The useful question is not whether AI can make an image. It can make thousands. The useful question is which parts of a production pipeline can tolerate a generated image, and which parts need a camera, a licence or a human decision.
We followed four studios through a quarter of client work: a packaging refresh, a hotel campaign, a SaaS rebrand and a magazine feature. The pattern was consistent. Generation accelerated exploration and background production, and slowed down anything that required a specific person, product or claim.
Where generation genuinely helps
Mood boards, texture studies, background plates and layout roughs are ideal. These are internal or low-stakes assets where speed matters more than provenance. A designer can explore ten directions before lunch and keep the conversation about composition rather than tooling.
Simple product-adjacent scenes also work when the product is added later. A soft gradient backdrop, a neutral surface, a stylised environment: these save a studio day and are easy to replace if the client objects. The key is that the generated element is not the thing being sold.
Where they fail
Hands, text, logos and precise geometry still break at production resolution. More importantly, generated images cannot testify. A hotel room that does not exist, a packaging prototype with the wrong cap, or a customer photo that is not a customer creates a legal and trust problem that no amount of retouching solves.
Consistency across a campaign is the second failure. The same prompt rarely produces the same face, wardrobe or lighting twice. Studios that need a recurring character or a product family still shoot or render, then use generation for variation around the edges.
Licensing is the third. Model terms differ on commercial use, indemnity and training data. A client contract may require warranties the tool cannot provide. Read the terms for the specific model and plan, and keep a record of the generation date and settings for every asset that ships.
A generated image is a sketch until a human signs off on what it claims. The signature is the expensive part, not the pixels.
A workflow that keeps client work safe
Start with a brief that separates expressive and factual assets. Expressive assets can be generated, reviewed and replaced freely. Factual assets, meaning anything that depicts a real product, person, place or result, need a real source or a clearly labelled illustration.
Then set a resolution and inspection standard. Generate at the largest size the tool supports, upscale deliberately, and inspect at 100 percent for text artifacts, repeated patterns and impossible geometry. A quick pass at thumbnail size hides exactly the defects that appear in print.
Finally, log provenance. Keep the prompt, model version, date, seed if available and the human editor's name. If a client asks where an image came from, the answer should take a minute, not a forensic investigation.
Key facts
- Generation is strongest for exploration, textures and backgrounds.
- It cannot verify a real product, place, person or result.
- Character and product consistency still require shoots or renders.
- Commercial terms and indemnity vary by model and plan.
- Log prompt, model, date and editor for every shipped asset.
Pricing the work honestly
Studios that priced AI work by generation time lost money. The saving moved from production to review, and review is the part clients are least willing to pay for. The credible model charges for art direction, selection, retouching and rights review, with generation as a cheap input rather than a discount.
Explain the difference to clients in plain terms. A generated background costs less than a location shoot because it carries less evidence and fewer rights. A campaign hero with a real product still costs what it always did, plus a little more for the provenance work.
Where to draw the line
Do not generate photojournalism, medical imagery, before-and-after results or anything a reasonable viewer would read as documentary. Do not use a generated face to imply a testimonial. Do not ship a product image without a real reference, even if the client cannot tell the difference. The risk is not that someone notices; it is that someone relies on it.
Used inside those limits, generation is a useful studio tool, not a replacement for one. For the system that keeps the outputs consistent, read design systems that survive a rebrand. For the channels that distribute the work, see cookieless analytics.
Bottom line
AI image tools save the most time on work that makes no factual claim. They cost the most time on work that does. Decide which kind of asset you are making before you open the prompt box, and the pipeline stays fast without becoming careless.
