausible statement produced by a language model that is not supported by its inputs or by fact. In a CRM context it means an invented meeting, quote or commitment.
Why it matters
A language model writes the most likely next words for a prompt. It isn't looking anything up in a database. Most of the time those words match reality closely enough to be useful. Sometimes they don't, and the model states something false with the same smooth confidence as something true. You might see a meeting that was never booked, a price that was never quoted, or a promise nobody made. In a CRM, that's no harmless quirk. A made-up promise can end up in front of your customer.
Here's the risk: a hallucination reads just like a right answer. The model puts no italics or warning label on its own guesses. Ask it to double-check and you often get another sure-sounding answer, not an honest "I don't know". So ground the model in real data wherever you can. Use retrieval-augmented generation, or have it work only from the record's own timeline. When the model states an outside fact, a date, a number or a promise, check it before it reaches a customer.
How to reduce hallucination risk
- 1
Ground the model in real data
Retrieval-augmented generation or the record's own timeline, not open recall.
- 2
Ask for specifics, not summaries
A model asked for an exact date or figure is more likely to guess than one quoting a source.
- 3
Keep external claims out of auto mode
Anything stated as fact to a customer should be reviewed first.
- 4
Cite the source
Output that names which document or record it drew from is easier to check.
- 5
Review anything customer-facing
Before a claim about dates, prices or commitments goes out.
The mistake to watch for
Questions
- How is a hallucination different from an error?
- An error is usually a mistake you can trace to a cause, like bad input data or a bug. A hallucination is the model making up a plausible claim with nothing behind it, and it needs no bad input to happen.
- Can retrieval-augmented generation eliminate hallucination?
- No. It cuts the risk when it pulls the right source passage. If that lookup misses, or the model ignores what it was given, it can still state something with no basis.
- Should every AI-generated statement be reviewed?
- Not every one. Anything that states an outside fact, a date, a price or a promise should be checked before it reaches a customer. That's where a hallucination does real damage.