The two previous articles discussed The Chinese AI Microdrama Experiment: Who’s Making Money When Production Costs Hit Zero? and AI Microdramas Aren’t a Get-Rich-Quick Scheme, But They Do Open a Door for Newcomers. The former looked at where the money flows after production costs drop, while the latter explored how ordinary creators can use AI as a low-cost sandbox for experimentation. But both pieces examined the external business and environment, leaving a very practical on-the-ground question in between: when images can already be generated by models, from idea to delivery on a short film, where exactly do you need AI, and where is it outright useless?
In the Netflix
Q2 2026 Shareholder Letter, the company noted that it already had
roughly 300 of our titles using GenAI workflows in 2026,
with the largest concentration in post-production. Shortly after, during
the Q2
2026 Earnings Call transcript, they explained specific use cases and
also updated production requirements for partners on the Netflix
Partner Help page.
What’s most compelling is that it lets us see how professional film and television teams work backward to determine where AI belongs. They don’t wait for a model to generate an entire film in one go. Instead, starting from budget, timeline, image quality, and delivery requirements, they embed AI into the production bottlenecks that create the most friction. Technical feasibility alone doesn’t make a good production; what truly bridges the two is the production judgment accumulated over years in the film industry.
Many people following AI video tend to fixate on model parameters and generation quality: Can the shot be extended a few more seconds? Will the character’s face stay consistent across different scenes? Can the generated footage go straight to release? This habit implicitly assumes that once model capability crosses some threshold, production value will naturally follow. But back on the production floor, getting a model to generate a visually appealing sequence is only the beginning. What determines whether a project moves forward is whether that footage fits into the pipeline of greenlighting, shooting, and final delivery.
Creators without film production experience often generate a mountain
of shots on their computers, but when it comes time to deliver or
release, they can’t tell which material is actually usable, let alone
calculate the downstream rework and compliance costs. The
roughly 300 of our titles that Netflix mentioned in its
shareholder letter as using GenAI workflows — the word
title here is merely the company’s disclosure term for
counting productions. It doesn’t specify the denominator or the exact
methodology, and should be understood neither as 300 works entirely
generated by AI, nor as 300 pure AI microdramas. The largest known
application is concentrated in post-production.
When professional teams use GenAI, they don’t sit around waiting for the day of “end-to-end one-click film generation.” They’re not rushing to replace the entire set with AI. Instead, they first deploy the technology in those localized stages where decisions can be made earlier and wasteful spending can be avoided.
In the earnings call transcript, Netflix discussed real-world scenarios where teams use GenAI, covering scene references, sequence preparation, shot planning, VFX visual effects, and pre-vis — previewing shots and movements before actual filming.
Take a 3-to-5-minute microdrama project as an example. Before the team spends money renting locations, booking actor schedules, or building sets, they first use GenAI to generate a set of mood boards, scene reference images, pre-vis motion drafts, and shot planning sheets. The creative leads stand together and review them: Does the composition feel right? Does the character blocking flow naturally? Is the rhythm between scenes hitting the right beats? Instantly, they have an intuitive visual reference. If they discover that a certain sequence looks awkward, or that a particular visual style simply won’t work, revising the script or switching approaches at this stage is typically far cheaper than tearing everything down and starting over after formal shooting begins.
This kind of usage never demands that AI-generated imagery meet final deliverable standards in one shot. Its real benefit is bringing flaws that could only be seen after filming and editing to the very first prep table. As long as the team can make the “should we shoot it this way?” decision early, the cost of trial and error drops.
What pre-vis solves is “how clearly do we see it?” But once the scripted shots are validated as feasible, the question that follows is “can we actually afford to shoot it?”
Some ambitious shots, even when they look stunning during pre-vis, often get cut during actual production because the location is too expensive to rent, there aren’t enough extras, or the schedule is too tight. Netflix specifically called this out in the shareholder letter: without GenAI, certain key shots and sequences in a project would simply have to be abandoned.
For example, in the Indian series Glory, the Brazilian
documentary Brasil 70: A Saga do Tri, and the American
docuseries The American Experiment, production teams used
GenAI to handle complex imagery, including crowd enhancement, historical
battle sequences, and wide establishing shots that convey the scale of
an era or environment. The American Experiment includes 17
minutes of AI-enhanced footage. Netflix noted that these 17 minutes
expanded the project’s visual scope; in this specific case, production
speed was roughly twice that of the prior approach, and costs were
reduced by about half. The comparison here is against an earlier option
the production team had previously considered, the details of which were
not disclosed.
This “twice the speed, half the cost” result is a case-specific outcome relative to an earlier approach in a particular project. It cannot be directly applied to all Netflix projects, much less treated as a universal rule across the microdrama industry. Its value lies in reminding us that AI’s cost-saving effect doesn’t necessarily manifest as a reduction in the total budget. Rather, within a fixed budget and timeline, it preserves visual effects that might otherwise have been cut for being too expensive.
If we extend the logic of that 3-to-5-minute microdrama further: if the script calls for a sweeping wide shot of an ancient battle, the cost of filming it for real could instantly blow the budget. Handing the background wide shot to AI keeps the shot in the film. So why do these successful attempts all land on localized complex shots, rather than using AI to generate every single shot in the entire show?
In a microdrama, different shots carry fundamentally different narrative responsibilities, and the risks they face are on entirely different orders of magnitude. Mature production teams rarely fantasize about using a single magic tool to do everything. Instead, they evaluate which tasks carry the lowest cost of failure, and which are the most expensive yet easiest to replace.
Let’s return to that 3-to-5-minute microdrama. If you’re only generating a few mood boards or shot planning sheets during the early concept phase, these materials stay entirely within internal team communication and never make it into the final cut. Even if the AI-generated hands or feet are slightly distorted, it doesn’t matter. Moving to mid-stage environmental shots — say, using AI to fill in a bustling crowd in the background, or adding a few grand buildings to a wide shot — while traditional filming would be extremely expensive, these background images don’t need to carry a protagonist’s nuanced expressions or maintain a stable face, so replacing them is relatively painless. But once you reach core shots like close-ups of the lead’s dialogue or emotional climaxes, the situation changes. If the character’s face flickers or the lip sync is off, the entire performance breaks immersion, and downstream editing may need to be redone as well.
This distinction is clearly reflected in the partner policies published on Netflix Partner Help. If material is only used as reference images for concept exploration, it’s classified as low-risk use and doesn’t require escalated approval. But once material enters final deliverables — meaning images, sound, or text that actually appear in the finished work — or involves an actor’s talent likeness, personal data, or third-party IP, written approval is required. Netflix likewise requires partners to escalate core protagonists, key visuals, and primary fictional scenes for written approval. Even for custom workflows that chain multiple models together, every link in the pipeline must comply with data protection and licensing rules.
Since using real actors’ likenesses and rights requires written approval, many people naturally reason: if we skip real actors entirely and use purely fictional virtual characters, can we bypass likeness rights, avoid approvals, and lower production costs even further?
With fully virtual protagonists, you indeed no longer need to obtain digital replica authorization from specific real actors — meaning you don’t need permission for a digital voice or likeness that can be recognized as a particular real person. On-set coordination, location rentals, costumes, makeup, props, and stunt filming costs may also decrease accordingly. But this doesn’t mean the difficulties magically disappear; the production pressure simply shifts from the physical set to character asset maintenance and post-production retouching.
From a rights and review perspective, even slightly altering the appearance of a source face doesn’t mean you automatically sidestep risks around likeness, copyright, contract, licensing, or misidentification. How specific risks are assessed still depends on jurisdiction, image recognizability, actual use, and the specific manner of presentation. Moreover, under Netflix Partner Help requirements, final character designs, key visuals, and core fictional scenes are likewise listed within the scope requiring written escalated approval. These are the platform’s production and procurement specifications for partners; they cannot be directly treated as universal law, nor used to infer the platform’s actual approval rate.
Looking at actual production, the fully virtual route essentially hands the hardest part of the show to AI: keeping the same character consistent across dozens of shots in a few-minute episode — maintaining stable continuity in facial structure, body shape, clothing, voice, emotion, and even interpersonal dynamics. Once the image starts to drift, the costs of building character assets, manual frame-by-frame fixes, lip-sync dubbing, and repeated rounds of rework begin to accumulate. When you do the math, the total cost of the fully virtual route may not necessarily be lower than the hybrid route of real actors combined with AI post-production.
Ultimately, every approach comes with its own price to pay. All routes are technically feasible. The key is to calculate, using real production standards, which approach delivers genuine net value for the project at hand.
The details and policies Netflix has made public don’t actually present a flashy AI report card. They present a set of unglamorous production judgments. Generative models provide the foundational capability, but what determines whether a project can be completed smoothly, whether costs are genuinely reduced, and whether footage can be delivered in compliance is still the industry experience of film and television production itself.
For microdrama creators, rather than checking daily for a new “one-click film generation” model, it’s better to take a concrete project and run a real comparison. Take the same 3-to-5-minute microdrama script and try three different paths: pure live-action filming, live-action with AI post-production, and fully virtual characters. Record how many generations each path required, the labor hours spent on fixes, where rework got stuck, how material licensing was obtained, the review results of the test version on the target platform, and audience feedback after a controlled release.
Run these three paths on the same script, and you’ll see clearly the real differences in cost and delivery between each approach. The film industry’s production judgment reframes the question from “What can AI do?” to “Which part of this project genuinely needs AI?”