merchantCENTRAL AI Setup
Every AI feature across the merchantCENTRAL apps shares one Azure OpenAI resource. It is entered here once — not once per connector. The deployment entered here is the standard model: every feature uses it unless you assign it another one under Model per Feature.
In-context help (F1)
Press F1 to open this page.
Open the page
- Via search: merchantCENTRAL AI Setup
Why it is central
AI features are appearing across the product family — a price explanation here, an item description there. Each needs the same three things: an endpoint, a deployment and a key. Maintaining those per connector would be one chance per connector to fumble a key, and one place per connector to look when it stops working.
What is not central is switching them on and off. Each feature registers its own Copilot capability and can be switched off by name under Copilot & AI Capabilities.
Standard Model (All Features)
| Field | Meaning |
|---|---|
| Azure OpenAI Endpoint | The endpoint of your resource, e.g. https://my-resource.openai.azure.com/. Leaving it empty switches every AI feature of the platform off. |
| Deployment Name | Name of the chat completion deployment in your Azure OpenAI resource. |
| API Key | Choose the field to enter or replace the key. |
The key is not in a table field
The API key goes through the credential store into the platform's isolated storage and cannot be read back. A masked text field would still be a text field — readable through a configuration package, a RecordRef or a permission set nobody tightened.
The deployment name is not the model name
Azure resolves nothing here. The deployment name is the name you gave the deployment when you created it — it may happen to match the model, but it does not have to. If it does not match Azure character for character, the service answers with The API deployment for this resource does not exist. This is the most common setup mistake.
Which model to deploy
The recommendation is gpt-4.1-mini (as of September 2026).
The platform's AI features explain, draft and classify data you already have, under fixed constraints: length limits for SEO texts, fixed output formats, temperature 0. What is needed is reliable instruction following, not reasoning — which is exactly what the mini tier is built for.
| Model tier | Suitability |
|---|---|
| mini of the current generation | Recommended. Holds the format constraints reliably and costs little. |
| nano | Not for text. Built for classification and extraction; it misses length and format constraints often enough to create rework instead of saving money. |
Large models, pro |
Unnecessary. These tasks gain nothing from more reasoning power and cost a multiple. |
| Reasoning models | Not as the standard model. Business Central's AI module cannot call them yet, and their thinking tokens count against the same answer budget as the generated text. For single features that benefit from them — alt texts from pictures, for example — assign them under Model per Feature. |
| Embedding models | Not used. The platform speaks chat completions only. |
As a rule that survives model generations: the mini tier of whatever generation is current. Only two reasons justify a switch — the current model is being retired, or its quality is demonstrably too poor.
Create the deployment
- Open Azure AI Foundry and pick the resource whose endpoint is entered above
- Deployments → Deploy model → Deploy base model
- Select
gpt-4.1-miniand give the deployment the same name — then both sides match without further thought - Choose the deployment type (see below) and leave the rate limit at the suggested value
- Wait a minute or two after creating it before trying again in Business Central
Deployment type: Data Zone for European data
Global Standard may process your requests anywhere in the world where capacity is free. Data Zone Standard keeps processing inside the EU data boundary and costs around 10% more. The prompts carry item, price and order data — for European tenants Data Zone is the right choice, and the premium amounts to fractions of a cent per call.
As starting values for the cost estimate below: gpt-4.1-mini costs USD 0.40 per million prompt tokens and USD 1.60 per million completion tokens on Global Standard, and 0.44 and 1.76 on Data Zone Standard (as of September 2026). The Azure pricing page remains authoritative.
Models get retired
The deployment list in Azure shows a retirement date per model. Once it passes, the deployment stops answering and every AI feature stands still. Note the date and deploy the then-current mini generation well before it — and only delete the old deployment once the new one works.
Model per Feature
Not every feature needs the same model. A small model writes listing texts well and cheaply; reading a product picture and describing it without inventing details takes a model that sees and reasons. Under Model per Feature you give a single AI feature a deployment of its own.
Three rules always apply:
- No line means the standard. A feature without an assignment runs on the deployment above — exactly as before this setting existed. After an update the list is empty and everything runs as it did.
- An assignment affects only its feature. It changes neither the standard model nor the model of any other feature.
- No silent fallback. If the assigned deployment was deleted or is incomplete, or a feature sends pictures to a deployment without Accepts Images, the call stops with a message. It is not sent to the standard model instead — an answer from the wrong model looks just like one from the right model.
AI Deployments
The deployments you can assign are entered through the AI Deployments action.
| Field | Meaning |
|---|---|
| Code | Short name features are assigned by, e.g. GPT6-SOL. |
| Deployment Name | Name of the deployment in Azure, exactly as in the Foundry portal. |
| Endpoint | Empty when the deployment is on the resource above — its key is used then too. If it belongs to another resource, enter that resource's endpoint and store a key of its own under API Key. That key applies to the current company only. |
| Accepts Images | Whether the model can read pictures. Features that send pictures run only on deployments marked this way. |
| Reasoning Model | Whether the model thinks before it answers (o-series, GPT-5 and later). merchantCENTRAL calls such models directly, because Business Central's AI module does not know them yet. |
| Reasoning Effort | How long a reasoning model may think. Thinking is billed as answer tokens; Low is enough for describing pictures. |
| Price per 1M Prompt/Completion Tokens | The prices of this model, in the currency of the setup. Calls to this deployment are priced with them when they are made. |
| Reported Model, Last Tested At, Last Test Result | What the last connection test found. |
Test Connection asks the deployment a short question and, if Accepts Images is set, one about a sample picture — a red square. A model that cannot see does not always refuse a picture; sometimes it guesses. The test then fails. It costs a few hundred tokens and does not appear in the usage log.
Direct calls for pictures and reasoning models
AI calls normally go through Business Central's AI module (System.AI). That module sends pictures only for Microsoft's own apps, and it cannot call reasoning models yet. merchantCENTRAL therefore sends such calls directly to Azure OpenAI. Everything else stays the same: the feature is registered as a Copilot capability and can be switched off, the budget is checked beforehand, and the call appears in the usage log — there with the model Azure reported, the number of pictures and the reasoning tokens.
Limits
| Field | Meaning |
|---|---|
| Monthly Token Budget | How many tokens the platform may consume per calendar month. 0 means no limit. Once the budget is reached, the AI features stop and resume on the first of the next month. |
| Monthly Cost Budget | What the AI calls may roughly cost per calendar month, in the price currency. 0 means no limit. It works like the token budget but counts money instead of tokens — once features run on different models, a token no longer costs the same everywhere. Only calls whose deployment has prices entered count. |
| Budget Warning % | At what share of either budget the AI Overview starts warning. It changes nothing about what the features do. |
Answer length is not a setting
How long a single answer may be is decided by each AI feature — a price explanation needs a short paragraph, a listing draft needs several. That limit is knowledge the feature has, and it nevertheless used to sit on this page as Max Tokens, where it did nothing. It is gone. What you want to steer is monthly consumption, and the token budget above does that.
Cost estimate
Azure bills by token, and what a token costs is known only to your Azure agreement. Enter the two prices here and merchantCENTRAL turns the counted tokens into an approximate amount — on this page, in the AI Overview and on the dashboard line.
| Field | Meaning |
|---|---|
| Price per 1M Prompt Tokens | What a million input tokens cost under your agreement. The figure is on the Azure OpenAI pricing page, for the model you entered under Deployment Name. |
| Price per 1M Completion Tokens | The same for the output. Azure prices the answer higher than the question, hence two fields. |
| Price Currency | Labels the estimate. Leave it empty for your local currency. Nothing is converted: enter the prices in the currency Azure charges you in. |
| Approx. Cost This Month | The calls since the first of the month, priced at these rates. |
| Approx. Cost Since Setup | Every call since setup, including log entries the cleanup has already removed. |
An estimate, not an invoice
The prices are the ones you entered, not the ones Azure charges. The token counts are close rather than exact on newer models. And changing a price reprices the past as well — the estimate always uses the rate currently in the field. The figure answers "is this cents or hundreds"; the invoice itself is in the Azure portal.
While either price field is empty, no cost figure appears anywhere. A nought would otherwise stand where "not entered" is meant.
Usage
| Field | Meaning |
|---|---|
| Last Answered | When an AI feature last actually received an answer. Failed calls do not move it — a run of failures should not look like the service working. |
| Tokens This Month | Consumption since the first of the month. Choose the field for the calls behind it. Red once the budget is reached, amber from the warning threshold. |
| Tokens Since Setup | Total consumption, including log entries the cleanup has already removed. |
Administrators only
This page belongs to the administrator permission set. Whoever can change the monthly budget can lift the cost ceiling and switch the AI off platform-wide, so the standard user set does not include it. The AI Overview and the usage log are open to everyone.
Housekeeping
Log Retention Days decides how long a call stays in the usage log. The default is 365 days; 0 keeps every entry.
Actions
| Action | Effect |
|---|---|
| Test Connection | Asks the standard deployment a short question and reports which model answers. Checks endpoint, deployment name and key before a feature fails on them. |
| AI Deployments | Opens the further deployments for Model per Feature. |
| AI Overview | Opens the consumption overview. It is also reachable through the Artificial Intelligence line on the dashboard once the AI is set up. |
| Usage Log | Opens every recorded call. |
| Azure Pricing Page | Opens the Azure OpenAI pricing page in your browser — the two figures for the cost estimate are there. Find the model behind your deployment and take the row matching your deployment type: global, data zone and regional are priced differently. |
| Remove API Key | Deletes the stored key. Every AI feature of the platform stops afterwards. |