How Anchor uses business records to answer questions

A language model on its own writes fluent text from patterns it learned, which is how an assistant ends up quoting a price you never set. Anchor answers customer questions from your catalog instead, so the quality of the answer depends on the quality of the record.

• Products: name, description, price, variants and attributes.
• Stock levels for each product or variant.
• Compatibility rules you record, for questions such as "does this charger fit that phone?"
• For an owner’s own questions: the business’s orders, quotes and customers.
• Only your business’s records. Every request is checked against the business it belongs to.

1. 1Step 1: A question arrives from your website widget or a connected messaging channel.
2. 2Step 2: Anchor searches your catalog, matching both keywords and meaning, to find the relevant products.
3. 3Step 3: It writes the reply using only what it found and shows the records it used.
4. 4Step 4: Where a compatibility rule applies, the rule decides the answer rather than the model.
5. 5Step 5: The conversation is saved, so you can read it in Conversations.

If the search finds nothing relevant, Anchor does not fill the gap with a guess. It tells the customer it cannot confirm that, and the conversation is handed off to your team: it is marked as needing a reply, it appears in Conversations and in the Today attention list, and a person follows up when available.

Questions that are not about your catalog, such as delivery areas, opening hours or special requests, will usually be handed off, because the customer assistant answers from your products.

A made-up shop that sells chargers, to show both paths.

1. 1Step 1: "How much is the 65 W USB-C charger?" The product record has a price, so Anchor answers with it and shows the product.
2. 2Step 2: "Will it charge my laptop?" The shop recorded a compatibility rule between the charger and two laptop models. The customer named one of them, so the rule answers yes.
3. 3Step 3: "Can I pick it up this evening?" Pickup times are not in the catalog, so Anchor says it cannot confirm that and hands the conversation to the team.

• Write descriptions a customer would understand, without internal codes.
• Use attributes for facts people compare: size, colour, capacity, material.
• Record compatibility rules for accessories and parts.
• Update stock after deliveries and sales, or record in-person sales in Anchor.
• Read handed-off conversations each week and add the missing product details they reveal.

• A reply is only as accurate as the record behind it; out-of-date stock or prices give out-of-date answers.
• The assistant does not browse the web or other systems for your products.
• It can still misread a question. Spot-check answers regularly.