A practical guide to grounded AI for your business

Grounded AI means the assistant looks facts up in your own records before it answers, instead of relying on what a model happens to remember. This guide walks through the three parts that make that work in Anchor: the records you keep, the answers built from them, and the points where a person reviews what the AI suggests.

It describes how Anchor works today. It is not a promise of perfect answers: grounding reduces wrong answers, and people still check the important ones.

• Add every product you sell, with a clear name and the price customers actually pay.
• Record variants (size, colour, pack size) separately, so each one has its own stock.
• Put the facts customers ask about into attributes as well as the description: dimensions, materials, ingredients, compatibility.
• Add many products at once from a CSV or Excel (.xlsx) file. Anchor shows a preview before anything is saved.
• Keep stock and prices current. An answer is only as fresh as the record behind it.

When a customer asks a question on your website widget or a connected channel, Anchor searches your catalog for the relevant products and writes the reply from what it found. The reply shows which records it came from, so you can check it.

Owners can also ask plain-language questions about sales and stock. Those answers are computed from your orders and stock records and show the date range they used.

If your catalog does not contain what a question needs, Anchor is designed to say it cannot confirm it rather than guess. In customer chat the conversation is handed off: it is marked for your team, it shows up in Conversations and in the Today attention list, and a person replies when they can. AI can still make mistakes, so check important details before relying on them.

Anchor can draft suggestions from your data: a follow-up message for a quote with no reply, a restock suggestion, or an alert about unusual sales. They wait in an approval queue, and only an owner or admin can approve them. A follow-up message is sent only after approval, and if you edit it, your edited version is what is sent. Separately, a quote becomes an order only when the customer accepts it.

The shop, product and numbers below are made up to show the flow.

1. 1Step 1: A shopper asks on the website widget: "Do you have the walnut desk lamp in stock?"
2. 2Step 2: Anchor finds the product record (Walnut desk lamp, $48, 12 in stock) and replies with the price and availability, showing the product it used.
3. 3Step 3: The shopper asks whether it can be delivered to their town. Delivery areas are not in the catalog, so Anchor says it cannot confirm that and hands the conversation to the team.
4. 4Step 4: The owner sees it in Conversations, replies, and because the same question keeps coming up, adds the delivery note to the product description.

• Every product has a name, a price and a stock level.
• Variants are separate products or variants, each with its own stock.
• The facts customers ask about are in attributes or descriptions.
• You have asked the assistant ten real customer questions and checked each answer against your records.
• Someone reads handed-off conversations every working day.
• An owner or admin reviews the approval queue.
• Your team knows what the assistant does not do (see the limits below).

• Anchor answers only from the records you add. It knows nothing about products you have not entered.
• It does not process payments, book appointments or give medical, legal or financial advice.
• Answers can still be wrong. Check important details.
• Which suggestions you receive depends on your plan and on how much order and stock data you have.