How to stop an online store AI agent from making up product answers

A shopper asks whether a jacket is available in their size and whether it is waterproof. The store’s AI agent gives a confident answer, but it has borrowed a feature from a different jacket and relied on an old stock figure. The chat looks successful until the customer discovers that the item does not match the answer, or is not available after all.

Preventing that outcome takes more than telling an agent not to guess. A store needs to decide which records support each kind of claim, when information must be checked again, and what happens if a reliable answer is unavailable. The goal is not to answer every question automatically. It is to help customers get dependable information without hiding uncertainty.

Set boundaries around the questions

Start by reviewing product enquiries arriving through the website, chat and other customer-service channels. Group them by what they require: published specifications, comparisons between variants, current availability, shipping and returns policies, or advice about a shopper’s particular needs. These categories may appear together in one message, but they should not all be handled as if they were ordinary catalogue lookups.

An approved product record may support a statement about fabric, dimensions or what comes in the box, provided the answer matches the exact item and variant. A current policy record can support a general returns answer. Availability is different: it may depend on a live inventory lookup for the right size, colour, sales channel or location. An old product description cannot establish that an item is in stock now.

Choose what the first version will not answer. If reliable records do not cover compatibility, performance in specific conditions or future restock dates, the agent should not infer them from a similar product. It can ask a clarifying question, state what is confirmed, or pass the enquiry to a member of staff.

Give every claim an authorised source

Much of the work happens before the agent responds. Assign a trusted source to each answer type: approved catalogue data for specifications, an inventory system for stock, and the current policy record for returns. Decide what takes priority when records disagree and who is responsible for fixing the discrepancy. A fluent response is not a substitute for a clear rule about conflicting data.

Check how products and variants are identified. A search for “blue jacket” may match several sizes or an older version, while a copied description may no longer describe the item on sale. Tie lookups to product and variant identifiers where possible rather than relying on similar names alone. Remove outdated material from the information available to the agent, and set an approval process for catalogue changes.

Language models are useful for understanding varied wording, finding relevant passages and expressing an answer clearly. Deterministic checks are better for confirming a specific variant, obtaining a stock reading or preventing an availability claim when the required lookup fails. If records cannot move reliably between systems, improving an integration or a structured product form may matter more than refining the agent’s prose.

Make uncertainty visible and handoffs useful

Set response rules that separate known facts from missing information. The agent may quote dimensions for the matching variant, but it should not extend those dimensions to the whole range. If a current stock lookup is unavailable, it should say it cannot confirm availability at that moment. Even a valid stock reading is not a reservation or a guarantee of delivery.

Specify when the agent must involve a person: there is no authorised source, records conflict, the variant is unclear, an inventory request fails, or the question calls for individual judgement. A useful handoff carries the product and variant, the customer’s question, the records checked and the reason the answer is uncertain. That saves the shopper from starting over and gives staff a way to spot recurring data problems.

Ownership matters here. A catalogue owner can resolve unclear specifications; operations can investigate stock discrepancies; customer-service staff can continue the conversation and check what the shopper was told. Assign those responsibilities before an incorrect answer becomes a complaint.

Test the awkward cases, not just the easy ones

Build a test set from actual enquiry patterns, removing personal details where appropriate. Add similar product names, out-of-stock variants, missing specifications, contradictory descriptions and questions with multiple parts. Test what happens when a variant is removed or the inventory connection is temporarily unavailable. Success includes refusing to make an unsupported claim, not just producing a quick reply.

Review answers against the authorised records. Check that the agent identifies the correct item and variant, that availability statements come from a current lookup, and that handoffs preserve enough context for staff to act. Ask catalogue and service teams to examine both answers and refusals. An agent that forwards every routine question may be safe, but it will do little to reduce repetitive work.

Record why each failure occurred. Some problems call for a corrected product record; others need a rule change, an integration fix or clearer response wording. Prompt changes alone will not repair inaccurate stock data.

Roll out narrowly and learn from the results

Begin with a product category whose records are maintained well and a limited set of question types. Keep the route to a person obvious. Expand only after reviewing conversations, uncertain answers and mistakes. If current stock cannot be checked dependably, leave availability out of the initial scope rather than asking the agent to sound confident about it.

Track how often answers have valid supporting records, how often the wrong product or variant appears, whether staff receive useful handoff context, and what corrections they make. Look for clusters of unanswered questions too: they may reveal missing catalogue information rather than a conversation-design problem. Review findings regularly and make someone accountable for approving changes.

A dependable product agent does not need to know everything. It needs to answer what the store can verify and make the next step clear when it cannot. To explore a suitable starting point, ask Cibercoding to review one product-enquiry workflow with your team.