Your AI Shopping Assistant's Wrong Answer Is Your Liability

Picture a shopper asking your site assistant whether she can return a swimsuit after her trip. The assistant says yes, within 60 days. Your policy says 30. She screenshots the answer, you decline the return, and a chargeback lands a week later. The real question is not whether the model "hallucinated." It is whose statement that was.

That question is about to get more expensive. Adobe's holiday forecast, released September 28, 2026, projects US online holiday spending of $275.1 billion (up 6.7% year over year) and a 130% year-over-year rise in traffic to US retail sites from AI chat and browser services. Adobe measures that traffic by shoppers clicking a link. More of your visitors will arrive having already asked one AI a question, and many will ask yours the next one: can I return it, does this code stack, will this fit my model.

The precedent: your chatbot is part of your website

In Moffatt v. Air Canada (British Columbia Civil Resolution Tribunal, February 14, 2024), the airline's website chatbot told a customer he could apply for a reduced bereavement fare after travel. The airline's actual policy did not allow that. Air Canada argued, among other things, that the chatbot was a separate entity responsible for its own statements. The tribunal rejected the argument, found the chatbot was part of the airline's website, and held that the airline had not taken reasonable care to ensure the chatbot was accurate. Correct information elsewhere on the site did not rescue the airline. The award was about CAD $650 in damages, plus interest and fees.

The dollar figure is small. The reasoning is the point. Two caveats, stated plainly: this is a Canadian tribunal decision, not binding on a US court, and US law on AI-generated statements is still unsettled. This is not legal advice, and counsel should review your terms. But the logic is easy to carry over: if the answer appeared on your site, in your voice, a shopper can say she reasonably relied on it.

Don't confuse this with Target's agent terms

On March 22, Target updated its terms for an upcoming Gemini-powered shopping agent so that transactions the agent makes after a customer authorizes it are "considered transactions authorized by you," and customers are responsible for reviewing what it does (as reported by TechSpot on April 6, 2026). That is a retailer allocating risk for a third-party agent acting on the shopper's behalf.

Your on-site assistant is the opposite case. It speaks for you. A line in your terms of use may help at the margins, but it will not turn a wrong return-policy answer into the shopper's problem.

Four kinds of answers an assistant can get wrong

Language models are fluent, and fluency hides errors: a wrong answer reads exactly like a right one. In ecommerce, the risky answers fall into four groups.

• Policy: return windows, exchange rules, shipping cutoffs, warranty terms.
• Price and promotion: whether a code stacks, whether an item is in the sale, the final total with shipping and tax.
• Product facts: compatibility, sizing, materials, ingredients, allergens.
• Order-specific facts: delivery status, address changes, cancellation eligibility.

Quick answer: who is responsible when an AI shopping assistant is wrong?

In the most-cited ruling, Moffatt v. Air Canada (2024), the business was held responsible for its chatbot's wrong answer, and correct information elsewhere on its site was no defense. US law is unsettled and this is not legal advice, but brands should treat assistant answers as their own published statements and engineer them accordingly.

Five controls that make assistant answers defensible

1. Answer from a versioned source of truth, and cite it. Policy answers should be retrieved from the current policy text, quote the relevant clause, and record which version was used. When the policy changes for the holidays, the assistant's answer should change the same hour, not after someone remembers to re-prompt it.

2. Take arithmetic and eligibility away from the model. Prices, promo stacking, shipping cost and stock should come from deterministic calls to your commerce platform, not from the model's reading of a page. The model explains the result. It does not compute it.

3. Build a refusal path. When retrieval finds nothing, or two sources conflict, the assistant should say it cannot confirm the answer and hand off to a person or a policy link. A graceful "I can't confirm that" costs one support ticket. A confident wrong answer costs a dispute.

4. Log the full chain. Store the question, the sources retrieved, the tool outputs, the answer shown and the policy version, so any disputed exchange can be reconstructed. Keep card data out of these logs; it keeps your PCI scope from growing.

5. Test before every release and every policy change. Maintain a fixed set of adversarial questions (edge-case returns, stacked codes, "your rep told me you'd price-match") and treat a regression as a release blocker. Our earlier piece on testing ecommerce AI search covers how to build that evaluation set.

How we approach it at MnT Future

At MnT Future, we design shopping assistants around one separation: language for the conversation, deterministic systems for any fact that carries legal or financial weight. MnT Commerce, our own reference build with semantic search, a shopping assistant and an ops agent, is where we apply that principle ourselves. We would rather ship an assistant that says "let me connect you with someone" on a hard question than one that answers everything and is wrong on the expensive questions. See our AI agents services.

A pre-holiday checklist before November 1

• Pull the last 30 days of assistant transcripts, flag every answer about returns, shipping, price or promotions, and compare each against current policy.
• Keep policy and promotion rules in one versioned place the assistant reads from, and retire stray PDF copies.
• Write down holiday exceptions (extended return windows, shipping cutoffs) before the assistant has to improvise them.
• Put a visible human handoff on every policy and order question.
• Have counsel review your terms and the assistant's disclosure language.
• Confirm your logs capture the full answer chain without storing card data.

The takeaway

An assistant's answer is a published statement. Treat it with the same care as the policy page it is quoting, and the holiday traffic Adobe is forecasting becomes an opportunity rather than an exposure.

If you want a second set of eyes before peak season, MnT Future offers a free agent-readiness audit and a free strategy session. We will review how your assistant answers policy, price and promotion questions, and show you where it is guessing. Book a free audit.

Sources: Adobe holiday forecast (Sept 28, 2026, via BusinessWire); Moffatt v. Air Canada, BC Civil Resolution Tribunal (Feb 14, 2024); TechSpot (Apr 6, 2026) on Target's agent terms.