CASE STUDY
pipeline types
Agentic video production across 11 pipeline types
pipeline types
provider swapping
An agentic video pipeline turning scripts into finished video at volume — programmatic assembly with generated components composited against real brand assets, and a human review gate before anything publishes.
Producing video variants at scale — per market, per language, per product — breaks any workflow built on human editors alone.
OpenMontage is an instruction-driven production system: an LLM agent reads declarative YAML manifests and orchestrates 11 specialized pipeline types, from explainers to avatar spokespersons to localization dubs.
A dynamic tool registry enables zero-code provider swapping and cost governance across video, image, TTS and music generation, with modern composition runtimes rendering the result.
What made this hard
Repeatability over craft
The value is producing the same kind of output reliably at volume. A pipeline optimised for a single impressive result solves the wrong problem.
Generated video fails visibly
The failures are obvious to a person and invisible to any automated check, which makes human review a design requirement rather than a precaution.
Render cost scales with output
At volume it becomes a real line item, so cost per rendered output had to be tracked from the first batch.
Brand assets are not replaceable
Generated components had to composite against real footage, logos and typography rather than substitute for them.
How we built it
Template definition
The structure every output shares — segments, timing, brand treatment — so the pipeline produces variation within it rather than starting fresh.
Content assembly
Scripts and assets pulled from source systems so output reflects current information rather than a snapshot.
Programmatic composition
Deterministic editing, timing and captioning, queued and parallelised for volume.
Review gate
Human approval before publication, on every output.
The stack
What should you take from this?
- This earns its place on repetition and volume — product variants, localisation, training modules — and on nothing else. For one flagship piece, hire a production team.
- Human review is not a transitional measure. Generated video fails in ways no automated check currently catches.
- Localisation is usually the clearest return, because the alternative is reshooting.
Frequently asked questions
Will the output look generated?
Some of it will, which is why every output goes through human review before publication. Where quality matters more than volume, this is the wrong tool and we will say so.
What does it cost to run at volume?
Render cost scales with output and becomes a real line item, which is why it belongs in the business case rather than a footnote. We track cost per output from the first pilot batch.
Can it use our brand assets?
That is the normal case. Generated components composite against your real footage, logos and typography rather than replacing them, which is what keeps output recognisably yours.
Related service: AI Video Production Systems
Build something at this level
Tell us what you are building. We reply within one business day.