For the last few years, one question has dominated almost every conversation about AI in marketing:
What can we automate?
Can AI write the content?
Can it analyse the campaign?
Can it qualify the lead?
Can it build the report?
Can it optimise the ads?
Can it respond to the customer?
They're reasonable questions.
I'm just not sure they were the most important ones.
Because AI didn't invent automation. Marketing teams have been automating parts of their work for years: email sequences, bidding strategies, CRM updates, audience creation, reporting, lead routing, campaign triggers and plenty more.
What AI changed was the perceived limit of automation ("perceived" being the keyword here).
Suddenly, we weren't just automating predictable tasks.
We could automate work that previously required interpretation, judgement or the ability to deal with ambiguity.
And that opened up a much bigger conversation and changed things in the way we thought about more complex tasks almost overnight!
Unfortunately, I think we became so fascinated by what the technology could do that we skipped over something more fundamental:
What needs to exist around the automation for it to work reliably in the first place?
Because AI isn't an operating model.
It's one component inside one.
We made automation look like a technology problem
The appeal of modern AI is understandable.
Prompt something.
Get an answer.
Connect it to another platform.
Give it access to some data.
Add a trigger.
And suddenly, something that used to take a person 30 minutes happens in seconds.
The latest generation of AI takes that even further. The conversation is increasingly moving from assistants that respond to individual prompts towards agents and workflow systems that can plan, use tools and execute multiple steps with less human involvement. Google and Adobe are both describing this transition in terms of end-to-end workflows rather than isolated AI tasks.
That is an enormous opportunity.
But it also creates a temptation to think about automation from the visible end backwards.
We see the output and ask:
How do we automate that?
Perhaps a better question is:
What system would need to exist for that output to happen reliably every time?
Those two questions lead you to very different places.
Automation isn't a tool. It's an operating model.
Let's say a marketing team wants to automate lead follow-up.
On the surface, the AI part is fairly easy to imagine.
A new lead arrives.
AI reads the information.
It researches the company.
It decides how relevant the opportunity looks.
It drafts a personalised message.
It sends it.
Great.
But then you start asking questions.
Where does the lead information come from?
Which source is considered accurate if two systems disagree?
How do we know whether the person is already speaking to sales?
What makes a lead qualified?
Who determines the messaging rules?
What happens if important information is missing?
When is AI allowed to send something automatically and when should a human approve it?
What happens when the prospect replies?
Where is that response recorded?
What should happen next?
And how does the system eventually learn whether any of this resulted in a meeting, an opportunity or revenue?
At some point, you realise that the AI writing the email is only a small part of the job.
The real work is designing everything around it.
And this is the point where actual work begins.
The intelligence is only one layer
A useful automation system needs several things to work together.
- Context and data — What does the system know, and can that information be trusted?
- Rules and objectives — What are we actually trying to achieve, and what constraints apply?
- Orchestration — Which systems need to communicate, in what order, and what triggers each step?
- Intelligence — Where does AI need to interpret, generate, prioritise or make a recommendation?
- Controls and ownership — What can happen automatically, where should a human intervene, and who owns a failure?
- Feedback — How does the system know whether the action it took produced a good outcome?
AI might be the most impressive part of that chain.
But it isn't the chain.
And if the other layers are weak, adding more intelligence doesn't necessarily make the system better.
Sometimes it just allows a poorly designed process to operate faster.
This matters more as AI becomes more autonomous
There is a big difference between AI recommending an action and AI taking one.
If an AI assistant produces a weak campaign analysis, someone can read it and decide not to use it.
If an automated system interprets the same information and reallocates budget across 200 campaigns, the consequences are different.
If ChatGPT drafts a bad email, somebody can rewrite it.
If an automated workflow sends that email to 10,000 customers, the problem has already happened.
If an analyst misunderstands a dataset, the mistake may affect one presentation.
If the same misunderstanding is embedded into an automated decision-making workflow, it can be repeated every hour.
The more autonomy we give software, the more important the system around it becomes.
That is why I suspect AI won't reduce the need for systems thinking.
It will increase it.
Even the companies building agentic marketing products are running into this. Adobe reported this year that 75% of organisations surveyed cited data integration and data quality as their biggest challenge to implementing agentic AI.
The technology is moving quickly.
The surrounding organisation still has to catch up.
Marketing is full of workflows that look simple until you map them
Take campaign optimisation.
The obvious AI use case is attractive:
Analyse the advertising data, identify what is underperforming and recommend what should change.
Potentially, let the system make the change itself.
But what is it optimising for?
ROAS?
CPC?
Lead quality?
Revenue?
Profit?
Pipeline?
What happens if Google Ads says one thing while your CRM says another?
Should a campaign that looks inefficient today be switched off if the sales cycle is 90 days?
How much historical information should the system consider?
Can it increase spend without approval?
What constitutes an anomaly rather than normal volatility?
How do you know whether the action it recommended last week actually improved anything?
The intelligence required to produce the recommendation is only one problem.
The system required to understand when that recommendation is useful is much larger.
And this applies well beyond campaign optimisation.
Content automation needs brand context, approval processes, distribution logic and performance feedback.
Customer lifecycle automation needs reliable CRM data, consent management, segmentation rules and clear handoffs between marketing and sales.
Reporting automation needs agreed definitions, dependable data pipelines and rules for how information should be interpreted.
AI can participate in all of those workflows.
It cannot magically provide the organisational structure around them.
The danger is automating the mess
There is another problem here that I think we'll hear a lot more about over the next few years.
Most businesses aren't starting from perfectly designed operations.
They're starting from years of processes layered on top of one another.
A spreadsheet someone created in 2022.
A Zapier workflow nobody wants to touch because they're not entirely sure what will break.
A CRM property that means one thing to marketing and another to sales.
An automated email sequence created by someone who left the company eight months ago.
Three dashboards with slightly different definitions of conversion.
A Slack notification everybody ignores.
Now add AI.
The temptation is to automate across all of it.
And initially that can feel like progress because people are doing less manually.
But automating a bad process does not turn it into a good process.
It can make the underlying problem harder to see.
You end up with more integrations, more dependencies and more decisions happening automatically, while fewer people understand how the whole thing actually works.
This is where automation debt starts to look a lot like technical debt.
Every automation has a maintenance cost.
Every integration introduces a dependency.
Every AI decision needs context.
Every exception needs somewhere to go.
Every workflow needs an owner.
Five isolated automations might save time.
Fifty interconnected automations can become an operating system nobody fully understands.
The most mature automation might actually look less impressive
I think we sometimes judge automation maturity using the wrong measure.
We look at the number of tasks being automated.
Or how autonomous the AI is.
Or how many agents a business has deployed.
But perhaps a better measure is:
How little fragile human glue is required to keep the system functioning?
A mature system isn't necessarily the one where humans disappear.
It's the one where their involvement is deliberate.
Humans handle the exceptions because those exceptions genuinely require judgement, not because the workflow broke.
Approvals exist where risk warrants them, not because nobody trusts the technology.
Data is entered once rather than repeatedly reconciled between systems.
Ownership is clear.
Feedback loops exist.
And when something changes, somebody understands the consequences for everything downstream.
That's less exciting than launching another AI agent.
It is also much closer to what reliable automation looks like.
AI changes the role of the marketer too
There is another implication here.
If more execution becomes automated, some of the most valuable marketing skills move one level higher.
The question becomes less:
Can you perform this task?
And more:
Can you design the system that performs this task well?
That requires different thinking.
Understanding objectives.
Defining inputs.
Designing handoffs.
Recognising dependencies.
Creating guardrails.
Choosing where automation should stop.
Evaluating output.
Handling exceptions.
Connecting what happens in one platform to the commercial outcome somewhere else.
Google has recently described this shift as marketers moving from individual prompting towards managing AI-driven workflow systems, with humans increasingly defining tasks, inputs and guardrails before evaluating the outcome.
I think that transition matters far beyond whichever generation of AI happens to be popular right now.
The tools will change.
The requirement to design how work happens won't.
The goal was never automation for its own sake
There is a tendency with every major technology shift to mistake adoption for progress.
We did it with digital transformation.
We did it with big data.
We did it with SaaS.
And we're doing some version of it with AI.
How much AI are we using?
How many processes have we automated?
How many agents have we deployed?
Those numbers aren't necessarily evidence of a better marketing operation.
A team with five well-designed automated workflows may be considerably more effective than one with 100 disconnected ones.
The goal isn't to remove as many humans from marketing as possible.
The goal isn't even to automate as much marketing as possible.
The goal is to build a marketing operation that can move faster, make better decisions and produce better outcomes without becoming more fragile as it scales.
AI can be an extraordinary part of that.
But only when the system surrounding it deserves the intelligence we're putting inside it.
For the last few years, we've spent an enormous amount of time asking what AI can automate.
Maybe the more valuable question for the next few years is simpler:
What kind of system are we actually trying to build?