Every agency in the market now describes itself as AI-powered. The label has stopped carrying information, which means you need a different way to tell them apart.
I should declare an interest: Apopo is an AI-native agency. So rather than write the version of this article where my side wins, I have tried to write the one I would want to read if I were buying – including the part where agencies like mine are the more dangerous option.
What actually separates the two models?
Not access to the tools. Everyone has the same subscriptions. What differs is where the tools sit in the workflow. In a traditional agency AI is bolted onto the edges – a first draft here, a summary there – while the shape of the work, and the headcount it needs, stays what it was. In an AI-native agency the work is designed around the leverage: fewer people, each covering more surface, because the volume tasks no longer need a person per channel.
That is the honest structural difference, and it is worth being blunt about what it is for. It is not that AI does the marketing. It is that AI does enough of the grunt work that one senior operator can run a whole funnel instead of one channel. Whether that is good depends entirely on whether the senior operator is any good – which is the same thing it depended on before.
Is AI-native genuinely faster, and faster at what?
At production and analysis, yes, substantially. Variant generation, research synthesis, reporting, tagging, first drafts, pulling a quarter of account data into something legible – these collapse from days to hours. That is real and it is not controversial.
At the things that actually gate a result, no. A media learning period takes as long as it takes. An A/B test needs the sample size it needs, and no model will manufacture statistical power. Your own approval cycle is unchanged. So when an agency claims to be ten times faster, ask ten times faster at what. An agency that has become fast at publishing and stayed slow at deciding has moved its bottleneck, not removed it.
Where are AI-native agencies worse?
Volume without judgement is the big one. When producing forty variants costs the same as producing four, the discipline that used to come from scarcity disappears, and you end up with output nobody actually chose. Related to it: publishing at a scale nobody can review, which is how brand voice drifts towards the model average and stays there.
Then the specifically dangerous one. Models invent details – dates, prices, what is included, availability. In most sectors that is embarrassing. In travel it is contractual. A Canadian tribunal held Air Canada liable in 2024 for what its own chatbot told a passenger about a fare policy; the airline argued the chatbot was responsible for itself and the tribunal disagreed. If an agency’s process lets an unreviewed factual claim reach a customer, that is your liability, not theirs.
And the quiet one: client data pasted into consumer-tier tools with no processing agreement behind them. Ask where your data goes. It is a fair question and the answer is usually revealing.
“AI can massively increase output. It cannot guarantee quality. Without expertise and judgement, all you have built is a faster way to produce slop.”
What governance should you demand?
This is the axis buyers underweight, and it is currently being oversold to them. The EU AI Act’s transparency obligations applied from 2 August 2026, and you will hear agencies use that date to sell you something. Read it carefully first. The duty to mark synthetic content machine-readably falls on whoever provides the AI system, not on you or your agency for using one. The separate duty to disclose AI-generated text covers material published to inform the public on matters of public interest, which ordinary marketing copy is not. Chatbots do have to say they are chatbots, which is the part that will actually touch a booking flow.
So the honest summary is that very little new law lands on your marketing copy, and everything old still does. None of this is legal advice and you should take your own. In the UK there is no blanket rule that you must disclose AI in an advert. The CAP Code applies to the claim regardless of how it was produced, and – this is the part people miss – a disclosure does not rescue a claim that was misleading in the first place. A claim has to stand up whoever, or whatever, drafted it. On data, the ICO’s position is unchanged in substance: a lawful basis, transparency, accuracy, and usually a DPIA. On ownership, UK law on computer-generated works is under active review, so a deliverable with no human author may carry weaker protection than you assume. Put the tools you permit, and who owns the output, in the contract.
- Which tools, on which tasks?
A specific list, not “we use AI across the business”. The answer tells you whether they have thought about it or bought a subscription.
- Who signs off, by name?
Not a team, a person. Ask what they are accountable for and what happens when something wrong gets published.
- What never goes out unreviewed?
Factual claims, pricing, availability, anything regulated, and brand voice. If there is no such list, there is no policy.
- Where does our data go?
Which processor, under what agreement, and does it train a model. Get it in writing.
- Who owns it on exit?
Accounts, pixels, data, creative and the prompts and workflows built around your business. Ask before you sign, not after.
Does Google penalise AI-generated content?
No – and the question is the wrong shape. Google’s spam policies target scaled content produced to game search rankings rather than to help anyone, and are explicit that this applies “no matter how it’s created”. A human writing forty thin pages is in exactly as much trouble as a model writing them.
Your readers are a different matter. In a survey of 8,000 consumers across eight markets fielded in late 2025, 7% said visible AI in a brand’s marketing made them trust it more and 31% said it made them trust it less. It was run by a marketing platform, so weigh it accordingly – but the direction is not surprising, and it is an argument for AI in the workflow rather than AI as the pitch.
There is a fair objection to all of this that I should answer rather than wait for. A leaner agency has less bench. If your one expert is ill, on leave or resigns, a pyramid has somebody to put in the chair and we have to have written enough down that the next person is not starting from zero. So ask any lean agency – including this one – what is documented, who the named second is, and what happens in week one without your usual contact. If the answer is a shrug, the leanness is a cost saving rather than a model.
How should you compare ROI claims?
With more scepticism than either side would like. McKinsey’s survey of just under 2,000 organisations found 88% using AI somewhere in the business, but only 39% able to attribute any earnings impact to it at all – and for most of those, under 5%. Roughly 6% were getting real value. Near-universal adoption, narrow demonstrated return.
The specific trap in agency pitches is the swap between cost per asset and cost per acquisition. AI reliably lowers the first. Only judgement moves the second, and a cheaper asset that converts worse will raise your blended cost of acquisition while every input metric improves. If an agency shows you productivity gains, ask what happened to CAC over the same period. That is also why we treat what happens after the click as part of the same job rather than a separate retainer.
So which should you choose?
A good traditional agency beats a bad AI-native one, comfortably. The label is not the variable. The variable is how much accountable human judgement is attached to your account – how senior, how much of their week you actually get, and whether they are the person who learns your business or the person who presents someone else’s work about it.
Ask that question of both kinds of agency and the useful ones separate quickly. It is the same question underneath everything else you should expect from a performance marketing agency, and the reason we cap how many accounts one expert can hold.