We’ve spent a lot of time talking about what AI means for content production. How quickly we can create, how much we can automate, and whether the quality is actually good enough to put in front of a consumer.
But I think we’re starting to move into a much more interesting phase.
AI isn’t just helping us make the content anymore. It’s starting to review it, score it and influence which creative gets made. And at the other end of the chain, AI is increasingly helping consumers decide what to watch, what to consider and, ultimately, what to buy. We could be heading towards a slightly strange situation.
AI makes or contributes to the creation of the ad. AI helps decide whether the ad is any good. And another AI may eventually be the thing the ad needs to convince. That sounds more futuristic than it actually is.
Synthetic audiences are already being used to simulate consumer response to creative and propositions. Predictive models can score work before it goes into market. Meanwhile, AI assistants are becoming another layer between brands and consumers, researching products, comparing options and making recommendations on our behalf.
Individually, all of these developments make sense. Put them together and the advertising model starts to look quite different.
This is something we’re increasingly thinking about in the work we do at ICP. We already work with brands to understand where AI should sit within content production, how new tools fit into existing workflows and, importantly, where human review and decision-making still needs to happen. Spoiler: pretty much everywhere.
In the hybrid content studio models we’ve been developing, for example, AI-assisted output still passes through defined gates for quality, accuracy and brand alignment. But the next question is bigger than simply where we put the human QC gate.
What happens when AI starts appearing at almost every point in the content chain?
Creation is one part of it. Judgement is another. Discovery is another again. And if all three continue moving in the direction they are today, we may need to rethink not only how advertising gets made, but who, or what, we’re making it for.
The first part of this shift is the one we’re probably most familiar with. AI is already moving into the production process at pace. Not just for moodboards, first drafts or internal concepts, but for work that is actually going into market.
A recent Cuisinart campaign using Amazon Ads’ Creative Agent is a good example. What would traditionally have taken between one and three months for motion design, or considerably longer for an original video shoot, was produced in around four weeks.
More importantly, the AI-generated creative wasn’t just quicker to make. It delivered a 14% lower cost per detail page view and a branded search rate 10 points higher than the comparison creative.
Traditional Medicinals followed a similar route for streaming TV. An AI-generated campaign went from concept to live in three weeks and, according to the campaign results, delivered a 1.24x higher new-to-brand purchase rate than a comparable traditionally produced campaign.
The point isn’t that AI has suddenly made traditional production redundant. It hasn’t.
What is changing is the range of things we can now create without automatically defaulting to the production model we would have used a few years ago.
We see a version of this in our own work at ICP. The conversation with clients isn’t really about whether AI can generate an image or a video anymore. It’s about where it actually makes sense within the production model.
Sometimes that means AI generation. Sometimes it means automation, templating or a more traditional production route. In the hybrid studio models we’ve worked on, that distinction is important: the technology should fit the production need, rather than rebuilding the production process around whichever AI tool happens to be getting the most attention.
We’ve seen this first-hand through our work with Google and Diageo, where we’ve been creating AI-generated content against real production use cases, while also testing the tools themselves and feeding what we learn back to help the capabilities develop.
One of the biggest lessons has been that the use case must come before the technology.
It sounds obvious, but it’s surprisingly easy to do the opposite with AI. A new tool or model launches, everyone gets excited about what it can do, and suddenly teams are looking for problems to solve with it. That can quickly lead you down the wrong path.
We’ve found it much more effective to start with the production need. What are we trying to create, at what volume, and for which channels and markets? What needs to remain editable? Where does brand control sit? What level of quality do we need, and how quickly does it need to get there?
Only then do we look at the technology. That might mean generative AI. It might mean a combination of different models and tools. In some cases, automation or established production technology is still the better answer.
That’s why we take a tech-agnostic approach. The technology is moving too quickly to build an entire production model around one tool or platform.
The hybrid studio model we’ve been developing combines marketers, creators, production specialists and AI technology rather than treating AI as a standalone production solution. And as we’ve tested AI tools in real production environments, those operational learnings have fed back into the tools, workflows and models themselves.
For me, that’s where AI production gets much more interesting. It stops being about what a particular tool can generate and starts being about what the production actually needs to achieve.
Get that order wrong and you can end up forcing a use case into a technology that isn’t ready for it.
Get it right, and you have a much better chance of finding the right combination of AI, automation, technology and human craft. And I think that’s where the conversation around AI production is starting to mature.
The early AI production stories tended to focus on the spectacle of it: a commercial made without a shoot, a campaign created in days, or thousands of images generated from a prompt.
Those examples still matter, but the question that matters more now is whether the work actually performs. We’re starting to get evidence that it can.
And once you can create viable work faster, cheaper and in far greater volume, the next problem becomes fairly obvious.
How do you decide which of it is good?
Because generating 100 creative options is only an advantage if you have a sensible way of deciding which five deserve to go any further. Increasingly, the answer to that question is also becoming AI.
If AI allows us to create more content, more quickly, it creates another problem. Someone still has to decide what’s good.
If a team can suddenly produce 20, 50 or 100 creative routes rather than five, putting all of them through the same traditional review and research process doesn’t really make sense. You might have removed the production bottleneck only to create a new one further down the line.
Unsurprisingly, AI is starting to play a role here too.
We’re seeing a growing number of tools that can analyse creative, predict performance and simulate how different audiences might respond before the work ever reaches a real person. Kantar, for example, describes AI-powered creative evaluation as a major trend for 2026, using models to predict things such as attention, emotion, purchase intent and longer-term brand impact.
Synthetic audiences take that idea even further. Instead of recruiting a group of consumers and asking them what they think, AI can simulate audience groups and generate responses based on the data and characteristics they’ve been built around.
And this isn’t quite as experimental as it might sound. Kantar says its own synthetic data approach can achieve 94–95% accuracy against its ground truth data, although that’s important to read as a result from Kantar’s own methodology rather than proof that every synthetic audience is 95% accurate.
Kantar itself says the industry isn’t fully ready yet and points to data quality, rigorous testing and guardrails as critical to making it work.
That distinction matters. Because I can absolutely see the value in this.
Imagine generating 100 creative routes and using AI to identify patterns, flag obvious weaknesses or help narrow them down to the ten worth spending more time on. You can test more ideas earlier in the process without turning every creative decision into a lengthy research exercise.
For the kind of high-volume content environments we work with at ICP, that could be really useful. The challenge isn’t just how you make more content. It’s how you maintain quality, brand consistency and good decision-making when you’re suddenly making much more of it, much faster.
In the hybrid studio models we’ve been developing, AI-assisted outputs still have defined review gates around quality, accuracy and brand alignment for exactly that reason.
But there’s a subtle shift here that I think we need to pay attention to.
There’s a difference between AI helping us make a decision and AI making the decision.If a model gives one creative route a score of 92 and another a 76, which one gets made?
On paper, the answer looks obvious. But what happens when the creative director thinks the 76 is the better idea? Do we back their judgement? Do we ask them to justify why they’re ignoring the data? Or, as these systems become more trusted, does choosing the 92 simply become the safer decision?
That’s where this becomes more than just a technology question.
Because once creative teams understand what the models reward, it’s not hard to imagine people starting to create work specifically to improve the score.
And at that point, we’re not just using AI to judge our creative. We’re potentially starting to create for the AI that’s judging it.
Until fairly recently, the basic model has been pretty simple. We create advertising for people.
The tools, channels and formats might change, but ultimately we’re trying to influence a human decision.
Now there may be AI involved at almost every stage of that process.
AI can help create the work. It can help predict which work is likely to perform. And increasingly, it may become one of the ways consumers discover, evaluate and compare products in the first place.
None of that means people disappear from the equation. But it does raise some interesting questions.
If we start relying on AI to score creative, how long before we begin optimising for the score?
If AI assistants become a major route to product discovery, how long before brands start thinking about how to persuade the assistant as well as the consumer?
And if AI starts appearing throughout the entire content chain, are we still creating solely for people, or are we starting to create for a system that sits between brands and audiences?
I don’t think we’ve reached that point yet. But I do think it’s a conversation worth having now.
Because if AI starts sitting across creation, evaluation and discovery, it will start changing how we make advertising whether we’ve really thought about it or not.