Generative AI in a Real Content Workflow, Not a Twitter Demo

You have seen the demo. Everyone has.

One prompt, one image, and the result is genuinely remarkable. The implicit promise underneath it is obvious: this replaces the shoot, the studio, the crew, the budget.

Then you try to deliver an actual campaign with it, and you meet the gap.

The gap is not about quality. The models are good. The gap is that a demo needs to produce one impressive thing, and client work needs to produce forty consistent things by Thursday, all of which have to survive being looked at closely by the person who owns the business.

Those are completely different problems.

Consistency is the hard part, not quality

A single generated image can be excellent. Forty generated images that look like they came from the same brand, on the same day, in the same room, under the same light, is a substantially harder ask.

This is where most attempts fall over. The individual outputs are fine. Put them in a grid as a month of social content and the inconsistency is immediately visible: the light shifts, the colour temperature drifts, the space is subtly a different space, the product is subtly a different product.

Audiences do not analyse this. They just get a vague sense that something is off, which is worse, because you cannot fix a feeling with a revision round.

Solving consistency is most of the actual work. It means locking references, reusing the same base assets, building a colour and light treatment that gets applied across everything, and being willing to throw away good images that do not match the set.

The hit rate nobody puts in the demo

You generate a lot to keep a little.

Depending on how specific the brief is, somewhere between one in five and one in fifteen outputs is usable without further work. For anything involving a real product that has to look exactly right, the ratio gets worse.

This is not a failure of the tool. It is simply how the process works, in the same way a photographer shoots four hundred frames to deliver twelve. But it changes the planning. The time cost is not in generating. It is in reviewing, comparing, rejecting, and regenerating. Anyone budgeting AI production by counting prompts has misunderstood where the hours go.

The skill is direction and selection

Prompting is the least valuable part of this chain, and it gets almost all the attention.

What actually determines whether the output is usable is knowing what the image should look like before you generate anything. Composition, light direction, focal length, what the frame is for. That is art direction, and it is the same skill it has always been. The tool changed. The judgement did not.

The second half is selection, and it is where discipline matters most. "Almost right" is worthless in client work. The client will see exactly the thing that is wrong, because it is their product and their space and they look at it every day. The hand that is not quite a hand, the label that is nearly their label, the chair that is not the chair they own.

Rejecting near misses is unglamorous and it is the difference between output and deliverables.

Where it genuinely wins

Being specific about this matters, because vagueness here is how people end up disappointed.

Restaging and relighting real assets. Photograph the real product under real conditions, fast, then rebuild the setting around it. This is the single highest value application for small businesses, because the constraint it removes is the one that actually blocked them: not the camera, the location and the lighting setup.

Variation at volume for paid social. Creative testing needs distinct concepts, not tweaks, and production cost was historically the thing that made this impossible on a small budget. This is now the cheapest part of a paid campaign rather than the most expensive.

Presenter video without a shoot. Explainer and promotional video with a consistent on screen presenter, in multiple languages, without booking a person, a studio and a day.

Localisation. Producing the same campaign in three languages without three shoots.

Where it still loses

Anything where the exact product must be accurate. If a customer will hold the thing in their hand and compare it to the image, the image has to be of the thing.

Text inside images. Still unreliable, still needs to be laid on afterwards.

Brand typography and precise brand assets. Generate around them, then apply them properly in design.

Specific real people. Use a camera.

The line we hold, and why we say it out loud

The product is real. The setting can be reconstructed.

We photograph the actual dish, the actual item, the actual thing being sold. We rebuild the room, the surface, the light. We do not generate a product that was never made, and we do not improve the product itself.

That line exists for a practical reason as well as an ethical one. A customer who orders based on an image and receives something that does not match does not blame the agency. They blame the business, publicly, in a review. And in the EU, responsibility for accurate representation of a product sits with the advertiser, which means the client carries the risk of a decision the agency made.

Tell clients what is generated, in writing, before it goes out. Not because anyone is likely to audit it, but because being asked about it afterwards is a considerably worse conversation than mentioning it first.

What the workflow actually looks like

Brief and art direction. Shoot the real elements. Generate variations against locked references. Select ruthlessly. Human editing pass for colour, retouching and brand assets. Client approval with generated elements disclosed. Publish.

Notice that the shoot does not disappear. It gets shorter, cheaper and less dependent on conditions, which is a genuinely large change for a small business. But the thing being sold still has to be photographed, and someone still has to decide what good looks like.

The point

Demos show what is possible. Production shows what is repeatable.

Only one of those is a service, and the distance between them is mostly made of rejected outputs, consistency work, and the judgement to know which one of forty near identical images is the one that will not embarrass the client.

That part did not get automated. It got more important.

Where to go next

If you want to see what these ideas look like in practice, our services pages break each one down.