The first version of an AI creative team is usually one enthusiastic designer with a subscription and a channel full of requests. It works for a month. Then the images stop matching each other, nobody can say which product photograph a given ad came from, a generated garment ships with a pocket it does not have, and the brand team asks who approved it. None of that is a tooling problem. It is the absence of an operation: roles, a pipeline, standards and a record.
This page is a design for that operation. It publishes no headcount, no budget and no throughput figure; teams differ too much for any of those to be honest. It names the roles, the order to fill them, the pipeline they run, the standards that keep generated imagery on brand, and the faults that appear in the first quarter.
How do you build an AI creative operations team?
Around a pipeline rather than a tool: brief, source, generate, check, approve, publish, measure. Fill five roles against it — a producer who owns the pipeline, a source owner who keeps the master photographs and brand assets generation starts from, a generator who makes the images, a guardian who checks every image against its source and the brand's rules, and an analyst who reads what happened after publishing. Hire the producer and the guardian before a second generator; write the standards down before the first campaign; keep a record for every published image of what it was made from. Measure by the pipeline's own named events, not by a number you cannot yet defend.
What is an AI creative operations team?
It is the group that turns a brand's source assets — product photographs, models, brand rules — into published imagery with generative tools, reliably, and can say afterwards what each image was made from. It differs from a design team in owning the sources and the record as much as the output, and from a photo studio in producing variants of what was photographed rather than photographs. Its product is not images; it is a pipeline that makes images the brand can stand behind.
The pipeline
- Brief. What is needed, for which channel, from which product, with which wearer if any, by when. A brief that does not name the source is not a brief.
- Source. The master photograph, the flat lay, the approved model, the brand scene rules. If the source is missing or poor, the pipeline stops here and asks for a photograph, not a prompt.
- Generate. The images, one change at a time where the product's fidelity matters, several variants where the choice is creative.
- Check. Every image against its source at full size, with the brand's checklist, by someone who did not generate it.
- Approve. The brand owner signs off on the set, with the record in front of them.
- Publish. With the record kept: source, tool, wearer, date, approver.
- Measure. What the image did in its channel, read back against the brief.
The roles
- Producer. Owns the pipeline, the queue and the calendar. Decides what is generated and what is photographed. The first hire, because without one the pipeline is a chat channel.
- Source owner. Keeps the library: masters, flat lays, approved models, saved looks, brand scenes. Decides when a source is good enough to derive from and commissions a reshoot when it is not. Often the same person as the producer at first; separated as the library grows.
- Generator. Makes the images: the studio workflows for product, apparel and jewellery, and a general canvas for campaign and ad visuals — in Photta, Canvas, where a prompt and reference photographs make several variants at once. Skilled in narrowing a request, not in writing a long one.
- Guardian. Checks every image against its source and the brand's rules before approval, and never checks their own generations. Owns the checklists and the fault log. The second hire, before a second generator.
- Analyst. Reads the published images' results back against the briefs and tells the producer which briefs to write next. Part-time at first; the role exists so measurement is someone's job.
- Brand owner. Not on the team; approves. Sees the record with the set.
Governance: who may generate what, and what is never generated
A creative operations team needs one document that is not a standard of quality but a standard of permission: who may generate what, from which sources, with what review, and which things are never generated at all. Without it, the first person under deadline decides.
- Sources. Only the brand's own photographs of its own products and the wearers built for it. No competitor images, no stock photographs of products, no customer photographs without a written consent that covers generation.
- The product's pixels. Never redrawn, recoloured, resized or completed from a description. A product image is made around a kept product or it is not a product image.
- People. Wearers are built in the model builder or photographed with a signed release; no real person's likeness is generated from a photograph without their written consent, and adult products stay on adult wearers.
- Claims in images. No text, badge, before-and-after or result in an image that the product page does not make in words.
- Review before publishing. Every image with a product in it compared to its master by someone other than the person who generated it, with the record kept.
- Disclosure. Where a channel or a regulator asks for generated imagery to be labelled, the label is applied; the team keeps a list of which channels ask.
- Ownership of assets. Masters, wearers, recipes and records are the brand's, stored where the brand controls them, versioned, and never rebuilt casually.
- Exceptions. Anything outside the list goes to a named person for a written yes, and the yes is kept with the asset.
The document is short, signed by whoever owns the brand's reputation, and read by every new member of the team on their first day, before they open a tool.
Standards to write before the first campaign
- The source rule. Every product image is derived from a master photograph; the product's own pixels are never generated. A request that would change the product is a reshoot.
- The wearer rule. Which models the brand uses — from a catalogue, or ones the brand photographs and keeps — who approved them, where they may appear, and that one wearer runs through a range.
- The scene rule. Backgrounds, surfaces and light the brand uses; what never appears — text, other brands' products, a claim drawn as a picture.
- The checklist. What the guardian compares, in what order, and what fails. For apparel, the quality-control checklist is a starting point; for a range, consistent models across a catalogue.
- The record. For every published image: the source file, the tool, the wearer, the date, the approver, and whether it is a photograph or derived.
- The disclosure line. What the brand says, where, about generated imagery, decided once and applied everywhere.
- The regeneration rule. A failed image is regenerated from its source, never retouched on the product or the person.
Source, generate, guard
The loop the checklist below serves, named so the other operations pages can point at it.
- Source — a library of masters, wearers and scenes good enough to derive from, owned by someone, with reshoots commissioned when it is not.
- Generate — images made from the library, one change at a time where fidelity matters, several variants where the choice is creative.
- Guard — every image checked against its source by someone else, approved with its record, published with the record kept.
Team design checklist
| Role | Owns | Fill it when |
|---|---|---|
| Producer | The pipeline, the queue, the calendar, the photograph-or-generate decision | First — before any tool is bought |
| Source owner | The library: masters, flat lays, wearers, looks, scenes; reshoot requests | With the producer; separate when the library outgrows one person |
| Guardian | Checklists, the check of every image, the fault log | Before a second generator |
| Generator | The images, one change at a time; variants for creative choices | After the producer and the guardian |
| Analyst | Results read against briefs; the next briefs | When there is a quarter of published work to read |
| Brand owner | Approval, with the record | Always; outside the team |
| Standard | Written before |
|---|---|
| Source rule | The first generation |
| Wearer rule | The first on-model image |
| Scene rule | The first campaign |
| Checklist | The first approval |
| Record | The first publication |
| Disclosure line | The first publication |
| Regeneration rule | The first failed check |
The first quarter's faults
- Generation without a guardian. The generator checks their own work, the product changes, and the brand finds out from a customer. Fix: the guardian hire, and the rule that nobody approves their own image.
- No source discipline. Images made from whatever photograph was to hand; the range stops matching. Fix: the source owner, and the stop at step two of the pipeline.
- Tool sprawl. Each person with a favourite tool and a private library. Fix: one workspace the team shares, with the looks and wearers saved in it.
- The long prompt. Requests that describe everything and fix nothing. Fix: one change per image, a narrower request on failure, the source in the request.
- No record. A published image nobody can trace. Fix: the record as a condition of publishing, not a report written later.
- Measuring too early. A figure claimed from a week of data. Fix: the analyst's job is named events read over a real period; the team publishes no number it cannot defend.
One brand, one quarter
A mid-sized fashion brand starts with a producer who was the content lead and a guardian from the merchandising team who already checks samples against specifications. They write the seven standards in the first fortnight and build the library: the season's flat lays reshot flat and complete, two approved wearers, three brand scenes saved as looks in a shared workspace. The producer decides that the hero campaign is photographed and the catalogue's on-model set is derived. A generator joins and makes the set one change at a time; the guardian checks each image against its flat lay and logs faults by type; two garments go back for reshoots because the flat lays were not good enough to derive from. The set is approved with its record and published. At the end of the quarter the analyst reads the channel's results back against the briefs and hands the producer the next season's questions. No figure is claimed for any of it, and every published image can be traced to a photograph.
Frequently asked questions
What roles does an AI creative operations team need?
A producer who owns the pipeline, a source owner who keeps the masters, wearers and scenes generation starts from, a generator who makes the images, a guardian who checks every image against its source and never their own, and an analyst who reads results back against briefs. The brand owner approves from outside the team, with the record.
Who should be hired first for AI creative production?
The producer, then the guardian, before a second generator. Generation is the easy part; the pipeline, the checking and the record are what make generated imagery something the brand can stand behind, and those are the producer's and the guardian's jobs.
How do you keep AI-generated imagery on brand?
By writing the standards before the first campaign — the source rule, the wearer rule, the scene rule, the checklist, the record, the disclosure line and the regeneration rule — and by checking every image against its source and those rules before approval. On-brand is a property of the pipeline, not of a prompt.
What should an AI creative team measure?
The pipeline's own named events first: briefs received, sources rejected for reshoots, images failed at check by fault type, sets approved, images published with a record. Then channel results read against briefs, over a period long enough to mean something. This page gives no target for any of them.
Does the team replace the photographer?
No. It depends on the photographer: the source owner's library is made of photographs, and the producer's first decision on every brief is what is photographed and what is derived. The team replaces the gap between one photograph and the many images a brand needs from it.
What this page does not claim
That a team of this shape produces a given volume, saves a given cost, or lifts any channel's results; this page measures none of it. What it claims is that generated imagery a brand can stand behind comes from a pipeline with owners, that the guardian and the record matter more than the generator, and that standards written before the first campaign are cheaper than the ones written after the first mistake.
To give a team one place to generate campaign and product visuals from a prompt and reference photographs, see Photta Canvas; for the source discipline a range needs, how to generate consistent AI models across a fashion catalogue; for what the guardian checks, the AI fashion photography quality-control checklist.