---
title: "50% Rejection Rate and Still Profitable: A Content Operator Crunches the AI Math"
author: "Unique Research"
sourcePublication: "Unique Research Substack"
originalPublishedAt: "2026-09-21T14:14:57+00:00"
canonical: "https://ffcap.cn/en/research/50-rejection-rate-and-still-profitable"
source: "https://uniqueresearch.substack.com/p/50-rejection-rate-and-still-profitable"
language: "en"
---

# 50% Rejection Rate and Still Profitable: A Content Operator Crunches the AI Math

Brand clients aren’t buying “one creative video.” They’re buying “a system that can continuously produce effective content.”

Let’s start with two sets of numbers.

Traditional approach: a creative team spends a week making 3 ad videos, runs them in ads, and 1 performs well — a 33% hit rate.

Industrial approach: same time, roughly same cost, make 30 videos, run them, and 8 perform well — a 27% hit rate.

The hit rate actually dropped 6 percentage points, but the total volume of good content produced is **8x** higher.

This is the math that Liu Siyang, founder of Xingbang, has worked out. Before the Chengdu AI Entertainment Conference, I sat down with this serial entrepreneur.

Xingbang is a full-service integrated marketing company that’s been running for ten years, built on data services, with short video and e-commerce as its anchors. Liu Siyang’s own career history is essentially a compressed version of the content industry’s last decade: he started in image social apps in 2015, then moved through short video marketing, data platforms, live-stream e-commerce, MCN, and performance advertising, before diving into AI in 2025.

Asked whether there’s a constant thread through these ventures, Liu says: he’s always been interested in how new technology changes content production, and how content generates commercial value.

After our conversation, I had a clear feeling: most AI content discussions still focus on “can the model generate a good image or video.” Liu Siyang is asking a different question — how does a probabilistic model become a business whose math works out?

Between those two questions lies an entire systems engineering effort, and a set of books many companies still haven’t figured out.

For context, Liu Siyang didn’t enter AI early — he only formally laid out plans in 2025.

I asked why that timing. He broke the judgment into three conditions: the model can now handle specific creative tasks; tool calling and code execution let the model turn understanding into action; and they have enough real tasks to test whether these capabilities actually work.

The third point is the most critical. Many AI companies have a hammer and are looking for nails. Xingbang, conversely, accumulated ten years of nails and was waiting for a hammer good enough to use.

On the state of the industry, Liu’s judgment is: the model as a foundational productive force has already crossed the “can it be used?” threshold. From 2023 to 2024, the industry’s core question was “is AI-generated content watchable and usable?” By 2025, that question basically has an answer.

His own words:

> “I still remember two years ago, people were debating ‘can AI write a coherent piece of copy?’ Now the debate is ‘does AI-produced text, images, or even video have differentiation, brand personality, emotional value?’ The first question is the model’s passing grade. The second is where competitiveness starts.”

Translated: generation capability itself is no longer the barrier. The barrier has moved elsewhere.

There’s still a gap between the model’s capability ceiling and what users actually get out of it. Closing that gap requires systems engineering — a stronger model alone isn’t enough.

Because a stronger model just makes users set more complex goals. Nail one video well, and they’ll ask for continuous narrative and multi-version creation. Once tasks get longer, context, state, tool coordination, and evaluation all become engineering problems.

So what exactly is Xingbang building?

Liu’s explanation is simple: a creative assistant that can complete a work alongside you. You tell it what to do, give it product materials, references, and requirements. It unfolds the steps and gradually produces scripts, visuals, and final cuts. You see intermediate results and can request changes at any point.

What does traditional ad production look like? Find a planner to write the script, then a storyboard artist to draw frames, then a director to oversee shooting, then post-production for editing and sound. Each role works in its own production program, and information degrades as it’s passed back and forth.

Xingbang’s approach is moving all of these steps onto the same “infinitely large table.” Planners, writers, visual artists, editors — whether those roles are humans or AI — work in parallel on the same table. You see the full creative picture at any moment and can adjust any step.

This canvas also supports real-time multi-person collaboration on the same project, with humans and AI working in the same space.

What does the actual workflow look like? For e-commerce clients’ feed ad creative: they start with a brief covering product name, core selling points, tone keywords, reference videos, and delivery requirements. Then an AI writer generates scripts, an AI storyboard agent breaks down shots, visual and editing work in parallel, the internal team reviews and revises, and the final output is a vertical HD cut ready for ad placement.

The detail I find most valuable: once a workflow runs through, the entire process can be saved as a template.

Next time the same brand makes a video for a different product, brand tone, visual standards, and voice style can all be called directly. The Agent automatically remembers these “brand memories.”

One-off work becomes reusable assets. Anyone who’s worked in content knows this used to exist only in the ideal version.

When the conversation turned to cost, there’s a common industry phenomenon: fake cost reduction.

Generating one image is indeed cheap, but to find one usable image you might generate 100 first, then add human revision. Unit price dropped. Total bill didn’t.

Liu says in an AI content generation system, cost doesn’t only come from model pricing. It also comes from how much wasted work the system does. Inaccurate context leads to a wrong direction and a full redo; poorly decomposed tasks make steps that could run parallel wait on each other; missing intermediate checks let small errors propagate all the way to the final cut.

So rather than switching to a cheaper model, they care about three things: reducing wasted calls, controlling rework scope, and reusing already-confirmed content.

Of course, some costs can’t come down. Complex creative judgment, high-quality evaluation, and long-term maintenance of creative methods still require investment.

Experience in the system isn’t permanently correct either. It needs to keep updating as models, tasks, and user needs evolve.

If Liu had to keep only three metrics to evaluate an AI content production system, he’d keep: cost per qualified deliverable, delivery cycle, and reuse rate.

Note: it’s “qualified deliverable” cost, and the cost must cover the full task — model inference, media generation, failed retries, human correction all included.

One call is cheap, but if the whole system loops in circles, actual cost isn’t low. Cycle also starts from task kickoff to qualified delivery — you need to see where time goes: inference, tool execution, dependency waits, or human confirmation — before knowing what to optimize.

As for generation speed, the metric every launch event brags about, it doesn’t even make the top three.

“Aesthetics can be engineered.” This claim is controversial. Isn’t aesthetics subjective?

Liu’s answer is practical: for professional creation and enterprise use, continuity of a style line across a project matters far more than how good a single video or image looks.

Color, layout, shot duration — these can all have clear rules. Team workflows, Agent usage preferences, approved process materials can all become Agent memory,沉淀 in collaborative projects.

In the end, the Agent knows which method to use in which situation, what input it needs, which features to maintain, and what to check after completion.

In plain terms: turning a master craftsman’s feel into the system’s default settings.

Someone like me, who can barely identify a color palette, actually finds this credible. The reason is simple: a brand needs its 100th piece to look like its first. A single stunning piece isn’t worth much.

Then we get to the math from the opening.

Content people hear “50% rejection rate” and their first reaction is that the system is broken. Liu Siyang’s math goes the other way: even with a 50% rejection rate, industrialization’s overall efficiency is still far higher than traditional methods.

Because industrialization trades quantity for quality. With a large enough sample, data tells you which one is optimal.

Traditional: one week, 3 videos, 1 performs well, 33% hit rate, total harvest: 1 good piece. Industrial: same time and cost, 30 videos, 8 perform well, 27% hit rate, but 8x the total good content.

Even someone like me, who does math with a calculator, gets it: the core logic is diluting customer acquisition cost per piece of good content. How much gets rejected is secondary.

Liu even wants to retire the term “rejection rate” in favor of “asset test pass rate.” His reasoning:

> “Every piece that doesn’t perform helps the client narrow the search range for ‘what works.’”

That’s a brutal perspective shift. Rejected pieces become tuition. Every dollar spent points the next round of generation in a direction.

And they have the Agent adjust the next round based on ad placement feedback. Objective placement data determines whether the iteration is actually making progress or spinning in circles.

Of course, not everything can be industrialized.

Xingbang has a clear internal criterion: when content’s core value is “reusable structure,” industrialization is an amplifier. When the core value is “unreplicable individuality,” industrialization is a destroyer.

Feed ads, e-commerce creative, and AI comic drama fall in the former category. They have clear narrative templates, defined conversion goals, and standardized delivery formats. You can run 10 versions simultaneously, let data find the best one, then scale fast.

These three directions have another layer of identity in Liu’s eyes: three task environments for testing the AI creation system.

Comic drama tests long-task capability — characters, worldview, timeline, and visual style must stay continuous. E-commerce content tests constrained expression — product appearance, selling points, and use cases must be accurate. Feed ads test version exploration and feedback utilization.

While picking tracks, they also built a set of test questions. Smart selection.

As for content that depends on a creator’s uniqueness or inspiration capture, scaling it up destroys its value.

> “Our AI canvas creation tool doesn’t produce inspiration. It just makes inspiration land faster, more accurately, and more consistently.”

That’s what their marketing team says. Blunt enough.

When we got to business model, I asked a question many people are asking: what exactly are brands willing to pay for AI content?

Answer: both models work, but the client mix differs.

Consumer brands prefer buying outcomes directly — as long as creative can be effectively placed and scale business, they keep paying. Creative production companies — ad agencies, self-media studios, AI content shops — more often buy SaaS subscriptions directly.

I asked a sharper question: if revenue grows 100% and headcount also grows 100%, is that really an AI company?

> “I don’t think so. That means the business is still scaling by adding people. AI hasn’t truly changed how the company produces.”

His follow-up: AI products should produce a different result — the same team can complete more tasks and serve more clients.

New clients can increase compute consumption, but shouldn’t require staffing a whole new team to run planning, production, and delivery from scratch.

Following that logic, their standard for rejecting requests is clear: customization is fine, but it must help the product grow. A problem solved for one client today should be solvable faster for another client tomorrow.

If every request needs separate development and dedicated maintenance, leaving no reusable capability behind, they’ll reject it even if it has revenue.

Rejecting revenue. In a market where everyone tells growth stories, that’s a rare breed.

Toward the end, we discussed the classic loop: creative generation, placement, data feedback, AI learning, regeneration.

Liu says the loop is already happening, just being refined. What Xingbang accumulated over ten years sits right on top of this loop: the full-funnel “content, placement, conversion” data from serving brands, attribution capability for “effective content,” and long-term trust with brand clients.

Clients are willing to co-create AIGC content tests and reviews with them on day one. Using the product in practice, iterating the product through application. In Liu’s words: Xingbang entered the AI industry carrying data assets and a commercial loop.

If this loop truly closes, what happens to the content industry? He gave three predictions.

Creative is no longer “I think this is good.” It’s “based on the data model, what’s the probability this direction runs above industry average.” Creative won’t disappear, but its risk will be greatly reduced by data prediction.

Content production marginal cost will keep falling, but the premium on top-tier content will rise. Truly excellent creative and emotionally resonant stories become more scarce.

Brand content budget structure will be restructured. “Data models + strategic insight” becomes the largest budget line item, while production itself gets compressed to near-zero.

In Liu’s words: brand clients aren’t buying “one creative video.” They’re buying “a system that can continuously produce effective content.”

Production fees compressed to near zero. Every company that makes a living on production should read that twice.

Over the past decade, every wave of technology in the content industry followed the same sequence: first change production tools, then change division of labor, finally change the ledger. Short video changed it once. Live e-commerce changed it once. Now it’s AI’s turn.

Only this time, what gets rewritten first may be the order of line items on a brand’s budget sheet, while the status of content itself comes later.

* * *

_Guest: Liu Siyang, Founder of Xingbang_

**Liu Siyang:** I’ve always been interested in how new technology changes content production and how content generates commercial value. Image social was about connecting images with consumption scenarios. Short video, MCN, and live e-commerce were about understanding new expression and distribution. Data and performance advertising were about making content effectiveness analyzable and verifiable.

At the AI stage, the change goes deeper: software can now understand goals, decompose tasks, call tools, and participate in the creative process. In the past, people had to translate experience into every step. Now Agents can take on some of that execution and judgment.

When judging opportunities, I look at traffic and budget, but more importantly at how capability boundaries shift. What new tasks technology can take on determines whether the next product form works.

**Liu Siyang:** Three conditions. First, models have the basic ability to handle specific creative tasks across text and images. Second, tool calling and code execution let models turn understanding into action. Third, we have enough real tasks to test whether these capabilities actually work.

With only generation ability, products easily stay as point tools. When models can decompose tasks, use tools, observe results, and continue, we start seeing a new possibility: a multimodal creation workbench that works continuously around one creative goal.

**Liu Siyang:** Think of it as a creative assistant that completes work alongside you. You tell it what to do, give it product materials, references, and requirements. It unfolds the steps and gradually produces scripts, visuals, and cuts. You see intermediate results and can request changes.

Previously these things were scattered across different software, requiring people to pass things back and forth. We put process and results on the same editable canvas so the next step continues directly from the previous result.

The “infinitely large table” puts planning, writing, visual, and editing — whether human or AI — working in parallel. For an e-commerce client, the brief covers product name, selling points, tone keywords, reference videos, and delivery requirements. Then AI writer generates scripts, AI storyboard breaks down shots, visual and editing work in parallel, internal team reviews, and the final vertical HD cut is delivery-ready.

Most importantly, once a workflow runs through it becomes a template. Next time the same brand makes a different product video, brand tone, visual standards, and voice style are called directly. The Agent remembers “brand memory.”

**Liu Siyang:** Simple criterion. When core value is “reusable structure,” industrialization is an amplifier. When core value is “unreplicable individuality,” industrialization is a destroyer.

Structure-driven content — feed ads, e-commerce creative, AI comic drama — has clear narrative templates, conversion goals, and standardized formats. You run 10 versions, data finds the best, then scale.

Individuality-driven content depends on creator uniqueness, irreplicable emotional connection, and moment-of-inspiration capture. In our marketing team’s words: our AI canvas tool doesn’t produce inspiration. It just makes inspiration land faster, more accurately, and more consistently.

**Liu Siyang:** We pursue controllability within constraints and recovery when things go wrong. Layer one is constraints — characters, product, aspect ratio, duration are all explicit. Layer two is automatic validation — file specs can be checked by rules; character consistency and expression adherence involve more complex judgment and may need human review. Layer three is replanning — after finding an issue, the Agent decides whether to add information, revise locally, try another method, or stop for human confirmation. Simply repeating the same call doesn’t necessarily solve the problem.

Another principle: preserve confirmed results and keep modification scope narrow. Randomness doesn’t disappear with more runs, but constraints, checks, and recovery reduce its impact on the whole project.

**Liu Siyang:** Cost per qualified deliverable, delivery cycle, and reuse rate. But with stricter technical definitions. Cost covers the full task: inference, media generation, failed retries, human correction. Cycle starts from task kickoff to qualified delivery, and you need to see where time goes. Reuse rate means Skills, assets, and confirmed experience actually reduce future work — if you saved a lot but didn’t use it next time, it’s not effective accumulation.

**Liu Siyang:** Even assuming a 50% rejection rate, industrial production’s overall efficiency is still far higher than traditional. Because industrialization trades quantity for quality. With enough samples, data tells you which is optimal.

Traditional: one week, 3 videos, 1 good, 33% hit rate, 3 videos total cost for 1 good piece. Industrial: same time and cost, 30 videos, 8 good, 27% hit rate, but 8x the good content.

The core logic isn’t “reduce rejects” — it’s lowering customer acquisition cost per good piece. I’d rather call it “asset test pass rate.” Every piece that doesn’t perform helps narrow “what works.”

**Liu Siyang:** I don’t think so. It means the business still scales by adding people. AI hasn’t changed production. An AI product should let the same team complete more tasks and serve more clients.

New clients can increase compute consumption but shouldn’t require a whole new team running planning, production, and delivery from scratch.

Customization is fine but must help the product grow. If every request needs separate development and dedicated maintenance and leaves no reusable capability, we reject it even with revenue.

[![cover](https://substackcdn.com/image/fetch/$s_!ZLrl!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc985bc9b-db2f-4a28-93b5-2c7a70f7dac4_2048x1152.jpeg)](https://substackcdn.com/image/fetch/$s_!ZLrl!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc985bc9b-db2f-4a28-93b5-2c7a70f7dac4_2048x1152.jpeg)

---

Original publication: https://uniqueresearch.substack.com/p/50-rejection-rate-and-still-profitable
On-site reading page: https://ffcap.cn/en/research/50-rejection-rate-and-still-profitable