To use AI in a retail business without multiplying unreliable answers, start with an accurate catalogue and approved sources. An assistant needs to explain where its answer comes from and when it applies. If evidence is missing, outdated or contradictory, it should explain the limitation and offer a real way to reach a person. An old document cannot confirm current stock or policy.
This method suits a small shop preparing bilingual listings, supporting its staff or piloting customer replies. It keeps three related areas in view: catalogue quality, dependable service and inventory rules. You can start internally, using fictional records and a person who reviews every answer.
Why catalogues are becoming connected to assistants
On January 11, 2026, Shopify announced the Universal Commerce Protocol, co-developed with Google for AI commerce. Its June 17 developer announcement described wider access to that infrastructure and Catalog API. These developments show how product catalogues and purchasing workflows are becoming connected to assistants.
Actual availability depends on the platform, country and merchant account. A launch announcement does not mean every Canadian store can use every feature. Before investing, check your account’s eligibility, supported channels and the way an order is processed. Treat vendor announcements as evidence of a product direction, rather than a promise of results for your business.
Your catalogue becomes an information source
A description that says “high quality and ideal for everyone” does not answer a specific buying question. Customers need reliable dimensions, materials, variants, compatibility information, care instructions and limitations. The details must match the product being sold and remain consistent across the website, point-of-sale system and other channels.
One possible AI use is preparing descriptions from an approved supplier specification. A responsible person checks that the draft adds no unsupported certification, assumed compatibility or invented delivery date. The tool can also identify missing attributes, but it should not fill a technical field with an educated guess.
For a bilingual store, pay particular attention to sizes, units and variant names. Translating a size label or measurement without context can produce an ambiguous listing. The same product identifier should continue to identify the correct variant in both languages. Keep the source specification available so reviewers can settle disagreements quickly.
A customer assistant needs to cite, abstain and hand off
Watch for specific symptoms: an old policy presented as current, availability inferred from an outdated catalogue, different answers across languages or an “I’m transferring you” message that reaches nobody. These situations warrant checking sources and logs. By themselves, they do not establish that your platform is broken.
A citation does not make an answer correct. The NIST Generative AI Profile, published in July 2024, describes incorrect responses supported by misleading reasoning or citations. Reviewers need to check both the cited passage and whether it applies to the customer’s situation.
Prepare the sources and access
Start with a narrow request, such as in-store pickup arrangements. You need an approved policy, an owner responsible for keeping it current, reviewed French and English versions and a service channel that someone actually monitors. Provide read access to the necessary information and a test environment without real orders or contact details. Someone who knows the policies can prepare the review sheet; configuring and checking integrations, permissions and technical records requires the relevant technical skills.
For each source, record its identifier, version, applicable location and channel, effective dates, language and owner. A document’s upload date is not its effective date. A question about an older order may require the policy that applies to that order’s date and terms. Check those conditions instead of automatically choosing the newest document.
Current prices and stock require an appropriate operational source. Agree with the owner on acceptable freshness and the meaning of the returned value. Displayed inventory is not a reservation. If a connection fails or a result is too old, the assistant should say it cannot confirm the information rather than silently reuse the last known value.
Choose the outcome before writing the answer
- Check the scope. Does the question concern this location, product, period and an allowed topic? Ask for clarification when needed; personal information requires the intended access check.
- Find applicable evidence. The passage must answer the question and be approved for that context. Finding more documents does not resolve a conflict between two valid sources.
- Choose the next step. Answer with a precise citation when the evidence is sufficient, ask a question when a detail is missing, or explain the limitation and offer the appropriate human channel. Do not invent an exception, deadline or refund to end the conversation.
- Check commitments separately. Explaining a policy does not grant permission to act on an order. Commercial promises and record changes must follow their own access and approval rules.
A useful citation identifies the title, version or relevant section and the passage supporting the claim. A customer-facing link must be accessible to that customer; internal evidence belongs in the adviser’s authorized view. A working URL proves neither that the passage exists nor that it supports the answer.
Retrieved text is data, not an authorized instruction to change the assistant’s behaviour. A product sheet can contain malicious instructions. OWASP describes this indirect prompt injection risk and recommends measures including limited privileges and human approval for sensitive actions. A sentence in the system prompt does not replace application-level access and action controls.
A fictional pickup-policy example
The imaginary shop in this example keeps two versions: PICKUP-v1, seven calendar days after the ready-for-pickup notice, effective October 1–31, 2026; and PICKUP-v2, five calendar days after that notice, effective November 1 onward. Reviewed French equivalents are available. These are test records, not a real retailer’s terms or a returns policy.
For “What is the pickup window for a new order today?”, asked on November 5, the expected answer is: “According to PICKUP-v2, you have five calendar days after the ready-for-pickup notice.” It points to the supporting section. If only v1 is available, the expected answer becomes: “I can’t confirm the current policy from the available information.” It then offers a real way to contact the team, without repeating seven days as though that rule were still applicable.
| Test situation | Expected result |
|---|---|
| v1 and v2 available; new order on November 5 | Cite v2 and answer five days in both English and French. |
| Only the old policy, or no answer in the sources | Abstain from a substantive answer and explain what is missing. |
| Two conflicting current policies | Have the conflict reviewed before answering. |
| Dated stock information or an unavailable price connection | Do not confirm current availability or price. |
| Private order without verified access | Do not disclose its contents; direct the customer to the authenticated process. |
| Requested exception or unavailable service team | Do not execute the exception or claim a successful handoff when none has occurred. |
On November 5, 2026, we ran 12 fictional scenarios using CPython 3.12.14. All produced the expected result against a deterministic source-selection and routing rule. The records were structured and pre-labelled, including permissions and flagged documents. This does not establish AI reliability: no generative model, search engine, injection detector, live connector or human handoff was tested. Reworded questions and free-text answers still need evaluation with the chosen tool.
Confirm that the handoff actually works
A message suggesting a phone call is not a conversation transfer. Microsoft’s handoff documentation distinguishes simple escalation from a connection to a service team. Verify that the case arrives in the team’s system, has an identifier and status, and can receive a reply. If delivery fails, tell the customer that the request was not transmitted and provide a valid alternative channel. Do not promise a callback or response time unless the team has arranged it.
The adviser should receive the question, language, sources checked, escalation reason and only the necessary information they are allowed to access. Avoid automatically copying an entire conversation containing personal information. Use fictional identifiers and redacted observations in the test log.
Validate the pilot before making it public
Microsoft’s generative answers documentation notes that answers can be incorrect and need evaluation before publication. Run ordinary requests and exceptions through your actual tool in both languages, using reworded questions and repeated trials. Repeat the tests after changes to policies, sources, configuration or the model.
Assess answer accuracy, citation relevance, appropriate abstention and confirmed receipt by an adviser separately. A strong average must not hide an order disclosure or an unauthorized promise: in this pilot method, either blocks public release until corrected. That is a preparation recommendation, not product certification. Afterwards, measure time to a useful resolution and reopened cases; closing a conversation does not prove that the issue was resolved.
Inventory decisions need records and explicit rules
Predictive tools can be evaluated for replenishment planning when usable history is available. That history should distinguish regular sales, promotions, returns and periods when an item was unavailable. Otherwise, low sales during a stockout can be mistaken for low demand. Cleaning that distinction may be more valuable than choosing a more complex model.
A first project should recommend an action rather than automatically send a supplier order. The person responsible checks lead times, minimum quantities and available cash against the recommendation. Some products work well with simple rules; others may justify a more sophisticated method. Complexity alone is not evidence that a tool will improve decisions.
Start with a stable category and compare recommendations with the current process. Track stockouts, slow-moving inventory and corrections made by the team. For a new product without sales history, make the assumptions behind a recommendation visible rather than presenting it as a dependable forecast.
Search visibility still depends on the fundamentals
Google states that established SEO practices remain relevant to its AI search features. It does not require a special AI file or dedicated markup. A page must be indexed and eligible for a search snippet to qualify, and inclusion is not guaranteed.
For a retailer, the practical work is to make products reachable through links, explain differences between variants, maintain availability information and align structured data with the visible page. A useful category guide can answer a buying question. A collection of repetitive descriptions does little to help someone make a decision.
Your website also remains a place where customers verify the business, its policies and its contact details. If that foundation is incomplete, a chatbot will not fix it: start with a website suited to your business and market before adding an assistant or another integration. Accurate information benefits visitors regardless of how they find the store.
A four-week starting plan
- Week 1: examine the records. Select a small category and check product identifiers, variants, descriptions, inventory and associated policies.
- Week 2: choose one use. Preparing bilingual listings or assisting with common customer questions is a clearer scope than a general AI transformation.
- Week 3: test exceptions. Include a stockout, a missing variant, an out-of-policy request and a question whose answer is not in the available records.
- Week 4: make a decision. Compare total time, errors and service quality. Keep, adjust or stop the workflow according to the results.
This is an example planning sequence, not a universal delivery promise. Catalogue size, existing systems and data quality can change the timeline. The important point is to schedule a decision based on real cases before expanding the scope. Agree in advance on what an unacceptable error looks like.
What belongs in the real operating cost?
Include data preparation, system connections, subscriptions, service usage, human review and maintenance. Add error handling and the time needed to train users. A feature that is inexpensive to activate can still require substantial work to keep reliable. Someone must own the source records and know when policies change.
Define who can access customer information and which actions each tool may take. Preparing a response does not automatically require permission to cancel an order. Suggesting replenishment does not necessarily require permission to approve spending. Keeping those responsibilities explicit makes the process easier to review and operate.
For each proposed integration, ask what happens if a connection fails halfway through. The team should be able to see whether an action was completed, still awaits review or needs attention. A clear status and recovery process matters more than an impressive demonstration that only handles the normal case.
A checklist before expanding the pilot
- Every source has an owner, version, scope and effective dates.
- Identifiers, variants and units remain consistent across the catalogue, inventory and languages.
- Citations support the answer; missing evidence and conflicts have defined outcomes.
- Prices, availability and orders use the appropriate sources and permissions.
- Exceptions reach an identified team; unavailability is not presented as a successful transfer.
- English and French cases, blocking errors and recovery criteria are recorded before public release.
A limited category is enough to start checking listings and answers. For inventory forecasts, the quality of the history matters more than catalogue size. No assistant or markup guarantees a recommendation or sale; measure observable visits and enquiries without inventing their attribution.
If you need to connect your store, inventory and customer service, explore our custom applications and integrations. For another industry example, read our guide to AI in Canadian construction businesses.