Creative production: how much does AI really take over?
Generated creatives have become cheap. That did not reduce the work — it moved it. A look at where the bottleneck sits now.
The promise of the last two years was unambiguous: one team produces ten times the creative in a fraction of the time. That is even true. Output of variants was simply never the actual problem.
Production was never the bottleneck
Anyone generating a hundred variants a week also needs a hundred test slots, enough budget for statistically meaningful signals, and someone who reads the results. Those three things do not scale with compute.
We suddenly had more creatives than traffic to evaluate them with. Head of Growth, D2C brand
The result is paradoxical: teams with generative production often test less cleanly than before, because they push too many variants into the account at once and can no longer say which change did the work.
Where the gain actually is
- Localisation: the same campaign in eight markets without eight production rounds.
- Format adaptation: one concept properly rebuilt for 9:16, 4:5 and 1:1 instead of cropped.
- Iterating on a winner instead of blind tests from a blank page.
- Early concept visuals, so internal debate happens over an image rather than a brief.
These are unspectacular applications, but they save real weeks. The big creative leap remains handmade: what gets said, to whom, and why anyone should care.
A workable division of labour
In teams where this works, a human defines the message and the visual idea, the model produces variants, and a human decides what enters the account. Quality control moves later in the chain but does not disappear. Cut it and you notice in the brand before you notice in the numbers.
Author
Keep reading
Attribution in the AI era: measuring what escapes measurement
More and more research ends in an answer instead of a click. What that means for reporting, budget allocation and the question of channel contribution.
Building a small RAG system: what you actually need
A RAG system sounds like heavy infrastructure. For a small project, four building blocks are enough: documents, an embedding model, a vector database, and a language model. Here's how they fit together conceptually, using Google Gemini and ChromaDB.
Agents in the mid-market: what really happens after the pilot
Almost every company is running an AI pilot by now. Few make it into day-to-day operations. Why that is — and what the exceptions do differently.