When you connect a customer-facing AI agent to your store, you must supply a bounded, defensible answer set. If the agent lacks reliable policy and compatibility evidence, incorrect answers are a risk to test rather than a behavior to assume away. Catalog teams need a structured register of product truth to keep these tools accurate and accountable.
Two major platforms have described customer-facing agents, with different products and rollout stages. Meta describes customer-facing Business Agents that answer business questions, recommend catalog products, and allow human handoff, as detailed in their Meta Business Agent introduction. Separately, Google has invited eligible US retailers into a beta for a Business Agent for YouTube ads, according to their Google holiday commerce updates. These are external tools meant to interact directly with your buyers. They are distinct from internal marketing assistants that help you write copy or analyze data. Because they speak to your customers, they require strict boundaries.
Define the boundaries of product truth
Your catalog team already manages price and availability to prevent disapprovals, a process covered in our guide on price and availability mismatches: why catalog-scale feeds get disapproved. Shopping agents require you to expand that discipline to material, compatibility, care, warranty, delivery, and exclusions.
An agent may infer an answer when explicit evidence is missing; its behavior needs to be tested in the actual implementation. This is where risk enters the system. When you manage thousands of SKUs, product details vary wildly even within the same category. A blanket statement about your brand's quality will not help a buyer who needs specific technical specifications.
Consider a hypothetical example. A customer asks an agent if a specific SKU of hiking boots is waterproof. The agent scans the brand's general description, sees the word waterproof applied to a different premium line, and tells the customer yes. The specific SKU in question is only water-resistant. The customer buys the boots, ruins them in a stream, and demands a refund based on the agent's promise. You now face a return and a frustrated buyer.
To prevent this, your team must define exactly what the agent is allowed to know and say about every product category. If a product has specific care instructions, those must be documented. If a warranty excludes certain types of damage, the agent must have access to those exact exclusions.
Build a structured answer register
Use an answer register as a controlled reference and evaluation set for the team. It records approved facts and the responses you expect to see. It is not, by itself, a technical control that forces an agent to comply. Check how the chosen product accepts knowledge, updates and escalation instructions.
Every entry in your answer register needs three components: a source, an owner, and an expiry date.
The source is the documented proof of the policy or product fact. This might be a link to a specific manufacturer specification sheet, a sizing chart, or your official shipping policy page. The agent must draw its response directly from this source material, rather than summarizing marketing copy.
The owner is the person on your team responsible for keeping that answer accurate. If the warehouse changes the cutoff time for next-day shipping, the owner must update the register. When you operate at catalog scale, policies change frequently. Without a designated owner, your agent will eventually distribute outdated information.
The expiry date is critical for time-bound information. Take a hypothetical scenario involving holiday shipping. You program the agent to guarantee December 24th delivery for orders placed before December 18th. Retire that answer at the actual order cutoff, including the time zone, destination and shipping-service conditions. A date without those details can still create a false promise. If it does not expire, the agent might continue promising pre-Christmas delivery to a buyer asking on December 22nd.
Map compatibility and exclusions at the SKU level
Stores with large catalogs often sell primary products alongside specific accessories, replacement parts, or compatible add-ons. Buyers frequently use chat tools to verify if item A works with item B. Your answer register must handle these compatibility questions with absolute precision.
You cannot rely on the agent to read product titles and guess compatibility. You must provide a structured mapping that links specific SKUs to their approved accessories.
Imagine a hypothetical situation where a buyer asks if a specific replacement battery fits an older drill model. If your register maps compatibility at the brand level rather than the SKU level, the agent might incorrectly confirm the fit. The buyer receives a useless battery, and your customer service team has to process a return that never should have happened.
The same logic applies to exclusions. If a promotional discount applies to the entire store except for three specific high-margin brands, record those exclusions in the approved reference and verify how the implementation applies them. The agent must check the exclusion list before confirming that a discount code is valid for a specific cart.
Program uncertainty and the human handoff
No answer register will cover every possible customer question. Buyers will ask highly specific, unusual, or confusing questions. Your agent must be programmed to handle uncertainty safely.
The default behavior for an unsupported question must be a polite refusal to guess. You cannot allow the agent to hallucinate a policy just to keep the conversation moving. A safe fallback script protects your margin and your brand reputation.
Define an escalation outcome for unsupported questions and test the product's available handoff mechanism. If it cannot route the conversation as required, provide a clear alternative contact path and restrict the pilot accordingly. Meta specifically notes that their Business Agents allow for human handoff. Your customer service team needs a clear protocol for picking up these escalated threads.
For a hypothetical example, a buyer asks if they can combine a first-time buyer discount code with a seasonal clearance offer, and then apply a store credit on top. If your answer register only covers standard single-code policies, the agent should state the standard policy and immediately route the complex query to a human representative. Do not assume that unsupported questions will be handled automatically or that the agent will figure it out. An answer outside the approved evidence should not be treated as verified simply because it sounds confident. Your operational plan must account for the gap between what the machine knows and what the buyer needs.
What good looks like
- Document exact warranty lengths and exclusions per category rather than relying on a sitewide average.
- Define explicit material care instructions for high-return SKUs to prevent post-purchase damage claims.
- Set hard expiry dates for promotional or seasonal shipping answers so agents stop making impossible promises.
- Map compatibility rules to specific SKUs so buyers know exactly which accessories fit their primary purchase.
- Create a strict fallback script for unsupported questions that prevents the agent from guessing.
- Test the human handoff routing with your customer service team before peak traffic hits.
- Assign a specific team member to own and update the answer register weekly.