AI can produce publishable content in 2026 when it runs as a pipeline with gates rather than as a single prompt. Our SlimAI Content Studio, an internal short-video pipeline built in TypeScript with Remotion and Gemini, runs a 21-step orchestrator in which every step ends as PASS, AUTO_RETRY, HUMAN_REVIEW or BLOCKED. Approvals are pinned to file checksums, spend is capped per project, and the first approved 32-second video cost $4.12 in API spend.
Generating content is cheap now. Publishing the wrong content is not. A model that drops one word from a health claim, invents a statistic or renders an off-brand frame creates work for legal, support and the brand team. This guide describes how to build a production pipeline that keeps quality, using our SlimAI Content Studio as the worked example, and what it really costs.
Why a prompt is not a pipeline
A single prompt that writes a script, picks visuals and renders a video has no place to stop. If step three goes wrong, you find out at the end, after paying for every step. A pipeline breaks the work into steps with contracts: each step takes defined inputs, produces a defined artifact, and is checked before the next one starts. That structure makes failures cheap, makes costs visible, and gives people a precise place to intervene.
The worked example: SlimAI Content Studio
SlimAI is an AI calorie app we built with Flutter, Firebase, Gemini and RevenueCat, with a web presence at slimai.ai. Its Content Studio is an internal, autonomous pipeline that produces short videos for SlimAI. It is written in TypeScript, uses Gemini for generative steps, and renders with Remotion, a framework that builds videos from React components, so layouts, captions and brand elements are code rather than hand edits. A 21-step orchestrator runs each project through to an approved file, and 279 automated tests cover its behaviour.
Four outcomes for every step
- PASS: the step's checks succeeded and the next step can start.
- AUTO_RETRY: the step failed in a way a fresh attempt can fix, such as a timeout or an output that missed a format check.
- HUMAN_REVIEW: the step produced something a person must judge, so the project waits in a review queue.
- BLOCKED: the step cannot continue without a change to the input, the brief or the rules, so nothing downstream runs.
Four outcomes sound bureaucratic. In practice they are what let the system run unattended: every failure has a defined next action, and none of them is "publish anyway".
Approvals pinned to checksums
When a reviewer approves an asset, the approval is recorded against a checksum of that exact file. If anything changes afterwards, even a re-render with the same settings that produces different bytes, the checksum no longer matches and the approval does not apply. This closes the most common gap in content workflows: approving version three and publishing version four.
Cost ceilings per project
Each project carries a spend ceiling that the orchestrator enforces, so a retry loop or an unexpectedly long script cannot run up a bill. The first approved 32-second video came in at $4.12 in API spend. That figure excludes engineering and review time, which is where most of the real cost of a pipeline sits.
Pipeline stages, tools and QA gates
| Stage | Typical tool | QA gate before the next stage |
|---|---|---|
| Brief and claims inventory | Structured brief file owned by the product team | Every product claim, number and disclaimer listed with its source |
| Script | Language model (Gemini in our pipeline) | Claims check: no number, negation or product claim missing or altered; length within limit |
| Storyboard and shot list | Language model plus brand templates | Brand rules: palette, type, logo placement and safe zones for each aspect ratio |
| Assets: imagery, footage, voice, music | Generative models, licensed libraries, text-to-speech | Rights and likeness check; realistic synthetic scenes flagged for disclosure |
| Composition and render | Remotion (React-based video rendering) | Caption timing, legibility, duration and frame checks |
| Automated QA | Scripts and model-based reviewers | Spelling, forbidden terms and audio levels; failures route to retry or review |
| Human review | Brand and claims reviewer | Approval recorded against the file checksum |
| Publish and log | Platform upload with disclosure settings | Record of who approved which file, when, and at what cost |
Brand safety and claims review
The dangerous edits are small. Removing "not" reverses a claim. Changing "up to 20%" to "20%" turns a ceiling into a promise. Deleting "as part of a calorie deficit" turns a conditional statement into an unconditional one. Our Content Studio refuses edits that drop a number, a negation or a product claim from approved copy: the edit is refused rather than silently applied, so any change to a claim has to be made deliberately by a person. Four more rules sit around that one.
- Keep a claims inventory. Every claim the content may make, with its source and approved wording, lives in one file the pipeline reads.
- Substantiate health claims. The FTC's Health Products Compliance Guidance, issued in December 2022, covers health-related apps and expects competent and reliable scientific evidence for health benefit claims.
- Never fabricate people or reviews. The FTC's rule on fake reviews and testimonials, in force since 21 October 2024, covers reviews from people who do not exist, including AI-generated ones.
- Disclose realistic synthetic media. YouTube has required creators to label realistic altered or synthetic content since March 2024, and the EU AI Act's Article 50 transparency obligations apply from 2 August 2026.
Where people stay in the loop
Automation should decide what is routine; people should decide what is risky. That means people approve the brief and the claims inventory, judge anything a step routes to review, and approve the final file. Reviews go faster when the pipeline hands over evidence rather than just a video: the script with its claims highlighted, the checks that passed, and the cost so far. A reviewer who can see why something was flagged decides in a minute instead of re-watching every frame.
Track the share of steps that end in HUMAN_REVIEW. If it stays high, the rules are too vague or the briefs are incomplete. If it drops to zero, check that the gates still fire.
Testing the pipeline itself
A content pipeline is software, and it fails like software. Our Content Studio has 279 automated tests. Whatever your number, the tests that matter most protect the gates: an edit that drops a negation must be refused, a changed file must invalidate its approval, a project at its spend ceiling must stop, and a failed step must route to the right outcome rather than fall through to the next stage. Test those with fixtures before you test anything that looks good on screen.
Starting smaller than 21 steps
You do not need a 21-step orchestrator to get most of the benefit. A team starting out can build a useful first version in five steps, then add steps wherever reviews keep finding the same problem.
- Write the claims inventory and brand rules as files the pipeline reads, not as a slide deck.
- Generate scripts with structured output, and diff every draft against the inventory for numbers, negations and product claims.
- Render from templates in code, so layout, captions and logo placement cannot drift.
- Require one human approval, recorded against a checksum of the exact file that will be published.
- Log the cost, retries and review time of every asset, and set a spend ceiling per project.
What AI content creation costs
API spend is the smallest line. Our first approved 32-second video cost $4.12 in API calls. For scale, Google's Gemini API lists Veo 3.1 video generation at $0.40 per second of 1080p video on the standard model and $0.05 to $0.12 per second on its Lite and Fast variants (October 2026 prices), so generating all 32 seconds as standard Veo footage would cost $12.80 before any retries. Composing in code and generating only what needs generating keeps both cost and brand drift down. Licences matter too: Remotion is free for individuals and teams of up to three people, and larger companies need a company licence. The biggest costs are building the orchestrator, writing the tests and paying for review time.
If you want a pipeline like this for your own content, our digital growth and AI ad creation team builds it around your claims, your brand rules and your review process.
Frequently asked questions
- Can AI create video content?
- Yes. Text-to-video models such as Google's Veo 3.1 generate short clips with audio, and frameworks such as Remotion assemble branded videos from code. Production quality comes from the pipeline around them: claims checks, brand rules, automated QA and human approval. Our SlimAI Content Studio produced its first approved 32-second video through a 21-step orchestrator with review gates.
- How much does AI content creation cost?
- Model spend is often a few dollars per short video: our first approved 32-second video cost $4.12 in API calls. Google lists Veo 3.1 generation at $0.05 to $0.40 per second of 1080p video depending on the model tier. The bigger costs are building the pipeline, testing it and paying reviewers, so budget per approved asset.
- Will AI replace content creators?
- AI replaces production steps, not accountability. Models draft scripts, generate footage and render variants quickly, but someone still decides what is true, what is on brand and what may be published. Disclosure rules on YouTube and in the EU AI Act, and FTC rules on fake testimonials, keep people responsible. Creators who run good pipelines publish more without lowering the bar.
- How do you keep AI content on brand?
- Turn the brand into rules a pipeline can check: approved claims with sources, banned phrases, and palette, type and logo placement as code templates. Refuse any edit that drops a number, negation or product claim, route judgement calls to human review, and pin every approval to a checksum of the exact file so nothing changes after sign-off.
About the author
Zoraiz Ejaz
Co-founder, Techparser
Zoraiz Ejaz is a co-founder of Techparser and leads its engineering and product practice. He has spent close to a decade designing, building and scaling mobile, web and AI products for startups and enterprise teams across health, fintech, payments, social and education, from first architecture and release pipelines through to launch and years of production support. He writes about how to scope, cost and ship software that lasts.
