About our AI usage
AI, and who signs.
There is AI at work across ToldUntold, in the studio and on the platform. This page says plainly where it works and what stays in human hands.
Every request runs through one gateway, operated by us, and nothing a maison gives us is used to train a model. The decisions that matter stay with people.
Our position
An instrument, not an author.
We treat AI the way the studio treats any serious instrument: it does the work it is good at, under the eye of someone who knows what good looks like.
In the studio, it reads faster than any of us: brand books, archives, craft records. What it finds becomes an episode only after editors have shaped and written it.
On the platform, it answers from what the maison has published, and it shows its sources. The decisions with weight, publishing a program, approving a window, setting the week's roster, sit with people.
The ledger
The division of labour.
The same rule runs through everything on this page. On the left, what the models produce. On the right, the people they answer to.
| Criterion | AI does | People do |
|---|---|---|
| Narrative extraction | Reads brand books, archives and craft records, and surfaces the stories buried in them. | Editors choose which stories deserve an episode, and shape them. |
| Scripts and drafts | Proposes outlines, titles and descriptions from the source material. | People write, design and edit every episode. The script is theirs. |
| Dubbing and subtitles | Generates subtitles from the audio and dubs video in the voices you choose. | Reviewers read every translated segment beside the original before anything is published. |
| Translation | Translates episodes and documents for each market. | Your market teams proofread for accuracy and brand voice, in a workflow built for it. |
| TU Agent answers | Answers advisors from your documentation, guidelines and training content. | The maison decides what the agent may read; every answer links to the episode behind it. |
| Merchandising photos | Checks submitted photos against the reference guidelines. | HQ reviews each submission, annotates it and answers the store. |
| Staffing suggestions | Reads patterns and suggests the week's coverage. | The manager decides. A roster is a person's call. |
| Questions to the data | Answers plain-language questions from the maison's own analytics. | HQ draws the conclusions and makes the calls. |
The current list of model providers is in the security pack, on request.
In the studio
AI reads the archives. People make the episodes.
A maison's knowledge is scattered: brand books, archives, product files, customer experience training, craft traditions. AI is good at reading all of it at once and surfacing the stories buried inside.
What it surfaces is raw material. Editors refine it, decide what deserves an episode, and do the writing. Every episode is scripted, designed and edited by people; that is the craft the studio exists for.
Localisation follows the same rule. Dubbing uses voices you choose, and each translated segment is reviewed beside the original before publication. Subtitles are generated from the audio, translated, then proofread by your market teams.
- AI surfaces, editors decide and write
- Dubbing reviewed segment by segment before publishing
- Subtitles proofread by your market teams
On the platform
It answers the floor and prepares the back office.
On the floor, TU Agent answers advisors from the maison's own material. The platform is its brain: the episodes, quests and programs the maison publishes, the rest of its content, and anything it chooses to index, automatically or by hand. Every answer names its sources and links back to the episode, program or product it drew on.
In the back office, it prepares the work rather than finishing it. Drop in a PDF, a deck or a video and it proposes chapters, descriptions and quizzes for HQ to edit and publish. Submitted merchandising photos are read against the reference brief, then reviewed and annotated by HQ. Rosters begin as its suggestions and become schedules only when a manager says so.
HQ can also put a question to the data in plain language and get an answer drawn from the maison's own analytics, rather than an export to dig through.
- Every TU Agent answer is sourced and linked
- Imports structured into chapters, descriptions and quizzes
- Suggestions from the system, decisions from your people
Governed AI
How our AI works.
Every AI request on ToldUntold — a translation, a subtitle, an answer from TU Agent — runs through one gateway, operated by us. Each request is logged, attributable, and routed to the model best suited to the task.
Some models run on our own infrastructure; for others we work with carefully selected providers under strict contractual terms. Whoever the model belongs to, the rules are the same: your data is never used to train models — ours or theirs — and it stays under the controls described on this page.
The routing is set per capability. Translation, subtitles, dubbing, TU Agent’s answers, the structuring of imported content, the analysis of merchandising photos: each goes to the model service assigned to it. A maison can also connect its own LLM or its own AI services in place of ours, capability by capability, and keep those requests inside its own perimeter.
- One gateway — every request logged and attributable
- No training on your data, ours or any provider’s
- Model choice per task, changed without touching your data
- Bring your own model: a maison’s LLM or AI services, connected per capability
- The current provider list is in the security pack
Questions
The questions we are asked.
Which AI providers do you use?
We operate our own AI gateway. Depending on the task, a request runs on models we host or on selected external providers, under contracts that forbid training on your data. The full, current list of providers is available in the security pack.
Is our data used to train models?
No. Nothing a maison puts on the platform is used to train a model, ours or a provider's. The contracts we sign with providers forbid it, and because every request runs through our own gateway, logged and attributable, there is a record of exactly what was asked, and where.
Who owns what the AI helps produce?
Our position is the one we hold for everything on the platform: your content, your analytics and your business intelligence remain yours, always. A model somewhere in the process does not change whose work it is. What the studio produces for your maison stays with your maison; if your legal team wants the wording, ask us and we will provide it.
Where do the models run?
Some run on our own infrastructure. Others run with selected providers, under contracts that forbid training on your data. For mainland China we offer a separate infrastructure option, with data kept inside the region. The security pack names who runs what, and we keep that list current.
Who reviews what the AI produces?
People close to the work. In the studio, editors shape what the models surface and write the episodes themselves. On the platform, HQ edits imported chapters and quizzes before publishing, and reviews every merchandising submission. Translations and subtitles are proofread by your own market teams. A roster is only a suggestion until a manager approves it.