Grounding, explained for business owners

Grounding means an AI's answers come from a specific set of facts you control (your services, prices, hours, policies, approved FAQs) rather than from whatever the underlying model absorbed during training. A grounded AI receptionist asked about a service you do not offer says it does not know. An ungrounded one describes the service. Grounding is the difference between an assistant you can put in front of customers and one you cannot, and it is the first thing to test in any demo.

The model knows things; that is the problem

Large language models are trained on enormous amounts of text and can produce a plausible answer to almost anything. Asked "do you take Delta Dental?", a model with no information about your practice will produce an answer anyway, because producing plausible text is what it does. The answer might be yes. It is not based on anything.

What grounding changes

A grounded system does not ask the model "what is the answer". It asks "here are the approved facts; using only these, answer the question, and if they do not cover it, say so". The model's job becomes reading and phrasing, not knowing. Your facts become the only source, and a question outside them gets an honest "I am not sure" rather than a confident guess.

What "your facts" should be

Short, specific entries, one per fact: services with durations and prices, locations with hours, policies, staff and what they do, and FAQs you have approved word by word. Not a pasted brochure or a website scrape used as-is; those are a starting point for drafting facts, which you then approve. The approval step is what makes it yours.

Testing whether a product is grounded

Ask about something you did not load. A service you do not offer, an insurance plan you did not mention, a location you do not have. The right answer is a variant of "I do not have that information; let me check with the team", and the question should appear in a gaps log afterwards. A confident answer is a fail, however good it sounds. Then ask something you did load and check the answer matches your fact exactly, including the price.

Grounding is not the same as a good prompt

A vendor may say "we instruct the model to only use your information". An instruction is a request the model usually follows. Grounding as an architecture means the model is given your facts and nothing else to work from, retrieval picks the relevant facts per question, and the answer is checked against them. Ask which one the product does. The difference shows up on the hard questions.

What grounding does not solve

Facts that are wrong (a price that changed in March). Facts that are ambiguous ("depends on the plan"). Questions that need judgement (whether a treatment suits someone), which no fact answers and which must be handed off. Grounding makes the AI honest about what it knows; it does not make what it knows correct or complete. That is the gaps log and the monthly review.

Why it decides trust

Every other failure of an AI receptionist (a wrong price, a promised slot that did not exist, a confident yes about coverage) is a grounding failure or a stale fact. A grounded system with a maintained fact list can be supervised for a week and then trusted. An ungrounded one cannot be trusted at any point, because the next question might be the one it invents.

Frequently asked questions

Does grounding make the AI less helpful?

It makes it say "I do not know" more, which feels less helpful and is more useful. A customer told "let me check" gets a correct answer later; one told a guess gets a wrong one now.

Can it be grounded on my website?

The website can be the source for drafting facts. Grounding on the raw site means grounding on marketing copy, which is often vague or out of date. Draft, then approve.

How do I keep the facts current?

One owner, updates the same day a fact changes, and a monthly review of the gaps log. See training an AI receptionist.

What Frontiva does here

Frontiva's AI agents answer only from structured facts and approved FAQs in knowledge; if nothing matches, the question is logged as a gap, counted, and not guessed. Every reply shows the source it came from. The AI does not read your website or PDFs yet, so facts are entered directly, seeded from your industry template. See knowledge gaps.