The video production process at a professional services firm follows a familiar pattern: a client has a presentation deck. They want a video. A creative team takes the deck, extracts the key messages, writes a script, generates a storyboard, records voiceover, selects music, assembles an animatic, reviews, iterates, and eventually delivers a rough cut. This takes weeks. Most of that time is spent on mechanical transformation — turning slides into narrative structure — not on creative judgment.
The question: what if the mechanical steps were automated, and the human only showed up for the creative decisions?
Bill designed and built the pipeline as a browser-based workflow automation tool at BCG Design Studios. It serves two purposes: a production tool that speeds up the deck-to-video process for real client work, and the live demo centerpiece for Bill's AI Systems Design workshops — a working factory that participants can open, inspect, modify, and learn from.
Eight steps. You used to do all of them. Now you do step 1 and step 7. The system handles the rest.
The pipeline is a chain of specialized API calls, not a conversation with a chatbot. Each step has a defined input, a defined output, and a contract between them. The brief is JSON. The script is timed segments. The storyboard is visual specs with source references. No step requires understanding the full context — each one does one thing well.
Human feedback doesn't restart the pipeline. It routes to the specific step that needs to change. If the voiceover is wrong, only the voiceover regenerates. If a visual is wrong, only that visual regenerates. The system maintains state across iterations — the parts that work stay fixed while the parts that don't get refined.
"You don't have to build the whole thing at once. Start with one repeatable part of your workflow. Define the input. Define the output. Write the blueprint. Let it run. Then build another one."
The pipeline automates the mechanical transformation — slides to narrative, narrative to audio, audio to timed video. The human shows up where human judgment matters: evaluating whether the creative output matches the intent, and giving directional feedback that the system can act on. Two of eight steps are human. That's the right ratio for this kind of work.
The pipeline is deliberately transparent. Every intermediate artifact is visible — the extracted brief, the generated script, the storyboard, the individual segments. Workshop participants can inspect any step, modify the prompt that drives it, swap one API for another, add a step, remove a step. The factory is glass-walled. You can see the machinery.
The pipeline is the live demo for Bill's AI Systems Design workshops — sessions for groups of 5–20 that teach how to design multi-step AI workflows. Participants don't just watch the demo. They open the pipeline, read the code, modify the prompts, swap tools, and build variations.
The pedagogical value is in the structure, not the output. The pipeline teaches the concepts that apply to any AI workflow: intent specification, structured intermediaries, chain architecture, quality checkpoints, loop-back refinement, and the distinction between mechanical transformation (automate it) and creative judgment (keep it human).
The system is the innovation, not the tools. Every API in the pipeline is commercially available. The image generation, the voice synthesis, the language model — anyone can use them. The value is in how they're chained: what connects to what, where the human checkpoints go, how feedback loops back to specific steps instead of restarting everything. That's the design work.
Structured intermediaries prevent cascade failures. Each step produces a defined artifact — JSON brief, timed script, visual spec. If step 3 produces a bad script, you fix the script. You don't have to debug a 2,000-word prompt that tried to do everything at once. The chain architecture makes the system debuggable.
Loop-back changes the economics. In traditional video production, a revision note means going back to the beginning. In this pipeline, a revision note means rerunning one step. The cost of iteration drops by an order of magnitude, which means more iterations, which means better output. Speed is not just efficiency — it's quality through more cycles.
The best teaching tool is a working system. Slides about AI workflows don't teach AI workflows. A working pipeline that participants can break apart and reassemble does. The pipeline exists to produce videos. It also exists to teach people how to think about AI system design. Both uses reinforce each other.