Copilot Drafts

Copilot drafts a title, meta title, meta description, keywords, slug and an HTML description for one listing. Two sentences describe the whole procedure:

The model writes; it does not know. Every fact that may appear in the draft is handed to it; the prompt explicitly forbids everything else.

No draft reaches a channel without a human. Every draft becomes a content proposal, and only applying it writes the hub fields.

The facts-only rule

A model that is allowed to "know" a cable's data rate will produce a plausible one — and the merchant will publish it. That is why it is not asked "what do you know about this product", but handed a list of what your company knows about the listing.

What the model gets

Facts	Source
Item number, variant code, marketplace	Marketplace item
Item description and item category	Business Central item card
Manufacturer part number	Item (vendor item no.)
Item attributes as name/value pairs	BC item attributes — the only specifications that may appear in the text
Brand, manufacturer, MPN, EAN, marketplace category	Marketplace item
The current texts: title, meta title, meta description, keywords, slug	Marketplace item
The description the channel shows today — including marketing text or Icecat content	Description resolver, cut to 4,000 characters so one long page does not eat the token budget
The findings of the last audit for this listing	SEO audit

The findings are the reason the draft usually hits the mark: the model receives not just the product, but the list of defects — and the instruction to resolve every point on it.

What the model does not get

No prices. No stock. No customer data. No supplier data. None of it is in the facts, and none of it belongs on a product page a search engine indexes.

Telemetry keeps to the same rule: it reports counters and the duration of a run, never texts and never item numbers.

What the prompt prescribes

The rules that travel with every call — and the limits from your setup are written into them with their numbers, so the model is told the same values the audit will later check the finished text against:

Only the facts. No invented specification, number, compatibility, certification or claim. If the facts are thin, less text is the right answer, not guessed text.
Language from Content Language, otherwise the language of the requesting user.
Tone: precise, professional, for a buyer who knows the product category. No superlatives, no marketing filler, no exclamation marks, no prices, no delivery promises.
Lengths as in the setup. The meta title and the meta description must each read as one complete statement — never a sentence cut off just to fit.
Keywords: three to six comma-separated terms taken from the facts.
Slug: the title as a URL path, lowercase letters, digits, single hyphens, at most 80 characters.
Description: HTML using only <p>, <h2>, <ul> and <li> — what the product is and does, then the specifications from the attributes as a list, then the typical use. No <h1>, no inline styles, no links, no images.
Every audit finding has to be resolved by the draft.

Your Style Instructions from the setup travel with every call and take precedence over the general tone rules.

The answer is expected as one JSON object; a code fence around it is tolerated, nothing else is. Every text is cut to the length of its field — a model that ignores a limit still lands, and the audit of the applied listing will say so. If nothing usable comes back, the attempt ends as a proposal in status Failed with the reason on the record.

The three ways to a draft

Way	Where	Behaviour
Single, interactive	Marketplace item card, Draft Content with Copilot	The Copilot dialog Draft Listing Content: an optional instruction at the top, the draft below, plus Draft, Draft again, Keep it and Discard
Single, from the finding	Findings list, Draft Content with Copilot	After a confirmation the draft becomes a proposal whose card opens right away
Batch	Dashboard Draft Content for Errors, proposal list Draft for Listings with Errors	Every listing with an error finding, up to the run limit, with a progress dialog and a result message

The Copilot dialog

The input field holds what the draft should emphasise, or what was wrong with the last one — optional. Three suggestions take the most common wishes off your hands: Lead with the typical use, Lead with the specifications, Keep it short. Every proposed field is editable in the dialog, the description as HTML.

Keep it in the dialog does the same as applying on the proposal card: the draft is recorded as a proposal, written onto the listing, the listing flagged for the next item update and audited again immediately. That way the trail is the same whichever door a text came through. If the write fails, the draft stays in the queue as an open proposal — a message says so and names the reason.

Prerequisites

Prerequisite	Where
Azure OpenAI — endpoint, deployment and key	merchantCENTRAL AI Setup. One resource for all merchantCENTRAL modules, not one per connector
Capability "Listing Content Draft" switched on	The Copilot & AI capabilities page of Business Central. Every AI capability of the platform can be switched off individually there
License `CONTENTENGINE` active	Without it there is no draft and no apply
Content Language, Style Instructions, Max. Proposals per Run	Content Engine Setup

If either of the first two is missing, the module speaks up before the click on "draft" with Copilot is not ready for listing content and the hint to ask the administrator for endpoint, deployment and key and to switch the capability on. Both halves are checked: a resource nobody entered and a capability an administrator switched off stop the call alike.

Cost per draft

One draft = one AI call. It counts against the monthly token budget of the AI setup, like every other AI capability of the platform. The size of one call:

In: the facts of the listing, among them the current description with at most 4,000 characters
Out: at most 1,500 tokens — enough for a title, the meta pair, keywords, a slug and a few paragraphs of description. A draft that needs more is not a product page any more, it is a brochure

A Draft again is one more call, and a batch over fifty listings is fifty. Once the budget is exhausted, the hub client blocks, the draft fails and the run reports it. How much was consumed and for what is in the AI Usage Log; the ceilings are set in the AI Setup.

Two drafts of the same listing say the same things

The temperature of the call sits a little above zero: just high enough that a draft does not repeat the existing text word for word, and low enough that two runs over the same listing arrive at the same statements. Whoever wants a different direction changes the Instruction or the Style Instructions — not the randomness.