Compliance Gateway: Why Your Bot Won't Invent

TL;DR: The worst failure of an AI salesperson is a confident lie: a product you don't stock, a price you never set. aiSTAFF sets the relevance gate on each search to a threshold of cosines close to 0.64. If nothing clears the bar, the bot says "we don't carry that" instead of inventing the item.

Costly Failure Mode

The general purpose language model wants to be useful, which is the problem. Ask him for "waterproof bluetooth under 80 GEL" and without grounding, he will gladly describe one: the name, the specifications, the price, everything is made up. The customer believes it, asks to buy, and your team is left to explain that the product doesn't exist. You have now turned the sale into an apology and damaged trust. For a store, a hallucinating product is worse than a bot that says "I don't know," because the lie goes farther before it breaks. Avoiding this is the whole point of grounding the bot in your actual catalog, which our AI Sales Bot does by design.

Grounded, not guessing

aiSTAFF does not answer product questions based on general knowledge of the model. It searches your catalog, built-in private vector store, and responds with only what it finds. This is a search-powered generation: the bot retrieves real products first, then talks about them and nothing else. The search itself combines semantic and keyword searches, combined so that the best match is increased. If you weigh this against retraining your data model, the trade-offs are RAG vs fine-tuning.

However, grounding alone is not enough. A search always returns something, even for a query that has no real match. A top result for a "gaming mouse" in a furniture store might be a mouse-shaped pillow with a low score. Put it back and the bot looks broken. The compliance gate is what stops this.

How Compliance Gate Works

Each search produces matches with a similarity score, the cosine value between the query and each product, where the higher value is closer. The gate sets the floor near 0.64. The logic is short:

If the best match scores exceeds the threshold, the product is indeed a match and the bot will show this.
If the best match scores below the threshold, there is nothing in the catalog close enough and the bot returns an honest blank: "We don't carry it."

Threshold adjusted between real match and weak match. A furniture store search for "gaming mouse" comes up low with all the chairs and desks, gets under the gate and comes back as an obvious no-no. A search for "office chair" in the same store scores high and returns cards. The gate is the difference between a bot that calls the limit and a bot that bluffs.

An empty response is a function

An honest "we don't carry it" sounds like a loss, but it's the opposite. It maintains customer confidence, tells you there's a gap in your catalog, and turns the buyer toward something you make instead of stock. The bot can follow up the blank point with a helpful nudge: "We don't have any gaming mice, but here are the office accessories we do have." This turns the deadlock into a redirect. The sales engine that makes it all boil is in the hub, AI Chatbot That Sells Your Catalog.

The same sincerity can be seen in the shopping cart and offers. A multi-item cart is never filled with a fictional add-on covered in Chat Chat and related item sales are pulled from actual inventory, explained in Context Aware Best Product Pick. When the match opens the gate, the customer sees the full card with price and inventory covered in product cards in the chat that converts.

Gateway stored in any language

Since a Georgian or Russian query is translated into a search, the door opens on the translated query as well as on the English one. Searching in Georgian for an item you don't carry returns the same honest, non-fictional product, so the guardrail isn't weakened when the customer changes language, how a Georgian customer buys your English catalog. Word forms and synonyms are processed before the gate, so the actual product is not incorrectly rejected because of grammar.

Tuning your store threshold

The gate is a dial, not a fixed wall. Set it too low and weak matches fall apart, set it too high and real products are rejected. The default near 0.64 is a reasonable starting point, and it's tailored by business against the messy requests your customers send: misspellings, vague descriptions, mixed languages. You look at deadlock and mismatch ratio and adjust until honest emptying occurs only when the catalog is missing an item. A healthy store keeps that gate tight enough to never lie and loose enough to never hide the products it sells. A more extensive build is in the Ecommerce Chatbot Guide.

Working example

The electronics store receives two messages per minute. First: "Do you have a 65-inch OLED TV?" The bot searches, finds two matching models above the threshold, and returns cards with price and inventory. Second, from another user: "Do you sell washing machines?" The store carries only TVs and audio. The search scores all the products low, the best match goes under the gate, and the bot replies that it doesn't have washing machines, then points to the audio gear it does have. The first customer got the real product, the second got the truth and a redirect and no fabrication. This is the gateway to its business, and that is why the grounded bot is safe in front of buyers.

FAQ

What is Compliance Door?

This is the similarity threshold, close to cosine 0.64, for each product search. If the best match is above it, the bot will show the product; If nothing clears, the bot returns an honest blank instead of inventing the item.

Does a chatbot ever calculate a product or price?

No. The bot only responds from your built-in catalog and uses a compliance gateway, so non-catalog inquiries are returned as "We don't carry this" and not as a fabricated product, spec or price.

Is a blank answer bad for sales?

It protects trust and points out a real flaw in your catalog. The bot can follow the blank and redirect to the items you stock, turning a dead end into another chance to sell something real.

Does the gate work on Georgian and Russian questions?

Yes. Queries are translated for search and the door passes on the translated query, so the same honest doesn't apply to whatever language the user is using.