AI in marketing: judgment first, automation second

The conversation about AI in marketing has settled on the wrong question. It asks: how much can the machine do? The honest answer — a great deal, and more each quarter — turns out to be the least useful thing to know. It tells you the size of the engine. It tells you nothing about who should be driving.

The better question, and the one a serious operator asks, is narrower: which decisions should a human still make, and why. Answer that well and automation becomes an enormous asset. Answer it carelessly — by letting capability decide scope — and you have built a fast, tireless system for making the wrong calls at scale.

The discipline is not in how much you automate. It is in what you refuse to.

Begin with the case for the machine, because it is strong and worth stating plainly.

Automation is excellent at the continuous, the repetitive, and the tireless. Scanning a domain every day for issues. Discovering and scoring prospects. Detecting a reply and qualifying it. Sending a follow-up on schedule, every time, without fatigue or forgetfulness. Drafting a first version of a post, an article, a set of talking points. Holding a clean record of what happened. These are tasks where consistency beats inspiration and where human inconsistency is a liability, not a virtue. A person doing this work gets tired, distracted, and uneven. A system does not.

This is real leverage, and dismissing it is its own kind of unseriousness. The heavy lifting of a modern growth function — the volume, the vigilance, the drafting, the never-letting-a-lead-go-cold — should be automated. To insist on doing it by hand is not craft; it is waste.

Now the harder half, where most of the value and nearly all of the risk lives.

Automation is poor at the things that depend on context it does not have and consequences it cannot weigh. It does not know which opportunity matters most to this business this quarter. It cannot read a market, a relationship, or a moment. It will optimise the metric it was given with perfect indifference to whether that metric is the right one. And it cannot be held accountable — you cannot put a name to a result that no person chose.

So the decisions that should stay with a human are the ones where judgment, context, and accountability are the whole point: what to prioritise, what to say in the firm’s name, what a number actually means, when to act and when to wait. These are not heavy in volume. They are heavy in consequence. And consequence is exactly what a machine cannot carry.

The error is to let the machine’s fluency in the first category leak into the second. A system that can draft a thousand posts feels as though it should be trusted to publish them. It should not. Drafting is heavy lifting. Publishing in the firm’s name is a decision. The line between them is the whole discipline.

The right architecture follows directly. Automation does the work; a human approves the decisions that carry consequence. Concretely: the system drafts, the person approves. Nothing of consequence is published, sent, or committed automatically.

This is not a compromise or a safety blanket bolted on for nervous clients. It is the design. The draft is where the machine’s leverage lives — the speed, the volume, the tirelessness. The approval is where the human’s judgment lives — the context, the taste, the accountability. Put them in the wrong order and you lose the value of both: a human doing the drafting wastes the leverage, and a machine doing the approving forfeits the accountability.

Approval is not a rubber stamp, either. A real approval gate means a senior operator reads the draft, decides whether it serves the business, and owns the consequence of publishing it. The system can make that gate fast and well-informed — surfacing the opportunity with its problem, impact, and recommendation attached — but it cannot pass through it. The human does.

It is tempting to assume that as models improve, the case for human approval weakens. The opposite is true. Better automation widens the gap between what the machine *can* do and what it *should* decide, because it makes the wrong decisions just as fluently as the right ones — and far more persuasively.

A crude tool announces its limits; you do not over-trust it. A capable one hides them behind polish. As the drafting gets better, the temptation to skip the approval gets stronger, precisely when the cost of a confident error is highest. Judgment-first is not a stance for the current generation of tools. It is a stance for the better ones coming, where the discipline will matter more, not less.

This is also the honest answer to the hype. The breathless version of AI in marketing promises to remove the human from the loop. The serious version puts the human exactly where the human belongs — out of the heavy lifting, firmly on the decisions that matter. That is not a smaller role. It is a more concentrated one.

For an executive, the practical question is simple: in your growth function, who approves the decisions that carry your firm’s name and your capital, and is that a person or a default?

If the answer is a person, working through a system that does the heavy lifting and surfaces the choices clearly, you have the right model — leverage without abdication. If the answer is a default — published because the tool published it, sent because the sequence sent it — you have automated the part you should have owned, and the better the tool gets, the more efficiently it will do so.

The firms that get value from AI in marketing will not be the ones that automated the most. They will be the ones that automated the right things and kept human judgment on the rest. Automation does the heavy lifting. Judgment makes the decisions. That order is the entire discipline, and it does not change as the machines improve. It becomes more important.

When you are ready to put judgment back at the centre of how growth is run, that is the conversation to have.

AI should do the work, not make the decisions that carry consequence. Automation is excellent at the continuous, repetitive, tireless tasks — the heavy lifting — but the choices about what to prioritise, what to say in the firm’s name, and what a number actually means should stay with a human, because those depend on context and accountability a machine cannot carry.

Automation excels at the continuous, repetitive, and tireless: scanning for issues, discovering and scoring prospects, detecting and qualifying replies, sending follow-ups on schedule, drafting first versions, and holding a clean record. These are tasks where consistency beats inspiration and human inconsistency is a liability — the heavy lifting of a growth function that should be automated.

The opposite is true. Better automation widens the gap between what the machine can do and what it should decide, because it makes wrong decisions as fluently as right ones, and more persuasively. As the drafting improves, the temptation to skip approval grows precisely when the cost of a confident error is highest, so judgment matters more, not less.

Automation does the work and a human approves the decisions that carry consequence: the system drafts, the person approves, and nothing of consequence is published or committed automatically. A real approval gate means a senior operator reads the draft, decides whether it serves the business, and owns the consequence of publishing it.