The operating model of artificial intelligence-based product descriptions

AI Commerce can create a missing long product description or improve an existing description in the background. The work leverages product data, product-related documents, and authorized websites, generating facts before the description, and storing references and verification data for post-review.

The AI-generated text is written directly into the product's default language description on a successful run. There is no draft approval or rejection step in the workflow before saving. Therefore, implement the functionality with a small pilot, check the results on the products, and agree on a return and correction process before a large run.

This is how artificial intelligence imagery is created and verified

The AI description backend searches for a product that is eligible for processing, gathers the product name, SKU, category, current description, category guidelines, allowed domains, and product document addresses. It first requests a structured fact list, where each piece of information has a value and a supporting URL, and writes a description using only the returned facts.

After this, a local post-check compares the description numbers, units of measurement, and IP, EN, ISO, and CE markings to the fact list. An unsupported list item can be removed. An unsupported claim in the body text marks the result as Needs Review . The check reduces the risk of errors, but the administrator still needs to evaluate the product's suitability, safety claims, and commercial text.

Select allowed categories and enter instructions

Product description automation is configured in the AI Automations view at /ai-automations , on the Product Description tab. Under Categories allowed for automatic AI description , add the categories for which products are allowed to create a missing description.

Automatic generation does not use only one main category: a product is eligible if it belongs to at least one allowed category and its default language long description is empty. So select categories based on all their product connections. A product is not eligible for automatic running if it already has an AI description error stored for it.

Under Category-specific guidelines , select a category and write guidelines for things like tone of voice, intended use, compatibility, standards, and claims to avoid. The guidelines are saved when you leave the field; it’s a good idea to start your deployment with the 5–10 most important categories and avoid conflicting guidelines from overlapping categories.

Limit web search to allowed domains

Global allowed domains for Product Description Automation are defined in the same AI Automations view. Add just the hostname without the path, for example manufacturer.example , and enter a sequence number. A lower sequence number places the domain first in the list. When the global list is empty, the web search is not restricted by domain.

You can define your own Manufacturer Allowed Domains override for a manufacturer, one hostname per line. A non-empty manufacturer-specific list replaces the global list with that manufacturer's products; it does not merge with it. An empty manufacturer value inherits the global settings. Limit your search for technical products primarily to the manufacturer's website, official documentation, and trusted distributors.

Online search uses the Finland location and Europe/Helsinki time zone. This controls the search context, but does not guarantee that all results are Finnish or intended for the EU market. Always check the results for regional models, standards, dimensions and availability.

Attach primary documents for the product

AI description can use PDF, CSV, TXT, JSON, and XLSX files attached to the product. Open the product in the Products view at /products and add a product-specific technical brochure, user manual, or other specific document to the product media. The background work is instructed to use these files first and complete the missing information with a web search.

Use a product-specific document rather than a broad general catalog. The work generates a web address for the attachment and also adds it to the reference list. Hiding the file from the store interface does not make the address confidential, so do not attach material that is confidential, contains personal information, or has restricted access rights.

Automatically create missing product descriptions

The missing product description is queued when the product's default long description is empty, the product has no AI errors stored, and at least one of the product's categories is allowed. By default, the job processes a maximum of five products at a time. Products in use are processed first, followed by the newest products, and finally by the lowest ID.

The description is written in Finnish with a target of 300–1,000 words when there is sufficient verified information. If the short description is also empty, the job also requests a short description of 50–150 words. The existing short description is not replaced. Descriptions in other languages must be produced in a separate translation workflow.

Processing time depends on the web search, documents and template service. Do not promise a fixed 1–5 minute completion time for an individual product. Track the progress of the queue based on product description review data and monthly usage.

Improve existing descriptions with a mass run

The improvement of existing descriptions is initiated from the AI Automations > Product Description tab under Improve with AI (Bulk Run) . Select the root of the category tree. The system sets an improvement flag for all products in the selected category and its subcategories and displays the number of products queued.

There is no product-specific Improve with AI button in the product editor. Mass run can affect a large set of products, and the queue itself does not perform pre-approval. In improvement mode, the job retains the information from the existing description, tries to expand it with verified facts, and saves the text that existed before the change in the Original field. Use a limited pilot category first.

A stored AI error prevents the product from being processed even when the improvement flag is set. There is no function in the control panel to clear the error and requeue an individual product. Ask technical support to investigate the error and reset the status in a controlled manner before rerunning.

Check the result in the product's Descriptions tab

To check the AI description result, open the product at /products and go to the AI description section of the Descriptions tab. Check at least the following:

• The review shows the status as well as unsupported assertions and deleted lines.
• Changes compares sentences added to and removed from the current description.
• The original shows a long description recorded before the latest AI enhancement.
• The error shows the most recent error in the background job.
• The reference list shows the websites and documents used.
• Fact checking shows the structured factual content returned by the background work.

If the status is Needs Review , review each of the reported claims before you release the description. Even if the status is Ready , compare the technical values with the manufacturer's up-to-date documentation, verify the HTML structure, and read the text from the customer's perspective. Correct the description directly in the product editor; the product description will be saved using the editor's normal auto-save.

Manage category data correctly with CSV

The AI eligibility and instructions for categories can be controlled using the Category Import profile in the Data Exchange view. The fields are openai_product_description_enabled with a value of 0 or 1 and openai_product_description_instructions as text. Export the categories first, make changes to the copy, and check the preview before executing.

Product import does not accept openai_product_description_error or openai_product_description_improve fields, although product export shows AI-related columns. Do not attempt to clear errors or queue improvements with product CSV import. Use the AI Automations view for bulk execution and escalate the error reset to technical support.

Track monthly limit and usage costs

The Monthly Limit and current month usage for Product Description Automation are displayed in the AI Automations view. A limit of 0 means that there is no monthly cap on the processing volume. This running limit and the free amount for billing are different: API usage calculation first deducts 10 descriptions per month from the usage of each product description price category.

The calculation lines for billable descriptions 11–100, 101–1,000, and 1,001+ are based on the number of allowed domains:

Check actual usage via Reports > API usage at /reports . Pricing and agreements may change, so make sure you have a valid agreement for your store before running a large scale. Don't base your cost on the monthly limit alone.

Solve the most common problems

If an empty product description is not created, check that the product belongs to at least one allowed category, the default long description is empty, the month limit is remaining, and there is no value stored in the Error field. Also, verify the product name, SKU, and manufacturer so that the online search can distinguish the correct model from other variants.

If the result is short or general, check the fact-checking and references. Attach a more specific product-specific document or correct the list of allowed domains. If a manufacturer has its own override, changing the global list will not affect its products. If the list is too strict, the information needed for the job may not be found.

If the job fails, do not queue the same category repeatedly. A saved error prevents automatic retry so that the queue does not loop endlessly. Resolve the error, check the usage limit and data, and if necessary, ask technical maintenance to clear the error before the new pilot.

Expand only after an approved pilot

A controlled rollout of AI descriptions starts with 5-10 important categories and 20-50 product pilots. Include both missing descriptions and existing descriptions that need improvement. Review facts, references, original text, changes, regional information, and Needs Review cases.

Before expanding, agree who will approve the text, who will correct product information, who will manage domains and documents, and who will handle errors. Expand one category at a time only when the sample descriptions are true, useful, and on-brand. Automation only saves time if the review process and ownership of exceptions are clear.