One of the promises of AI is scale. Give it a task that would take a human hours, days or weeks, and it can potentially complete it across thousands of customers, products or pieces of content in minutes.
That’s incredibly powerful. But there’s a step we’re increasingly tempted to skip: working out what we actually want AI to scale.
I was reminded of this recently during a free consultation with a US-based founder. He was developing a tool for real estate agents and planning an outbound email programme to acquire users.
We’d been discussing how he could make his emails genuinely relevant to individual prospects. Rather than simply inserting readily available information such as someone’s name or location, we started exploring whether he could analyse the videos each agent was already publishing. How frequently were they posting? What type of videos were they creating? How complicated were those videos to produce?
Then he told me he’d already watched around 30 real estate videos himself and started categorising them. Some were simple one-take videos. Others used multiple cuts. Some required considerably more production effort.
And that’s when it became clear he’d already done the most important part: he’d done the thinking.
The next question wasn’t whether AI could come up with a clever way to analyse real estate videos. It was whether AI could reliably apply his framework to thousands of them. There’s a big difference.
We’re asking AI to think before we’ve thought
One of the easiest things to do with generative AI is start with a blank page and ask it to analyse something, find customer segments, identify patterns, suggest A/B tests or personalise an email.
Sometimes the results will be useful. But we’re also handing AI an enormous amount of strategic responsibility without necessarily realising it.
We haven’t defined what matters. We haven’t established the criteria. We haven’t formed a hypothesis. We haven’t decided what good looks like.
We’ve effectively said: You work it out.
Then, because the answer arrives quickly, confidently and in a beautifully organised list, it can feel as though the hard work has been done. It hasn’t necessarily.
The framework is where the value lives
Think about what the founder had actually done by watching those videos. He wasn’t simply gathering information. He was learning.
He’d seen enough examples to start recognising meaningful differences between them. He’d worked out which characteristics might indicate how much editing effort was involved. He was beginning to create a taxonomy based on his understanding of the problem his product solved.
That’s expertise being formed.
Once you have that framework, AI becomes incredibly useful. Instead of asking it to analyse the videos and tell you what’s interesting, you can ask it to identify specific characteristics you’ve already determined matter.
Now AI isn’t deciding the strategy. It’s applying it. And that distinction matters far beyond this particular example.
Do it manually before you automate it
I’ve always believed there’s value in doing something manually before trying to automate it. Not forever. Just long enough to understand it.
Take customer segmentation. You could give AI a dataset and ask it to identify some segments. You’ll almost certainly get some. But are they useful segments? Do they reflect meaningful differences in customer motivation or behaviour? Can you actually do something different with them? Will treating those customers differently create value? AI doesn’t automatically know.
The same applies to testing. Ask AI for ten A/B test ideas and it can give you twenty before you’ve picked up your cup of coffee.
Subject line length. CTA wording. Button colour. Hero image. Emoji versus no emoji.
Congratulations. You have a testing programme. Except you don’t.
A meaningful testing programme starts with understanding what you’re trying to learn. Are customers primarily motivated by saving time or saving money? Does reassurance matter more than urgency? Does demonstrating ease reduce a perceived barrier?
Those are hypotheses. Once we’ve decided what we need to learn, AI can help us create very different executions against those hypotheses.
Again, the marketer does the thinking. AI helps scale the execution.
AI makes a good framework more valuable
This is the part of AI I find genuinely exciting. Historically, some excellent strategic ideas were difficult to execute because they simply weren’t scalable.
You might know that manually reviewing every customer, product, piece of content or interaction would produce better marketing. But nobody has the time or budget to do that across a database of hundreds of thousands of people.
So we simplified. We created broad segments, built rules, used proxies and accepted “good enough” because the alternative was operationally impossible. AI changes some of that.
A framework that previously could only be applied manually to 50 customers may now potentially be applied to 50,000. A human can develop the classification, and AI can classify. A human can identify the questions customers need answered, and AI can help identify where those questions apply. A human can define meaningful motivations, and AI can help create executions against them.
AI doesn’t make the thinking less important. It makes good thinking more scalable.
But first, prove that AI can actually do it
Once we’d established what information would be useful in the founder’s outreach, I didn’t suggest blindly automating the whole thing. We needed to find out what AI could obtain reliably.
Could it count the number of videos someone had published? Probably relatively straightforward. Could it calculate useful engagement information? Potentially. Could it accurately classify the style of video using the founder’s framework? That needed testing.
Because there’s little value in scaling an insight if the insight becomes unreliable at scale.
This is another reason doing the work manually first matters. You have something to compare the AI against.
If you’ve personally classified a sample, you can test whether AI reaches the same conclusions. You can find the ambiguous cases, refine the criteria and decide which tasks are safe to automate and which still require human judgement.
Without that groundwork, how do you know whether AI is right?
The danger of plausible strategy
This, for me, is one of the biggest risks of using AI in marketing.
AI is exceptionally good at producing things that look like strategy: customer personas, segments, testing plans, content strategies, journey recommendations and personalisation ideas.
All neatly structured, convincingly explained and ready to paste into a PowerPoint. The problem isn’t necessarily that they’re wrong. It’s that they’re plausible.
And plausible is dangerous because it doesn’t trigger our scepticism in the way an obviously bad answer does.
A generic segmentation can sound sensible. A weak hypothesis can sound intelligent. A completely invented customer motivation can sound psychologically convincing.
If we haven’t done enough thinking ourselves to recognise the difference, AI hasn’t accelerated our strategy.
It’s replaced it.
The marketer’s job is moving upstream
As AI takes on more execution, I think the marketer’s value increasingly moves earlier in the process.
Away from asking, “Can I produce this?” and towards asking:
- What problem are we solving?
- What do we need to understand?
- What matters and what doesn’t?
- What framework should we use?
- What hypothesis are we testing?
- What does good look like?
- How will we know whether the AI output is correct?
These aren’t production questions. They’re judgement questions. And AI’s ability to produce more, faster, makes them more important, not less.
Because a poor framework applied manually has limited consequences. A poor framework automated across a million customers has considerably more.
Don’t outsource the bit that makes AI useful
There’s a temptation to look at AI and ask: “What work can this save me from doing?” And, it’s a perfectly reasonable question.
But perhaps there’s a better one: Which work do I need to do once, really well, so AI can scale it?
That’s what happened with the founder. Watching those first 30 videos wasn’t wasted manual effort that AI could have saved him from. It was the work that made automation possible.
He was learning the problem, identifying patterns, developing the framework and deciding what mattered. Once that thinking exists, AI can potentially take it somewhere a human never could.
That’s the division of labour I think marketers should be aiming for: You create the framework. AI applies it. You validate the output. AI scales it.
So yes, use AI to save time, analyse more, create more and automate work that previously couldn’t be automated. But don’t be too quick to automate away the part where you learn what you’re doing.
Because AI can scale your thinking remarkably well. But first, you need to do some.
