An agent that runs the LinkedIn presence of an education provider, with an approval step in front of it, because nobody publishes unchecked in the name of a chamber of commerce.
- The situation
- An education provider has a constant supply of things to communicate: new courses, information evenings, registration deadlines, topics from continuing education. In practice this rarely fails for want of will, but for want of regularity. Editorial work is the first task to slip when operations call. What was wanted was a way to run the channel dependably without creating a post for it.
- Approach
- Set up as a pilot with clearly limited scope: one channel, one format, one workflow. The approval was the decisive part. A post in the name of a chamber does not go online unchecked. Between generation and publication there is therefore a step where a human decides.
- What we did
- Content from their own recordsThe workflow pulls course data and a curated pool of topics from the organisation's own tables. Only what actually exists gets proposed, no invented offerings.
- Text and image generated separatelyOne agent writes the post, a second step produces the visual. Separately, because the two are steered differently and can be corrected individually.
- Approval before publicationEvery post lands first in an approval view with a preview. Only after confirmation does it go to LinkedIn on schedule. Without that step the system would not be usable for a chamber.
- Operated in houseServer, n8n and the publishing layer run on their own infrastructure rather than at a provider. Data and credentials stay in the building.
- Result
- The channel has been served regularly since launch without anyone sitting at it daily. Approval stays with the organisation, the workflow takes over the repetition. Images produced along the way are labelled in the post as AI-generated: a requirement of Art. 50 of the EU AI Act, designed in here from the start. Laid out as a pilot so that further channels and formats can follow if it holds up.
