How to train an AI on your product catalogue
A practical guide to importing product data, writing clear policies, reviewing wrong answers and fixing the five failure modes of catalogue-trained support.
How to · 11 July 2026 · 10 min read · By The OctaDezx team, Builders of the OctaDezx AI customer care platform
Updated
Key takeaways
- Accurate source information and a review process matter more than a confident-sounding answer.
- Import structured product data rather than describing your range in prose. Summaries produce answers that generate returns.
- Write policies as decisions with conditions attached, because ambiguity becomes invention.
- Your inbox is the best training material: the same handful of questions repeated is empirical rather than imagined.
- Review an initial sample of real conversations. Correct the source and approve any proposed lesson before relying on it.
An AI support agent needs accurate information it can use. A thin knowledge base leaves room for unsupported answers, even when they sound confident. Start with product-specific facts, clear policies and a review process before relying on the model to talk to customers.
Here is the order of work that actually gets an agent to a state where you would let it talk to customers.
Step one: import the catalogue, do not describe it
Start with structured product data: product ID, name, description, price, availability and any size or variant details customers need. Import through the supported store connection or supply verified records. Check what the import actually contains; a short description is not a complete variant catalogue.
The temptation is to summarise your range in a paragraph. Resist it. A summary produces answers like most of our jackets are water resistant, which is exactly the kind of sentence that generates a return.
| Source record | Shopper question | Allowed answer |
|---|---|---|
| Example SKU JKT-NV-M: navy jacket, medium, $80, available; water resistant; not waterproof | Is the medium navy jacket waterproof? | The medium navy jacket is $80 and listed as available. It is water resistant, not waterproof. |
| Example SKU JKT-NV-L: no current availability recorded | Can you guarantee large is in stock? | I cannot confirm the large variant from the current listing. I can ask the team to check. |
These records and answers are illustrative teaching examples, not screenshots or results from a customer account.
Step two: write policies as decisions, not vibes
Policies are where most knowledge bases fall apart. Businesses write what they wish were true, in language soft enough to leave room for judgement. An AI has no judgement, so ambiguity becomes invention.
Write each policy as a decision with conditions attached.
- Weak: we are flexible about returns. Strong: unworn items can be returned within 30 days of delivery with the receipt, refunded to the original payment method within 5 working days.
- Weak: delivery is fast. Strong: orders placed before 4pm ship the same working day. Domestic delivery is 2 to 3 working days, international is 7 to 14.
- Weak: we can sometimes do discounts. Strong: the only published discount is the 10% first order code. Anything else is escalated to a human.
Step three: feed it the questions you already answer
Open your inbox and read the last two hundred messages. You will find the same handful of questions repeated with different wording. That list is your highest value training material, because it is empirical rather than imagined.
For each recurring question, add the exact answer you want given, in the voice you want it given in. Upload the files you already have: FAQ documents, spreadsheets of product specs, PDFs of care instructions, a text export of your policies. Anything already written down is faster to import than to rewrite.
If a question comes up ten times a week and nobody has written down the answer, that is not an AI problem. That is a decision your business has not made yet.
Step four: set the tone deliberately
Tone is not decoration. It determines whether customers trust the answer. Decide whether you are warm and chatty or brisk and factual, whether you use the customer's first name, how you handle bad news, and whether you apologise or simply fix.
Give the agent two or three example exchanges written the way you would write them. Examples move tone far more reliably than adjectives do.
Step five: read the first hundred conversations
This is the step everyone skips and the one that matters most. Go through the first hundred real conversations line by line. You are looking for three things: answers that were wrong, answers that were right but sounded wrong, and questions that should have been escalated and were not.
Separate a missing fact from a weak policy and a tone problem. Update the underlying source first. In OctaDezx, review a proposed lesson, approve it only if it is correct, and repeat the question in a test conversation. Customer messages do not authorize policy changes.
| Wrong answer | Source correction or lesson | Review check |
|---|---|---|
| All our jackets are waterproof | Only claim waterproofing when the specific product record states it | Approve the correct rule; test a water-resistant and an unspecified jacket |
| We can refund any used item | Use the written condition for eligible unworn items; send exceptions to the team | Review the policy; test an eligible return and a used-item exception |
The five ways a catalogue-trained agent still gets it wrong
| Failure mode | What goes wrong | Practical fix |
|---|---|---|
| Missing variant data | A product exists but the requested size is absent | Add variant facts or require a human availability check |
| Stale stock | The source changed after the last import | Refresh the records and verify the date and returned availability |
| Ambiguous policy | Flexible returns leaves eligibility undefined | Write the window, item condition, proof and exception owner |
| Unlisted product | The shopper asks for something outside the catalogue | Say it is not confirmed; ask the team instead of inventing a listing |
| Overconfident extrapolation | One item's features are applied to the whole range | Require product-specific evidence and retest the unknown case |
Keeping it accurate as the business changes
- Refresh imported records when prices or availability change, then verify that the changed facts actually reached the agent
- Add a knowledge base entry the first time a new question is asked twice, not the tenth time
- Review escalated conversations weekly, they are the clearest map of what the agent still cannot do
- When a policy changes, change it in one place and check the agent repeats the new version
- Retire old promotions explicitly, an agent that still offers an expired discount will honour it in front of customers
Frequently asked questions
- How do I train an AI agent on my own products?
- Import structured product data directly from your storefront rather than describing it: names, variants, prices, stock, images and categories. Then add your policies as explicit decisions and the exact answers to your most common questions.
- What data does an AI support agent need?
- Your product catalogue, your written policies for returns, delivery, payment and hours, the answers to your most repeated questions, and two or three example exchanges that demonstrate your tone.
- How do I keep an AI agent accurate as my business changes?
- Re-import when source prices, availability or policies change, and check what your connector actually refreshes. Review unknown answers and escalations, retire expired promotions and test changed records before assuming the agent knows the update.
- How often should I re-import the catalogue?
- Use your rate of change, not an arbitrary monthly schedule. Re-import after relevant price, availability or product changes and verify the resulting records. Do not assume continuous inventory sync unless your configured connector demonstrably provides it.
- What should happen when a product is discontinued?
- Mark it unavailable or remove the active listing, retire outdated knowledge entries and check that the agent no longer recommends it. Offer an alternative only if its features and availability are verified.
- Can I correct a bad answer?
- Yes. Correct the underlying product or policy data, review the original exchange and test the question again. OctaDezx can propose a lesson from corrections, but an operator must approve an AI-written lesson before it affects future responses.
- Does this process apply outside OctaDezx?
- Yes. Structured facts, clear policy conditions, test questions and human review are useful regardless of the AI support tool. Import formats and approval controls vary by product.
Where this fits in OctaDezx
- Answers from your product catalogue
- Connect your Shopify store
- Review and approve lessons
- See the customer care platform
Keep reading
- How to reduce Shopify support tickets
Find the sources of Shopify support tickets, remove repeated questions at the source and keep complaints, disputes and uncertain answers with a person. - How to set up AI customer service without any technical skills
You do not need to write code, hire a developer, or understand how the model works. Here is what setting up an AI support agent actually involves, and why the hard part is not technical at all. - How to move your customer service to omnichannel without the chaos
Consolidating channels sounds like a big, risky project. Done in the right order it is a series of small, safe steps. Here is a sequence that gets you there without dropping messages along the way.
About the author
The OctaDezx team, Builders of the OctaDezx AI customer care platform. We build OctaDezx, an AI customer care platform used by online stores, restaurants, agencies and clinics to answer customers and take orders around the clock. Everything here comes from running that product and reading real support conversations across those businesses, not from a keyword brief.