A recent study found that consumers could not tell AI-generated ads from human-made ones, yet the AI versions performed 14% weaker on short-term sales and 17% weaker on long-term brand equity. Reading it, we recognised the same pattern in our own work across design, content and paid media. AI has genuinely changed how fast and how far we can move. But the risk was never that AI-made work looks bad. It looks fine. It clears review, meets deadline, and still quietly underperforms. This piece is our attempt to name what we protect, what we hand over, and why the goal was never to produce more ordinary marketing, only faster.
Three thousand consumers watched twenty video ads. Half were made by people. Half were made almost entirely by AI, built from the same creative briefs. When asked which was which, the consumers largely could not tell. But when the researchers at Ipsos and Syracuse University's S.I. Newhouse School of Public Communications measured what actually happened in the market, they noticed a gap. The AI-made ads were, on average, 14% weaker on short-term sales potential and 17% weaker on long-term brand equity than their human-made counterparts.
We read that study the way most people in our line of work probably did, with a mix of recognition and unease. Vajra Global is not an advertising agency. But between design, content and paid media, we do the same kind of work these researchers were testing: briefs, concepts, drafts, campaigns, and judgement calls about what is good enough to put in front of a client's audience. And the finding did not surprise us so much as it named something we had already started noticing on our own desks.
AI-made work does not look obviously wrong. It usually looks polished, on-brief, ready to ship. That is exactly the problem. Work that looks finished is easy to approve without checking it properly. So the right response isn't to dismiss AI as inferior, but at the same time, you cannot trust it just because it looks done. It's to check the work carefully either way, especially when it looks fine at first glance.
Let’s start with the real gains.
Take a recent example from our design work. A client needed a visual of a survey boat analysing the depth and underwater terrain of a large body of water, a specific, technical brief. Before GenAI, this would have meant searching stock libraries and manually compositing several images together, easily a few hours of work, sometimes longer. Instead, the visual was generated using the client's own website as reference, along with a detailed prompt.
The client then clarified that the boat needed to be unmanned and controlled remotely by a technician on shore. A follow-up prompt incorporated that change within minutes, rather than the hours a manual edit would have taken.
The real value was not just speed. It was the ability to produce something specific, contextual, and genuinely customised and revise it almost instantly when the brief changed, leaving more time for concept and direction, and less for sourcing and manipulating assets.
The change has been visible in paid media too. A few years ago, AI's role in a campaign began and ended with generating ad copy from manual inputs. Today, a single trained conversation can carry an entire client's context, business model, audience, competitive position, and support us end-to-end, from understanding a brief through to strategy and targeting.
For complex, less familiar industries in particular, this is where AI adds the most value, cutting through unfamiliar terminology and giving us data to work from instead of hours of cold research. Roughly 80% of the tasks in a paid media workflow can now be handed to AI with close to 95% accuracy, and the tools have their own strengths: some are stronger at keyword optimisation, placement and copy, others at the creative variations themselves.
None of this is a small efficiency story. It is a real change in how much ground a small team can cover.
But this is precisely where the HBR study's warning becomes useful, because it puts language to something we had felt without quite naming it. None of the AI-made ads in that study were obvious failures. They were credible enough to sit in a real media environment. That is what made them risky. Work that clears internal review, meets deadline and looks entirely acceptable can still be quietly weaker where it counts, in the market, with real audiences, against real competitors.
We see this pattern quite clearly with weak or vague inputs. Content that says very little while using a great many words to say it is often the clearest sign of AI-assisted work with no real thinking behind it, and it is the first thing an experienced reviewer learns to spot. The deeper risk can be seen in the writing itself - AI can make unverified information sound entirely credible, which means the old habit of trusting a source because it reads well no longer holds. That habit has to change into something closer to checking every reference before it goes anywhere near a client.
The HBR study's own Fiat example makes the same point from a different angle. The brief asked for something energetic, defiant, playful and rebellious. The AI version took those words largely at face value and produced a moody, cinematic car chase, technically accomplished, emotionally flat. The human version built the same idea into an actual joke: engineers in lab coats "testing" the car by simulating the anger of an angry partner. Same instruction, same underlying idea, very different result, because one version had a point of view behind it and the other had a description.
The inverse is just as telling. When a brief already carries a strong, specific insight, AI can execute it close to a human standard, sometimes level with it. The study's Cheerios example, built around a simple, well-researched idea (a daughter's literal way of showing she cares about her father's heart health), performed strongly in both its human and AI-made versions, ranking near the top of the entire sample. The strength was never really in who executed the idea. It was in how strong the idea was before either version got made.
Some of our own work makes the same point even more directly. A recent packaging project needed a level of human connection with a client whose instincts were deliberately old-school and personal. This is the kind of brief that cannot be reverse-engineered from a prompt because it depends on reading a person, not a pattern. Another client's brand name carried colloquial, culturally specific nuances that meant something to the people it was written for in a way no generic output could have replicated without that context already sitting in someone's head.
"AI designs are really good if used in the right scenarios. They save time for simpler, no-brainer tasks, but they can't be trusted with something that needs real emotional or cultural sensitivity." - Gokul
It is the same story the HBR study tells us with the Chewy example. The human-made ad followed a dog ageing from puppy to senior, using a simple visual trick (a sweater pulled over the dog's head, and when it comes off, the dog has suddenly grown older) to make years pass in seconds. The AI version kept the same basic idea but then lost the specific device that made it work - it was not even clear if the two dogs were meant to be the same pet. The idea survived. The precision that made people feel something did not.
Put together, a fairly consistent line emerges, and it holds regardless of whether the work in front of us is a layout, a paragraph or a campaign.
AI leads on first drafts, structural skeletons, resizing, localisation, variations, and straightforward, pattern-based execution - the work that has a known shape and a clear answer. We lead on validating whatever it produces, on deciding what is strong enough to go live, and on the layer of judgement that sits underneath everything else: understanding the business, reading the market, defining the audience, setting strategy, allocating budget and interpreting what the results are actually telling us. None of that middle layer gets handed over wholesale, because that is exactly where a five per cent gap in accuracy, or a missed cultural read, turns into a campaign that technically ran but never really worked.
"Right now, there's no tool mature enough to fully automate a campaign end to end. Human judgement is still essential, in both organic and paid work." - Rajesh
We also do not treat AI as something we hand a brief to and wait on. We work with it, generating a first pass or a skeleton and then building on it directly, because full ownership of an asset, knowing exactly what sits where and why, is the only way to update it properly later, or to catch the moment it has drifted off-brief.
The first change worth adopting across design, content, and paid media is to stop treating AI like a single-purpose tool, the way you'd treat a piece of design software you open only when you need to design something. Treat it more like the internet: something you turn to for many different reasons across a single day, not just one. You use the internet to research, to check a fact, to find a reference, to communicate, to browse for ideas, and so on.
AI works the same way when used well. It should show up at multiple points across a project, helping with research before a brief is even written, helping during drafting, and helping again during execution, rather than being switched on for one single step and switched off for the rest. Working this way changes how a brief gets read in the first place, because the question becomes where AI can help throughout the project, not whether to use it for one task.
The second change is more personal, and harder to shortcut. You need to be able to do the work without AI before you can judge whether AI has done it well.
"If you don't know how to write, think, and structure an idea yourself, you're not really going to know how to use AI well either." - Priya
That applies well beyond writing. A designer who has never had to solve a layout problem by hand will struggle to spot when an AI-generated one is quietly wrong. A performance marketer who has never built a targeting strategy from scratch will not know when an automated recommendation is missing something the data cannot see.
The third change is about depth over speed, i.e., training AI on the actual business rather than starting fresh with a generic prompt every time. A single, well-maintained conversation that carries a client's context, history and prior decisions produces work that a one-off prompt never will, because it is building on accumulated understanding rather than reconstructing it from nothing each time.
None of this makes creative or strategic skill less relevant. If anything, the opposite is true. Creative judgement, empathy, cultural fluency and the discipline to validate a piece of work before it goes out the door are becoming the actual differentiators, precisely because the more routine work around them (resizing, versioning, first drafts, etc.) is increasingly automated.
These are the same areas the HBR research points to directly - brand strategy, briefs grounded in real insight, emotional connection, cultural interpretation, and the final call on whether a piece of work has actually earned its place in front of an audience. That call has always required a person who understands both the craft and the people the work is for. It still does.
Marketing teams are undergoing a lot of quiet pressure right now. Lower production costs look good to a CFO. Faster campaign cycles look good to a CMO. More visible AI adoption looks good to a CEO. Put those three incentives together, and the temptation is to ask for more assets, more variations, more channels, faster, without asking the harder question of whether any of it is working better than what came before. Over time, "fast enough" quietly becomes the standard, and volume starts getting rewarded over distinctiveness.
The goal was never to produce more ordinary marketing at a pace nobody could manage before. AI earns its place when it protects the time and attention that judgement, insight and craft actually need, not when it fills that time with more output that looks the part but does not move anyone.
There is a governance side to this too - one we think is easy to overlook in the rush to adopt. Consumers are largely unbothered by AI-made marketing itself, but they consistently want to know when it has been used, and regulation is beginning to catch up with that expectation, from the EU's AI Act to new disclosure requirements for synthetic performers in the US.
Heinz's own transparent use of AI in its "A.I. Ketchup" campaign is a useful reminder that being upfront about AI use can strengthen a creative idea rather than undercut it. Treating disclosure as part of good practice, not as an afterthought once regulation forces the issue, is its own kind of discipline, and it is one we think belongs alongside everything else in this piece.
We are not arguing for less AI. We are arguing for more discipline about what it is asked to do, and a clearer sense of what should never be handed over in the first place.