Every marketing team I talk to is trying to solve the same problem: how to craft winning creative. I spoke about this at CommerceNext recently. The answer isn't another AI model, another prompt, or more tools. It's going back to your process.
Almost every AI tool on the market is built to speed up one step: making the actual asset. That's where the industry has focused, and it's where most of the investment has gone. But asset production is one stage in a longer process. Before the asset gets made, there's audience research, competitor analysis, and concept development. Those earlier stages determine whether the creative wins. The tools just haven't caught up to them yet.
Key takeaways
|
|---|
The problem isn't a missing model. It's an unchanged process.
Here's what I mean by that. A team adopts an AI image generator. They still write the brief the same way, still hand it to the same one or two people, still run it through the same three-week approval chain. The image gets made faster. Everything around the image doesn't change at all.
That's not a criticism of the team. It's just where most AI adoption in creative starts, because generating something is the most visible, most demoable part of the job. But visible isn't the same as valuable. If you map out where the actual time and judgment go in a creative cycle, research, concepting, testing, production, review, you'll usually find production is one stage out of five. AI applied only there caps your upside at a fifth of the process.
The teams getting real leverage from AI apply it at every stage, not just the one that's easiest to point a demo at.
Your audience research is already available
Before you write a single line of creative copy, you probably already have more audience insight than you're using. Product reviews on Amazon, Trustpilot, and Trust Radius. Reddit threads where your category gets discussed unprompted. Your own historical A/B test results, sitting in a dashboard nobody revisits once a test ends.
This is exactly the kind of unstructured, high-volume information an LLM is good at working through. Feed it into a model and ask it to build out what actually motivates each audience segment, not what your team assumes motivates them, and you get something closer to what your customers actually said, at a scale a person reading one review at a time never gets to.
Competitor research, and research outside your category, at a scale a person can't do alone
The same logic applies to competitor research. Facebook, Google, and TikTok's ad libraries are public. Every brand's creative, including yours, is sitting there for anyone to study. AI agents can pull that data and analyze it: how each competitor tells their story, which formats they lean on, which of their assets appears to be running longest (a decent proxy for performance).
Here's the part that's easy to miss. Looking only at your own category tends to produce the same handful of ideas everyone in that category already has. Studying a completely different industry, one with a different customer, different constraints, different creative conventions, tends to surface angles your own category's blind spots hide. A person can do this manually for a handful of brands. Across hundreds or thousands of data points, that's not a research task anymore. It's a data problem, and that's exactly where AI has the advantage over a person doing it by hand.
Pre-testing a concept before it costs you media dollars
New creative direction can be judged by instinct, or it can be judged against real audience data before you spend a cent putting it in front of anyone. Take the audience segments you've already built from public and performance data, and ask an LLM: given everything we know about this segment, how is this creative concept, this copy, this angle, likely to land?
That's not a replacement for a real A/B test. It's a filter before the test. It won't tell you exactly how a concept performs, but it will tell you when a direction is clearly off before you've spent money finding that out the hard way. Skip this step and you're still testing blind, just faster.
Production is the most automated step. Review is sometimes skipped.
Production, the part where the design actually gets built, adapted into every size, and versioned for every test, is genuinely where specialized AI tools do more than a general-purpose LLM. This is the tedious, mechanical layer: resizing, platform compliance, file naming, and the dozens of small variations testing requires. We built AI Studio because we hit this exact wall ourselves. Colgate uses it to get transcreation and adaptation for seven country teams done in under two days, work that used to run through multiple agencies at roughly $75 per asset for two dozen product variants. Alliance Pharma gets launch-ready product video in under four days, saving around 20% against other GenAI tools and 90% against agency rates.
But the step most teams quietly skip is review. When brands are listing on a dozen retail platforms and advertising on half a dozen more, checking every asset against every platform's rules by hand doesn't scale. Most of it either gets done manually at real cost, or it doesn't get done at all, and inconsistency slips through.
That work can be offloaded to AI agents that are given explicit rules to check against. What can't be offloaded is the sign-off. If there's one thing I'd want every team to take from this: don't let AI publish anything as the final, unreviewed step. Put guardrails on the AI and put a person on the output. That's not caution for its own sake. It's the difference between AI that scales your judgment and AI that quietly removes it.
The tool was never the constraint
None of this requires switching platforms or waiting for a better model. It requires looking at your process end to end and asking where AI is actually being applied versus where it's just sitting on top of the same old sequence. Most teams will find they've automated the one stage that was easiest to demo, and left the four stages that actually determine whether the creative works untouched.
