AI Video Production: Our Workflow from Sources to Editing

At X Mind Solutions, we brought together different stages of production using AI tools for our first video experiment. From preparing sources to generating a script and video, through voiceover and final editing, the process took approximately 1.5–2 hours. The focus was not on getting a result from a single tool, but on building a production workflow with complementary tools. This experience also provided a concrete starting point for assessing the difference between AI-assisted content production and end-to-end automation.

First, we prepared a set of topics compiled from relevant subjects covered on YouTube over more than four years. We used these topics and legislative sources in NotebookLM to generate a script and video. The core approach was to base production on specific sources. When designing a similar workflow, it is important to define the scope clearly and check that the sources are up to date. Particularly when content covers legislation, the generated text must also be checked against the sources; using sources alone does not guarantee accuracy.

Next, we transferred the transcript of the video generated by NotebookLM to Minimax. We cloned our own voice to record a new voiceover for the text, then completed the editing in CapCut. The narrative generated from the sources thus went through separate voiceover and editing stages. This structure provides a workflow in which text, audio and visuals can be assessed independently. In similar projects, pronunciation, narrative consistency and alignment between audio and visuals should be checked; voice cloning should only be carried out with the necessary usage rights and permissions.

The fact that we produced this video entirely using AI tools does not mean that the process ran without human intervention. Preparing sources, transferring outputs between tools and completing the edit were all part of the production workflow. The approximately 1.5–2-hour duration is also an observation specific to this experiment; it should not be interpreted as a general time or efficiency commitment for other videos. The scope of the content, the state of the sources and the corrections required may affect how long similar projects take.

The next question to assess is the extent to which these steps could be connected through an automation platform such as n8n. Before determining whether end-to-end automation is possible, each tool’s programmatic access options, data transfer formats and terms of use must be examined. Source ingestion, transcript transfer and output routing could be considered for automation; however, we are not claiming to have implemented such an integration in this experiment. A sound design should also retain human approval checkpoints, particularly for legislative accuracy and pre-publication quality review.