Skip to main content
← All articles
AIGC

How 27 AIGC Films Were Produced In-House, and What It Taught Us

Short answer. With a repeatable pipeline — human-written scripts, generative video and voice synthesis, cultural adaptation, and a dual-layer human review with authority to reject — run…

Mumu D. · September 2026 · 11 min read

Download PDF

Short answer. With a repeatable pipeline — human-written scripts, generative video and voice synthesis, cultural adaptation, and a dual-layer human review with authority to reject — run entirely by our own teams. Twenty-seven films later, the biggest lessons were not about the tools: the script is the product, culture is harder than language, and a film is only finished when it is structured to be found and cited. Here is the pipeline, and the six lessons we paid for.

What did we actually make — and why in-house?

Twenty-seven working films, not twenty-seven showpieces — each one doing a real job for a service line, an office or a hiring pipeline.

The library on lifewood.com spans everything the company actually does. Service explainers cover autonomous driving annotation (down to a film on 3D point cloud segmentation and our 99.9% accuracy benchmark), global scanning and indexing, genealogy digitisation ("To Remember Us All", "Unlocking the Roots"), edge intelligence, data collection and the intelligent virtual assistant. Office films document the GPT centres in Benin, Indonesia and China and the Hong Kong Technovation hub. Culture pieces — core values, the international edition, "Hymn for the Future 2025" — carry recruitment and identity. And a run of thesis films ("Search Is Dead" on AEO/GEO, "Cultural Voice Synthesis", the P-R-M-A-C-E framework) argue positions we wanted on the record. Every one is scripted, voiced and quality-reviewed under human creative direction, and every one is published openly labelled as AIGC.

Why in-house, rather than an agency? Partly economics: industry analyses put traditional production around $4,500 per finished minute against roughly $400 with an AI pipeline — reported reductions of 70–91% — and the timeline for a 60-second piece collapsing from about 13 days to under half an hour of generation time.

At those unit costs, a 27-film library stops being a marketing indulgence and becomes ordinary operating output. But the deeper reason was that we sell AIGC production as a service, and we hold the view that you should not sell a pipeline you have not run on yourself. Our own films are the standing demo: when a client asks what human-directed AIGC looks like at enterprise scale, we point at the library and the process behind it rather than a slide.

One industry caveat framed the whole programme, and it is worth stating before the lessons: cost is no longer the differentiator, quality is. Survey data finds 65% of US adults somewhat or very uncomfortable with generative AI in advertising — a 37-point gap between what executives assume audiences feel and what audiences report. Cheap, obviously synthetic video is now abundant; the scarce thing is AI-generated work an audience trusts. That gap is what the human layers in our pipeline exist to close.

How does the production pipeline work?

Four stages, with humans owning the first and last: script, generate, adapt, review. Days per film, not weeks — but never zero humans.

Script under human creative direction. Every film starts as a written argument: what question does this piece answer, for whom, and what must it say precisely? Scripts are drafted and edited by people who know the service line — the point cloud film was written with the annotation teams, not about them — because generation quality downstream is a function of script precision upstream. Vague scripts produce beautiful, generic footage; specific scripts produce films that could only be ours.

Generate. The generative stage produces visuals, motion and narration — text-to-video and image generation for scenes, voice synthesis for narration. This is the stage the industry talks about most and the one that consumed the least of our learning. The tools improved month by month underneath us; the constraint was never "can the model render it" but "did we specify it well enough to render".

Adapt for culture and language. Lifewood operates in 50+ languages across 40+ delivery centres, and the films had to work in that world. Voice synthesis with cultural adaptation runs across 30 languages — and this stage taught us more than any other, because pronunciation is the easy part. Idiom, pacing, register, what an image connotes locally: those are review problems, and our region-native teams became a formal checkpoint rather than an optional polish.

Review with authority to reject — then publish for retrieval. Every film passes the same dual-layer human-in-the-loop review we apply to AI training data: a first pass produces, an independent second pass audits against the script and brand facts, and the reviewer can send it back. Approved films are then published as answer-engine assets, not just media: consistent "AIGC:" titling, entity-consistent descriptions that state what the film covers in the first sentence, and placement on both YouTube and the relevant The Lifewood AIGC film pipeline 1 2 3 4 SCRIPT GENERATE ADAPT REVIEW & PUBLISH Human-written, serviceteam-reviewed argument; the precision here decides everything downstream Text-to-video, image generation and voice synthesis produce scenes and narration from the script Cultural and language adaptation across 30 voice-synthesis languages, checked by region-native teams Dual-layer review with authority to reject; released AEO-ready with consistent titles and descriptions Humans own stages 1 and 4; the machine owns the middle. Days per film — but a person always signs the release.

What were the six hardest lessons?

Almost none of them were about generation quality. They were about scripts, culture, authority, consistency, findability and honesty.

Lesson 1: The script is the product. Our worst early drafts came from treating the prompt as the creative act. It is not; the script is. Once we started writing scripts the way we write annotation guidelines — one claim per beat, concrete nouns, numbers with sources — regeneration cycles dropped sharply. If a competent stranger could not storyboard your script without asking a question, the model cannot either.

Lesson 2: Culture is harder than language. The thesis of our "Cultural Voice Synthesis" film — most AI speaks the language, few understand the culture behind it — was earned, not written. Synthetic narration can be phonetically perfect and still land wrong: pacing that reads as impatient in one market, imagery that carries the wrong association in another. The fix was structural: cultural review by region-native staff became a named stage with the power to change the cut, not a courtesy comment at the end.

Lesson 3: Review needs authority, not just eyes. A reviewer who can only annotate problems produces a list; a reviewer who can reject produces quality. Films went back — for a mispronounced place name, an outdated statistic, a scene that oversold a capability — and the standard tightened with each rejection because decisions were recorded. This is the same dual-layer principle our data QA runs on, and it transferred without modification.

Lesson 4: Twenty-seven films need a system, not memory. Around film ten, drift appeared: terminology wobbling between films, visual styles diverging, the same service described three ways. The answer was boring infrastructure — a shared glossary of entity names, approved prompt and style libraries, naming conventions ("AIGC:" prefixes every title) — so that consistency stopped depending on who happened to remember the last film.

Lesson 5: A film is an answer-engine asset or it is invisible. AI engines do not watch video; they read the text around it. Industry citation research finds 94% of AI citations of YouTube content going to long-form, reference-style video, with views and subscribers showing near-zero correlation to citation — structure beats popularity, which suits a corporate channel fine. So every film ships with a question-led description, consistent entities, and a home on a crawlable page. The films now do double duty: media for people, citable surfaces for engines — the same AEO/GEO discipline we sell, applied to our own library.

Lesson 6: Label it, loudly. With most consumers uneasy about undisclosed AI content, we made the opposite bet: every film is titled AIGC, and the human direction behind it is part of the story. Nothing in the library pretends to be conventionally filmed. The label converts a trust risk into a proof point — the work demonstrates that AI-generated does not mean unsupervised.

What generalises to any team starting AIGC video?

The economics are proven and the tools are commoditising; the differentiators left are script discipline, human authority and honest labelling.

The adoption backdrop says a programme like this is now normal, not novel: industry reporting puts 73% of Fortune 500 companies using AI video tools in content workflows, 78% of marketing teams using AIgenerated video at least quarterly, and firms reporting an average 4.2x return within six months — while the top adoption barrier marketers name is in-house skills (43%), not cost. In other words, the constraint has moved exactly to where our lessons sit: the human layer.

The transferable playbook, compressed: make working assets, not showpieces — films with a job hold themselves to a testable standard. Spend your best people on scripts and reviews, and let the machine own the middle. Give the final reviewer authority to reject, and record the decisions so the bar rises over time.

Build the consistency infrastructure — glossary, prompt library, naming — before film ten, not after. Treat every film as a retrieval asset with a text surface engines can read. And label the work honestly, because in a market where 65% of adults distrust undisclosed AI content, disclosure plus visible human direction is a competitive position, not a confession.

A caution on the numbers. The figures about Lifewood's own library and process are ours, drawn from economics and adoption figures come from vendor statistics compilations with commercial interests in AI video tools, methodologies of varying transparency, and cost comparisons that do not always compare like with like — a $400 AI minute and a $4,500 agency minute are not the same product. Treat directions as reliable, precise multiples as indicative, and verify anything you intend to quote against its original source.

The economics that made 27 films rational Traditional production, cost per finished minute (reported avg.)

~$4,500 AI-pipeline production, cost per finished minute (reported avg.)

~$400 And the catch that shaped ours 65% 60-second film: traditional timeline ~13 days of US adults are uncomfortable with generative AI in ads — quality and trust, not cost, are the differentiators now 60-second film: AI-pipeline timeline ~27 minutes 73% of Fortune 500 companies already use AI video tools — a programme like this is table stakes, not novelty 43% of marketers name in-house skills, not cost, as the top adoption barrier — the human layer is the constraint Cost and timeline figures as reported by AI-video statistics compilations; the comparison is directional — the two products are not identical.

Key takeaways

    • Lifewood's 27 in-house AIGC films span service explainers (autonomous driving, scanning and indexing, genealogy, edge intelligence), GPT-office documentaries, culture films and thesis pieces — every one scripted, voiced and reviewed under human creative direction, and openly labelled AIGC.
    • In-house made sense on economics (reported ~$4,500 vs ~$400 per finished minute; ~13 days vs ~27 minutes for a 60-second piece) and on principle: we sell the pipeline, so we run it on ourselves first.
    • The pipeline is four stages with humans at both ends: human-written scripts, generative video and voice synthesis, cultural and language adaptation across 30 voice-synthesis languages, and dual-layer review with authority to reject.
    • Lesson 1: the script is the product — write it like an annotation guideline, and regeneration cycles collapse.
    • Lesson 2: culture is harder than language — region-native cultural review became a formal stage with power over the cut.
    • Lesson 3: reviewers need rejection authority, and recorded decisions tighten the standard over time.
    • Lesson 4: past ten films, consistency needs infrastructure — glossary, prompt and style libraries, naming conventions — not memory.
  • • Lesson 5: a film is an answer-engine asset — 94% of AI citations of video go to long-form reference content and popularity barely matters, so every film ships with a crawlable, entity-consistent text surface.
    • Lesson 6: label AI work loudly — with 65% of adults wary of undisclosed AI content, disclosure plus visible human direction is a proof point, not a confession.
    • Industry adoption (73% of Fortune 500, 78% of marketing teams) says the capability is table stakes; the skills gap (43% cite it as the top barrier) says the human layer is where programmes now win or fail.

Sources and further reading

Frequently asked questions

The visuals, motion and narration are generated; the scripts, creative direction, cultural adaptation and final approval are human. Every film is labelled AIGC and passes the same dual-layer review Lifewood applies to AI training data before release.

Concept to delivery runs in days rather than weeks — generation itself is the fast part. The schedule is set by the human stages: script development with the service team, cultural adaptation, and review cycles when a cut is rejected.

A changing set — deliberately. Generation models improved month by month under the programme, so the pipeline is built around stable human stages (script, adaptation, review) with swappable generation in the middle, rather than around any single tool.

The honest-labelling bet has held: because every film discloses AIGC and foregrounds the human direction, the library functions as a demonstration of supervised AI rather than an attempt to pass synthetic work off as filmed. Given most consumers' unease with undisclosed AI content, we consider the label a trust asset.

Yes — that is the AIGC service line. The client pipeline is the one described here: human-scripted, generated, culturally adapted across supported languages, dual-layer reviewed, and delivered AEO/GEO-ready with the structured metadata that makes video citable.

Have an AI or visibility project in mind?

From AI evaluation and human-in-the-loop review to GEO and AEO strategy, our team can help you deploy with confidence and get found in the AI search era.

Talk to our team