How AI Assistants Handle Size-Out and Discontinued SKU Scenarios Without Losing the Sale

What Happens When a Customer Asks for Something You No Longer Sell?

Most ecommerce stores lose the sale silently. The chatbot says "I'm sorry, that item isn't available" and the conversation ends. The customer leaves. No alternative is offered, no substitute is surfaced, and no revenue is captured.

This failure happens because most AI shopping assistants are designed to answer questions—not to manage inventory. They do not know what is actually in stock at the moment a conversation happens. The result is either a wrong recommendation (the AI confidently suggests a sold-out SKU) or a dead end (the AI admits it does not know what to suggest instead).

There is a better architecture. When the server—not the AI—holds responsibility for catalog decisions, size-out and discontinued SKU scenarios become recoverable moments instead of lost sales.

Defining the Problem: Size-Out vs. Discontinued SKUs

These are two distinct inventory states that require different handling:

• Size-out: The product exists in your catalog, but the specific variant a customer requested (size, color, configuration) is currently out of stock. Other variants may still be available.
• Discontinued SKU: The product has been permanently removed from the catalog. No variant will return. The customer's intent must be redirected to a different product entirely.

Both scenarios are common. A footwear store running a seasonal sale will regularly encounter size-outs. A beauty brand reformulating a product line will face discontinued SKUs. The question is whether your AI layer knows the difference—and whether it can act on that difference in real time.

Why AI-First Architectures Fail at Inventory Accuracy

Large language models are trained on static data. They do not have access to your Shopify catalog at inference time unless they are explicitly connected to a live data source. Even when a retrieval layer is added, most implementations cache product data—meaning inventory levels can be minutes or hours out of date.

This creates a specific class of error: the AI recommends a product with confident, persuasive copy, and the customer clicks through to find it sold out. Trust erodes. Cart abandonment rises. The AI has actually made the experience worse than a simple search would have.

The problem deepens for discontinued items. A model trained or fine-tuned on older product data may reference SKUs that no longer exist in any form. The customer is sent on a chase for something that cannot be purchased anywhere in your store.

How Server-Side Catalog Authority Solves the Problem

The architectural fix is a clean separation of responsibilities: the server selects the product, the AI writes the copy.

In this model, when a customer expresses a purchase intent, the request is first routed to a recommendation engine that queries the live catalog. The engine checks real-time stock levels, applies business rules (margin thresholds, preferred substitutes, promotional priorities), and returns a qualifying SKU. Only then does the AI layer receive a product to describe.

This means the AI is never in a position to recommend something unavailable. It does not know the difference between a size-out and a discontinued item—it does not need to. That logic lives in the server layer, where it belongs.

SmartBrain is built on this principle. When a shopper in a SmartBrain-powered conversation asks for a size that is not in stock, the engine does not pass that SKU to the language model. It selects the closest qualifying alternative—same style, nearest available size, within the customer's stated budget—and the AI describes that product. The conversation continues. The sale is recoverable.

Concrete Examples: What Good Handling Looks Like

Scenario 1: Size-Out on a Bestseller

A customer opens a chat and says: "Do you have the canvas tote in olive, size large?" The large in olive is sold out. A server-driven system checks the catalog and finds the same tote is available in large in sand and sage. It returns the sage option as the closest match by color proximity and stock depth. The AI responds: "The olive large is currently sold out, but the sage version in large is in stock and ships today—same design, very similar tone." The customer can proceed.

Scenario 2: Discontinued SKU with a Natural Successor

A returning customer asks for a specific serum by name—one that was reformulated and retired six months ago. A catalog-aware system flags the SKU as discontinued and looks for the configured successor product. The AI says: "That formula has been updated—the new version uses the same active ingredients with improved texture and is available now." No dead end. No broken link. No disappointed customer.

Scenario 3: No Viable Substitute Exists

Sometimes there is no good alternative. A server-driven system can be configured to surface a waitlist option, a bundle that partially addresses the need, or a transparent "we don't carry that" response alongside a category suggestion. The AI handles the language; the server decides the fallback path. This is honest commerce—and it protects brand trust better than a hallucinated recommendation.

Server-Driven vs. AI-Driven Recommendation: A Direct Comparison

• AI-driven (no live catalog): Recommends based on training data or cached index. Risk of suggesting out-of-stock or discontinued products. Copy is fluent; accuracy is unreliable.
• Server-driven (live catalog authority): Recommendation is always in-stock and within configured rules. AI writes copy for a pre-validated product. Accuracy is structural, not dependent on model knowledge.

For merchants running high-SKU catalogs or frequent promotional cycles, the server-driven model is not just preferable—it is operationally necessary. Inventory changes faster than any AI can be retrained or its cache refreshed.

How SmartBrain Implements This in Practice

SmartBrain sits between the customer conversation and the Shopify catalog. Every product surface in a chat session is resolved against live inventory before the AI generates a single word of description. Substitution rules, discontinued-SKU redirects, and fallback hierarchies are configured at the merchant level—not hardcoded into a model.

This means a store can define: "If size M is out, offer size S before offering a different colorway." Or: "If this SKU is discontinued, always route to the successor SKU first, then to the category page." These are business decisions, and SmartBrain keeps them in business logic—where merchants can update them without touching the AI layer.

FAQ

Can an AI chatbot know when a product is out of stock in real time?

Only if it is connected to a live inventory feed at the moment of the query. Most AI assistants are not. The reliable approach is to have the inventory check happen server-side before the AI generates a response.

What is the difference between a size-out and a discontinued SKU for an AI system?

A size-out means a variant is temporarily unavailable; substitution within the same product is usually possible. A discontinued SKU means the product is permanently gone; the system must redirect to a different product. These require different fallback rules, which should be managed at the catalog layer, not inferred by the AI.

What happens if there is no substitute for a discontinued product?

A well-configured system surfaces the most relevant alternative (by category, price, or attributes), offers a waitlist or notification option if available, or clearly communicates unavailability without generating a fake recommendation. Transparency preserves trust even when the sale cannot be completed.

How do substitution rules get configured in a server-driven commerce system?

Merchants define substitution logic at the product or category level—successor SKUs, fallback categories, margin thresholds, and priority rules. This configuration lives outside the AI model and can be updated without retraining or re-prompting.

Does this approach work for stores with thousands of SKUs?

Yes. Server-side catalog queries scale with the catalog. The AI layer only ever receives a pre-validated product recommendation, so the complexity of a large catalog does not increase the risk of inaccurate suggestions—it is filtered before the AI is involved.