AI That Cannot Make Things Up: Grounding Answers in Your Data

A customer asks the chat assistant on a store’s website which model fits their needs. It answers warmly and confidently, and recommends a product the store has never carried. Another customer asks a restaurant’s assistant what pairs with the salmon and is told about a sauce the kitchen does not make. A third asks about returns and gets a policy the business does not have.

None of this is the AI being broken. A general AI model knows a great deal about the world and nothing about your business unless you tell it. Asked a question, it will produce a plausible answer whether or not that answer is true for you. The fix is not a smarter model. It is a different way of building the assistant around your own information.

The problem with a general assistant on your site

Put a general AI model on your website and it will draw on everything it learned in training. That makes it fluent and helpful sounding, and it makes it a risk. It can name products you do not sell, describe services you do not offer, quote hours you do not keep and promise things your staff would never promise.

Customers do not know which parts came from your business and which came from the model’s general knowledge. To them, it is your website talking.

Make your own data the only source

The approach we use is called grounding: the AI answers from your information, and only your information.

We built The Somm, an AI sommelier for Enology Wine Bar & Bistro in Saint James, this way. It pairs food and wine, but only from what is on the menu tonight. Here is the pattern, and it applies well beyond wine:

• Hand the model your current data as the source. The live menu, the product list, the policy text. That list is the only set of things it is allowed to talk about.
• Have it refer to items by number, not by name. Each item gets an ID. The model answers with IDs rather than writing out product names, which is where invention tends to creep in.
• Let your own code check every answer. Your software maps each ID back to a real record and drops anything that does not match. If the model offers a dish where a wine was asked for, that gets dropped too.

The customer only ever sees things that exist. The model can still write a warm, useful explanation, but the specific items in that explanation come from your records, checked by code you control. We wrote more about how The Somm came together in Creating The Somm.

Keep the source live

It is tempting to paste your catalog or menu into the AI’s instructions once and move on. Months later, that copy is out of date, and the assistant is confidently recommending last season’s items.

Connect the assistant to the same live data your business already maintains. When The Somm answers, it loads the menu the bar keeps current for its website, so when a wine sells out or a new plate goes on, its answers follow within minutes. Nobody has to remember to update the AI separately.

Write down what it must never say

Grounding stops the AI from inventing items. You also need explicit rules for questions it cannot answer safely, even about real items.

The Somm never says a dish is gluten free, vegan or safe for an allergy, because it does not know the kitchen’s ingredients and the server does. Every answer carries the same reminder to tell your server about allergies or dietary needs before ordering. It also never mentions prices, because prices are not the point of a pairing and the menu already shows them.

Every business has questions like these: medical, legal, safety, pricing, availability, anything where a wrong answer causes real harm. Write each one down as a refusal with a clear next step, usually “ask a person,” and include it in the assistant’s instructions.

Give answers a defined shape

Free-form text is hard to check. A defined structure is easy. Instead of letting the model write whatever it likes, tell it exactly what an answer contains: a headline, a short explanation, and specific fields for each recommendation.

The Somm’s pairings, for example, each carry a type of pairing, the flavor that links the food and wine, and a score. That consistency makes answers easier for customers to scan and much easier for your code to validate, because a missing or malformed field is obvious.

Keep the business in control

A few practical controls matter as much as the technology:

• An on and off switch the owner controls. The Somm can be switched off in the same admin where Enology’s team edits its menu. If something looks wrong, nobody waits on a developer.
• A way for customers to report a bad answer. Real use will surface cases nobody tested. Make it easy to tell you.
• Speed tuned for a waiting customer. Someone asking at a counter or a table will not wait long. A good answer in a few seconds beats a perfect one that takes a minute.

Where this works

The value of an assistant like this is not the AI’s general knowledge. It is how well the assistant uses your specific information. The same pattern fits a product finder that only recommends what is in stock, a help desk that answers from your own documentation, a policy assistant that quotes your actual terms, or an appointment helper that only offers real open times.

If you are weighing where AI fits in your business and how to build it safely on your own data, start with our AI readiness assessment.