Multi-brand content production
An AI content factory for many brands, with a human sign-off on every piece
An AI content factory is a production system that turns a brief into reviewable drafts of posts and short video for several brands. AI agents write and generate, automated QA gates check each piece, and a person approves it. It supports two approaches: brand-critical work with a sign-off on every piece, the current focus, and volume work with approval at template and rule level plus sample checks, which is in development. Every output today stays a draft, and the system has no publish step.
Content factory is a generic term, and this page does not describe The AI Content Factory agency or the Optimizely Content Factory product.
- [ Posts ]
- [ Images ]
- [ Short video ]
- [ Telegram ]
- [ LinkedIn ]
- [ Instagram ]
- [ QA gates ]
- [ Human approval ]
- [ Brand voice ]
- [ Fact-check ]
How the pipeline works: brief, draft, QA gates, approval, handoff
Every piece follows the same five steps, whatever the brand or the format.
- 1
Brief
A topic, a free-form brief, an operator draft, or a competitor post to rewrite for your brand. A strategist agent shapes the angle, the audience and the format.
- 2
Draft
The text chain runs in a fixed order: copywriter, brand editor, lawyer, skeptical reader, fact-checker, humanizer. Image and video steps start from the same brief.
- 3
QA gates
Automated checks run on model responses, generated images and every video clip. A failed check blocks, rerolls or returns the piece. A check that cannot be read counts as unknown, never as a pass.
- 4
Approval
Today a person reviews each draft and decides. Nothing moves on without that decision. In volume mode, in development, the person approves the template and rules and reviews samples instead.
- 5
Handoff
The approved piece leaves the factory as a finished draft. Publishing happens in your own tools, done by a person. The factory has no publish step.
One factory, many brands: per-brand voice, claims and assets
Each brand is an isolated workspace, not a shared prompt. Adding a brand means loading its brandbook and factbook, not building a new system.
- Every brand has its own factbook, voice rules, sources, media and audit trail.
- Numbers come only from the brand factbook, and a separate fact-check step enforces it.
- Each brand lists its channels, such as Telegram, LinkedIn or Instagram, each with its own language.
- Every artifact is traceable to its brief, inputs, prompts, models and QA result.
- A monthly plan can be loaded once, and its jobs run in parallel as independent runs.
- An operator edit can become a brand rule after confirmation, with the before, the after and the reason shown.
Posts, carousels and short video from one brief
One brief can feed several formats. The status of each is stated, so you know what runs today, what runs in pilots and what is next.
Text posts
Written by the text chain in each brand's voice and channel language. The brief can be a topic, your own draft, or a competitor post to rewrite for your brand.
Images
Candidates from several image models are shown side by side. The operator picks the frame that moves on, and the pick is stored with its reason.
Short vertical video
Produced within pilots. Script, storyboard, animation, voiceover and subtitles for vertical clips. Each clip passes the video QA gate. Whether generated video meets a given brand's bar is exactly what a pilot tests.
Carousels
In design. Roadmap, not shipped.
Footage remix
Re-cutting your own footage to a script, with voiceover and subtitles. Roadmap, not shipped.
The QA gates, with owner and fail action
Each gate has an owner, a pass condition and a defined result when it fails.
| Gate | Owner | Pass condition | Fail action |
|---|---|---|---|
| Editorial critics | Brand editor, lawyer and skeptical-reader agents | The draft fits the brand voice, raises no legal concern and holds up to a skeptical read | The draft goes back for revision before the next step |
| Fact-check | Fact-checker agent | Every number and claim traces to the brand factbook | The unsupported claim is blocked and the draft uses source-backed wording |
| AI-artifact check | External artifact-check service | Every model response and generated image has passed the check before it is saved | Fail-closed in production: nothing is saved |
| Video QA | Automated video QA gate | A frame strip, a frame-difference pass and a multimodal scan return the verdict usable | Reroll, at most 2 times. An unreadable or failed scan returns unknown, which never counts as a pass |
| Image selection | Operator | One candidate is chosen after a side-by-side comparison of several models | No frame moves on until one is picked. The pick and its reason are stored |
| Human approval | The reviewing person | The person accepts the draft | The piece stays a draft. The factory cannot publish it |
Human approval: per piece or per template
There are two valid ways to place approval, and the right one depends on the brand, the network and the format. Brand-critical work gets a sign-off on every piece. Volume work, such as many variants of a repeatable format, moves approval to the template and rules, with automated gates on every piece and a person checking samples. Many teams mix both.
Today the focus is brand-critical work, and volume mode is in development. Every output is a draft. The system has no publish step. A person reviews each piece, edits it if needed and publishes it from your own tools.
The factory keeps the audit trail up to the draft. Operator edits to a prompt log who changed it, when, the old value, the new value and the reason.
Who it fits: in-house multi-brand teams and agencies
- In-house teams that run several brands and want one engine with separate brand knowledge, instead of one shared prompt.
- Agencies that produce for several client brands and need each brand's facts and voice kept apart.
- Teams with a footage archive, who will want remix before generating from scratch once footage remix ships.
- Teams that need many variants of repeatable formats, once volume mode with template-level approval ships.
It is not a fit yet if you need volume mode in production today. Template-level approval with sample checks is still in development.
How to start: a pilot with two brands
A pilot runs the real pipeline on two of your brands before anyone commits to scale.
- 1Load two brands: the brandbook and the factbook for each.
- 2Run real briefs through the pipeline in the formats you choose.
- 3Your team reviews the drafts and the QA results.
- 4Decide go, no-go or pivot against criteria agreed in advance: artifacts, product fidelity, reviewer acceptance, measured cost and measured operator time.
Pilot length and scope are agreed per case.
To choose between volume and brand-critical production, read AI content production: volume vs quality. For the agent infrastructure and governance we build for enterprise teams, see AI agents for enterprise workflows. If the marketing strategy comes before the production line, see our fractional CMO service.
AI content factory questions
What is an AI content factory?
A production system that turns a brief into reviewable drafts of posts and short video for several brands. AI agents write and generate, automated QA gates check each piece, and a person approves the work before anything is published, per piece today or per template in volume mode, which is in development. Each brand has its own workspace with its own facts and voice rules.
How is an AI content factory different from ChatGPT or an AI content agency?
A chat assistant works one prompt at a time. A factory runs a fixed chain of agents and checks, with each brand's facts and voice rules loaded as workspace data. Compared with an agency, drafting is automated and every step leaves an audit trail, while a person still approves.
Does it publish automatically?
No. Today every output is a draft and the system has no publish step. A person reviews each piece and publishes it from your own tools.
Is it only for high-quality, low-volume content?
No. Both approaches are valid and the factory is built for both. Brand-critical work gets a sign-off on every piece, which is the current focus. Volume work moves approval to templates and rules with automated gates on every piece and sample checks, which is in development. Many teams mix the two by brand, network or format.
How does it keep different brands' voices separate?
Each brand is an isolated workspace with its own factbook, voice rules, sources, media and audit trail, so nothing flows through one shared prompt. Numbers come only from that brand's factbook, and a separate fact-check step blocks claims it cannot support.
What QA checks run before a piece reaches a person?
The text chain runs a brand editor, a lawyer, a skeptical reader, a fact-checker and a humanizer. Model responses and images pass an external AI-artifact check, fail-closed in production. Each video clip gets a frame strip, a frame-difference pass and a multimodal scan.
Can it produce short video, not just text?
Yes, within pilots. It makes vertical clips with script, storyboard, animation, voiceover and subtitles, and each clip passes the video QA gate with at most 2 rerolls. How well generated video fits a particular brand is what a pilot measures. Carousels and footage remix are next.
Who approves content and how is approval recorded?
Today a person on your side or the operator team reviews each draft and publishes it from your own tools. In volume mode, in development, a person approves the templates and rules and reviews samples. The factory records everything up to the draft: brief, inputs, prompts, models and QA result, plus who edited a prompt, when and why.
How long does a pilot take?
Length is agreed per pilot, because it depends on the brands and formats. The exit is fixed in advance: go, no-go or pivot, judged on artifacts, product fidelity, reviewer acceptance, measured cost and measured operator time.