Reef TRHA customer once asked a company's shiny new support assistant whether they could return something after 45 days. The assistant replied instantly, warmly, and with total confidence: yes, of course, you have 90 days. The real policy was 30. Nobody had ever told the model the policy, so when the question landed on a gap, it did what these models do and produced something that sounded exactly right.
That confident invention is the fastest way there is to lose trust in an AI, and it has a single root cause: the assistant was answering from its own memory instead of from the company's facts. The fix is not a cleverer model. It is handing the model your sources and insisting it answer from them.
Answer from the sources, not from memory
A grounded system does something a plain chatbot doesn't. Before it answers, it retrieves the passages that actually bear on the question from your own material, the returns policy, the signed contract, the customer's own record, and hands them to the model along with the question, with one instruction: answer only from what you were just given, and show where each claim came from. Now the customer hears "30 days, under the returns policy," with a link to the clause. The citation is not decoration. It is how a person checks the answer in seconds instead of taking it on faith.
Teach it to say "I don't know"
The other half of grounding is permission to refuse. When the retrieved sources genuinely don't contain the answer, the system should say so and pass the question to a person, not smooth over the gap with fluent fiction. That is the same confidence-floor discipline behind proving an AI is right before launch: below a set bar, it defers rather than guesses. An assistant that knows the edge of what it knows is far more useful, and far safer, than one that is always sure.
The answer is only as good as what it can read
Grounding quietly turns accuracy into a retrieval problem. If the right document is missing, mislabelled, or scattered across three systems, the model has nothing correct to ground on and you are back to guessing. This is why serious grounded-AI work usually starts with the unglamorous part, cleaning and structuring the sources and scoping who is allowed to see what, so the assistant only ever retrieves what the person asking is permitted to see. One insurance group's assistant kept inventing answers until exactly that groundwork was done, and the same rebuild that fixed the accuracy is what made it safe to put in front of staff.
The gap between a chatbot that demos beautifully and a system you can put in front of a customer, or a regulator, is not eloquence. It is provenance. Retrieve the answer from a real source, cite it, and let the system admit the gaps it can't fill. Confidence is cheap and a model will hand it to you for free. Being right, and being able to show your work, is the thing worth building.
Common questions
What does it mean to "ground" an AI's answer?
It means the system retrieves the relevant passages from your own documents and has the model answer only from those, citing the source, instead of relying on the model's general memory. The answer is traceable back to a real document.
Will grounding stop the AI from making things up?
It reduces it sharply, especially when paired with a confidence floor that makes the system say it doesn't know and hand off when the retrieved sources don't cover the question, rather than inventing a fluent answer.
Do we need perfect data before we start?
No, but the answer is only ever as good as what the system can retrieve. Cleaning, structuring and scoping the key sources is part of the work, not a prerequisite you must finish first.