When it comes to AI workflows, it is important to manage expectations from the beginning.
The CEO of a cultural organization with limited staff and resources asked me how he could automate a workflow to develop documentary-style video shorts that would generate traffic to his organization’s website. To maintain credibility, it was important not to generate culturally inaccurate representations. Our conversation led me to build the Cultural Research Engine.
The Cultural Research Engine is part of a multipart workflow that produces output to be assembled for a 90-second video and blog post for social media and website content. The following video prototype was produced with the workflow.
The research engine functions as a research assistant, fetching resources and assets from UNESCO, Europeana, and various museums and institutions in the cultural heritage field. The system includes the following:
- a human-readable script
- a script formatted for a voice generator
- a prompt for background music
- resource links
- a blog post
The output also includes confidence levels labeled as Verified, Inferred, and Contextual to help source the information. As I mentioned in my article, “What Most People Get Wrong When Using AI,” human review of all AI output is essential for reducing slop.
A staff member then collects the assets, reviews for accuracy and quality, and assembles them in a nonlinear editor. Any additional B-roll needed can be generated or manually sourced. Parts of the assembly process can be automated, but this would require a budget.
We are still talking about whether to move forward with the project. The biggest issues are cost and labor. The monthly fixed subscription costs are US$30 to US$40, with additional variable costs per video depending on how much GPU time is used for B-roll generation. The blog post generation costs almost nothing. So he may end up using the research engine just for blog posts; it will still require review, but the labor would be dramatically cut overall for content production.