Through MCP, the assistant can design an AI workflow with you, add its action cards, test them on a few products and tune the prompt until the output is right. Starting the workflow is always done by a person in Elovate.
1. Start with the end result
Tell the assistant what you want to achieve and for which store views. For example: “Every product in Garden furniture needs a description of about 150 words in our tone of voice, in Dutch and German.” The assistant first checks the data, so it knows which source fields are available and how complete they are.
2. Build the chain
A workflow can chain several actions. The usual order is:
Attribute extraction and/or category mapping: structure the data first.
Content enrichment: write or improve texts, using that structured data.
Translation: translate the final texts to other store views.
The order is set by the order in which cards are added. Cards of the same type run in parallel in the same step. If the assistant warns that the order is unusual, for example translation before enrichment, it means the old text would be translated. That can be intended, but check it.
3. The chain rule
A later step only runs for a product after the earlier step's result for that product has been approved by the editor, approved by the moderator and synced. Generated or editor-approved alone is not enough. A chain of three steps therefore needs three review rounds per product. A workflow that seems to “do nothing” is usually waiting for review.
4. Test and tune
The assistant runs a card on 1 to 5 products.
You compare the generated value with the source, and read the AI's reasoning.
If it is not right, change one thing in the prompt and test again, so you can see the effect.
Choose a varied test set: one product with rich data, one with poor data, and one edge case. Use a fresh set for the final check, so you don't tune the prompt to three examples.
What a test is not: a test is a preview. Nothing is synced, nothing enters the review queue, and the next step of the chain does not start. Tests cost AI budget, so there is a limit per card per hour.
Tests do not chain
A test of a later card reads the values as they are in Elovate now, not the test output of earlier cards. If you test enrichment and then translation, the translation test translates the old description. Use the translation test to judge tone, terminology and settings, not the final text.
5. Start the workflow in Elovate
When you are happy with the tests, open the workflow in Elovate and start it. From then on the cards are locked for changes through MCP. Results appear for the assigned editors and moderators.