Suggesting Item Attributes and Item Category with AI
Product Data lets an AI suggest values for item attributes – and an item category for items that have none yet. For this, the AI reads only what your company already knows about the item, and suggests only what is stated there.
Nothing is written to an item before you have reviewed, accepted and applied the suggestions.
You start the suggestions with Suggest Attributes with AI:
- on the Item Card and in the Item List in the Copilot area – for this item or the selected items,
- on the Product Data Dashboard in the group AI – for items by filters of your choice.
You review the suggestions on the page AI Suggestions for Item Attributes. You open it with AI Suggestions on the dashboard, on the item card (group Product Data) and through search.
Prerequisites
| Prerequisite | Where |
|---|---|
| Product Data is switched on and has a license or a running demo | Setup |
| An Azure OpenAI resource is entered – endpoint, deployment and key | merchantCENTRAL AI Setup. One resource for all AI features of the platform |
| The Copilot capability Item Attribute Suggestions is switched on | Business Central page Copilot & AI Capabilities. It is switched on after installation; an administrator can switch it off at any time |
The calls go through the central merchantCENTRAL AI connection: with the key stored there, within the monthly budget set there, and every call appears in the AI Usage Log. Under Model per Feature in the AI Setup you can assign the feature a model of its own.
If anything is missing, neither a suggestion run nor applying starts. The message names what is missing and takes you to the right page.
Starting a run
The page Suggest Attributes with AI shows at the top the notice that AI-generated content may be incorrect. Below, you choose:
- under Items the items, with the same filters as for the tools: item no., item category (with Include Subcategories), vendor, manufacturer and Only Items With Value,
- under Also Suggest These Attributes attributes the AI should ask about beyond the attribute rules.
Suggest asks for confirmation and starts the run. AI Suggestions opens the results.
What the AI reads
Per item, the AI receives only these facts:
- description and description 2
- vendor item no., manufacturer, GTIN
- item category (if any)
- the attribute values the item already has
- the item's key features from merchantCENTRAL
- the item's extended texts that apply to all languages (cut to 2,000 characters)
The AI is instructed to use no knowledge of its own about product, brand or manufacturer, to leave an attribute out if the facts state no value, and to quote for every suggestion the passage in the facts word for word (Evidence).
What is asked for
Attributes. The AI is asked for the required and recommended attributes of the item category from the Attribute Rules – including those inherited from parent categories – and for the attributes you marked under Also Suggest These Attributes. The following applies:
- It is asked only for attributes the item has no value for yet, that are not blocked and for which no accepted suggestion waits to be applied – at most 40 per item.
- For every attribute the AI learns type, unit, whether several values are allowed and, for option attributes, the list of allowed values.
- The value lists of all option attributes of one call are limited to 20,000 characters together. An option attribute whose list no longer fits is not asked in that call.
Item category. Only for items without a category. The choice is every category without subcategories, as long as there are at most 200; otherwise the categories of the items of the same manufacturer (or, failing that, the same vendor), at most 200. If an accepted category suggestion is already waiting, it is not asked again. An existing item category is never replaced.
An item with nothing to ask about costs no call.
The run
- A run asks about at most 50 items, one AI call per item. If more items are in scope, the confirmation says so, and the first 50 are asked.
- Every item is saved on its own. If the run stops, the suggestions of the items already asked stay.
- If the AI gives no usable answer for an item (budget exhausted, say), it appears as a line of type No Answer with the reason. After three failures in a row, the run ends.
- A new run for an item replaces its suggestions with status New for the attributes (and the item category) it asks about, and earlier No Answer lines. All other suggestions stay.
The result message counts answered items, suggestions created, discarded answers per reason and items with nothing to ask.
What is checked before a suggestion is created
The AI's answer is checked strictly. Whatever fails is discarded and counted per reason in the result message:
| Check | Discarded if … |
|---|---|
| Attribute | it was not asked for (unknown, blocked or already with a value) |
| Option | a value is not in the attribute's value list – then the whole suggestion for this attribute |
| Decimal, Integer, Date | the value cannot be read; an integer has decimals |
| Text | the value is longer than 250 characters |
| Multiple values | several values come for an attribute that takes one value only. The facts are ambiguous then, and picking one would be a guess |
| Evidence | the evidence is missing |
| Confidence | it is missing or outside 0 to 1 |
| Item category | it is not among the categories offered |
Numbers are expected with a dot as decimal separator; a decimal comma and the attribute's unit after the number ("2,5 m") are tolerated. A comma before exactly three digits ("1,500") may just as well be a thousands separator – such values are discarded. Option values are stored as spelled in the value list, even if the AI spelled them differently.
Evidence Found shows whether the quote appears word for word in the facts that were sent. A suggestion without evidence found deserves a closer look.
Reviewing and applying
The page AI Suggestions for Item Attributes shows at the top the notice that AI-generated content may be incorrect and that every suggestion is to be checked against its evidence. Per line you see the item, Type (Attribute, Item Category or No Answer), the attribute or item category, the Suggested Values, the Confidence % (how clearly the facts state the value, 0 to 100), the Evidence, Evidence Found and the Status with Status Reason. The page opens with the open lines; Show All also shows the processed ones, Show Open only the open ones again.
| Action | Effect |
|---|---|
| Accept / Reject | For the selected lines. Nothing is written yet |
| Accept from Confidence | Accepts every new suggestion in view from the Minimum Confidence % on the Options FastTab (default 80) |
| Apply Accepted | Writes the accepted suggestions in view to the items |
| Delete Processed | Deletes applied, rejected and failed lines in view |
Every action works only on the lines the view shows with its filters. In the Show Open view, Delete Processed therefore deletes only failed lines.
When applying
- Attribute values go through the central merchantCENTRAL writer with the origin AI, the lowest rank. An AI value never replaces a value written by hand, by an import, by a supplier connector or by Product Enrichment (Icecat). The suggestion then becomes Failed with the reason Kept - the value was written by a higher-ranking source, for example by hand or by an import. The result message counts written, unchanged, kept and rejected values.
- Item categories are set after a confirmation of their own, only on items that have none yet. As when choosing it on the item card, the item gets the category's default attribute values where it has no value yet. Categories are applied before the attribute values: a default value of the category takes precedence over an AI suggestion for the same attribute.
- To set the category, the user needs the usual permissions to modify items. If they are missing or you decline the confirmation, the category suggestions stay accepted, and the result message says so.
| Status | Meaning |
|---|---|
| New | From the AI, checked, not decided yet |
| Accepted | Is written when applying |
| Rejected | Is never written |
| Applied | Written to the item, or the item had exactly these values already |
| Failed | Could not be written, or no usable answer. The status reason says why |
Costs
An item with an open question is one AI call. It counts against the monthly token budget of the AI Setup, like every other AI feature of the platform. A run over 50 items is at most 50 calls. What was used is in the AI Usage Log.
Data protection
Only the item facts named above go to the AI service, no customer or personal data. Prompt and answer are not stored. Stored are only the checked suggestions and, in the merchantCENTRAL AI Usage Log, the consumption and outcome of every call. The AI resource is your own Azure OpenAI resource.