When AI gives the wrong price or stock level, start by checking the source used in the answer. The problem may be an old table, product variant identification, or how an integration delivers the data. Fixing it depends on finding where the information stopped matching what the store can sell.
A customer asks about white sneakers in size 37. The answer confirms availability, but the stock checked was for size 38. In this hypothetical example, AI may have received a correct number and applied it to the wrong item.
In the ChatFunnel knowledge base, you can review content and versions, define what is active, and choose the materials linked to each agent. This helps correct documentary sources. For price and availability that change during the day, also check the scope of the connection supplying those data.
Save the question and locate the information used
Before changing AI instructions, collect the customer message, the answer received, and the correct data at the time of review. Record the complete product, including size, color, or another relevant variant.
If the agent's lookup was logged, check which item was queried and what result returned. If not, ask the person who configured the solution to help locate the source. Without this step, it is easy to change the wording while keeping the same incorrect information.
Also consider timing. A price may have changed after the conversation. Comparing the old answer only with the current page is not enough to conclude an error occurred. Look for the change history or record that limitation in the analysis.
While investigating, restrict answers that depend on the questionable source. The team can handle those lookups until the fix is tested. Guidance should state what can still be answered and what needs checking, avoiding interruption of service using reliable information.
Separate stable information from changing data
Composition, washing instructions, and product measurements generally change less often than promotional prices or available quantities. These differences matter when choosing a source.
A reviewed product sheet may explain a shirt's fabric. A table sent last week may be unsuitable for saying how many units can be purchased today.
| Information | Source to check | Error worth investigating |
|---|---|---|
| Measurements and composition | Product sheet and material version | Document for another model or outdated content |
| Variant price | Corresponding listing and offer terms | Another variant's price or incorrect field |
| Discount | Current rule and application requirements | Expired coupon or benefit applied without its condition |
| Availability | Stock data appropriate for sale | Total quantity confused with sellable quantity |
| Timing | Estimate applicable to the address and shipment | General timeframe treated as a promise for every order |
Define a primary source for each piece of information. When a document and a lookup return different values, the agent needs a clear rule for handling the conflict. Adding more files to the knowledge base without reviewing earlier versions can increase ambiguity.
Check the product and variant before the value
A similar name does not necessarily identify the same item. A T-shirt may have different prices by edition, and two visually similar garments may have different measurements.
When testing, use questions that omit part of the identification. "Do you have the black one?" requires knowing which product the conversation concerns. If context is unclear, customer service should ask for the missing information.
When checking the integration, compare the queried identifier with the variant displayed in the store. See whether the answer uses the requested size and retains that choice after the customer changes color.
The meaning of the price field also matters. Shopify, for example, distinguishes the selling price from the compare-at price used to present a promotion. Its documentation also handles these values by variant. Reading the wrong field can produce an inconsistent answer even when the lookup worked. Shopify sale pricing documentation.
If a benefit depends on payment method, give that condition with the value. A Pix-only price should not appear to apply to every payment method.
Find out what the stock number represents
Having units recorded at a location does not mean all are available for a new purchase. Some may be tied to orders, reserved, or unavailable for sale.
Shopify distinguishes available inventory from total on-hand inventory. The total includes available, committed, and unavailable units. This definition is platform-specific, but illustrates why the field name matters. Inventory states documentation.
Imagine a warehouse records 5 units of an item, but 4 are already committed to orders. If the integration sends only the total, the answer "we have 5 available" may be wrong. These numbers are illustrative; check the actual states in your platform.
Also verify which location fulfills that sale. Stock at another location may not be available for the channel or requested delivery. The integration team should explain which quantity reaches the agent and what it represents.
Even a correct lookup is a snapshot of that moment. If the system does not reserve the item during the conversation, avoid saying the unit is guaranteed until the customer completes the purchase.
Correct the source and define the response to missing data
If the source is an old document, update the content and remove the prior version from agent use. In ChatFunnel, knowledge base controls allow review of versions and active materials. After the change, repeat the question that revealed the problem.
If the source is a connection, review the received fields and refresh interval with the person who configured it. Instruction text alone does not fix a swapped identifier or a lookup that stopped responding.
Also define how customer service should act when information is unavailable. Useful guidance might be:
To confirm price, first identify the product and variant. Use only the source defined by the store. If the lookup fails or values conflict, say the information needs checking and forward the question to the team. Do not use a price remembered from another conversation.
This example is writing guidance that must be adapted to the tool's capabilities. It does not replace lookup configuration or independently guarantee the agent follows the rule in every case.
In ChatFunnel, you can combine reviewed materials with handoff to human customer service. Define who receives the question and what data should accompany the conversation so the customer does not start over. To define those cases, see when AI should hand customer service over to a person.
Repeat the test with question variants
Answering the original question correctly is the first sign of a fix. Then test the same topic with different wording and incomplete information. The guide to testing an AI agent for an online store offers other useful variations.
For the sneaker example, request size 37, change to 38, and return to the first. Include an unavailable color. Also check behavior when the source does not respond: the agent should follow the defined failure path without presenting an old value as current.
Record the expected source and each test result. If the error reappears only for one variant, investigate that listing before changing instructions that work for other products.
After resuming automatic replies, review a sample of conversations on the topic. Look for continuing discrepancies and whether handoffs reach the team with enough context. A decrease in complaints alone does not prove every data point is now correct.
This process makes each answer verifiable: which product was identified, where the data came from, and which path was followed when information was missing. In ChatFunnel, a reviewed knowledge base and human service help organize these checks and keep materials used by agents under control.
