The Five Decisions AI Can Make in Your Marketing Today, and the One It Should Never Make Alone.
Most writing about AI for marketing argues at the wrong altitude. Should AI decide things, or only assist? Should a person review everything, or nothing? Framed that way the argument never ends, because a marketing week is not one decision. It is a chain of specific ones, made in order, each with a different cost of being wrong.
We run our own marketing as a governed system and have done since September, so we can answer the narrower and more useful question from experience: which links in the chain can a system hold on its own today, and which one must a person hold no matter how good the model gets. The answer is five and one.
Decision one: what to measure, every day.
The first decision in any marketing week is what the numbers say, and in most companies it is not made at all. Someone pulls the ad platform, someone else screenshots the analytics, the CRM is asked at month end, and the three never meet. The decision that gets made is whichever chart the loudest person brought.
A system can make this decision completely. Ours reads the ad platform, the analytics property, the search console, the search-rank data and the CRM every morning into one table, keyed so a dollar of spend can be followed to a session, a lead and a qualified lead. The decision "what happened yesterday" is then a read, and it is the same read for everyone. This is the least glamorous decision on the list and the one that makes the other four possible.
Decision two: what to say next.
Given the month's themes, the programs running under them, what has already been published and what the market is searching for, which pieces should be produced this week? This is a judgment, and it is a judgment a system makes better than a busy human because it can hold the whole corpus, the whole keyword set and the whole plan at once, and a person cannot.
Ours proposes twenty blog topics and ten video scripts a week against the plan, dedupes each against everything already published by argument rather than by title, and states for each one what it will say that the pieces already ranking do not. A person still approves each piece. But the decision "what is worth saying" is made by the system, with its reasoning written down, and the person's job is to disagree, not to originate.
Decision three: how to say it.
Drafting, imagery, the search block, the internal links, the related-content block, the review copy. A system makes all of these decisions in production, within a voice it has been given and rules it cannot break: no invented figures, no invented client, no claim about a competitor that their own site does not support, no article schema that fights the platform's own. The voice rules are a document, and a draft that violates one does not reach the reviewer.
Where people usually go wrong here is in imagining the alternative is a human writing from scratch. The alternative in most mid-market companies is nothing: the post that never got written because the one marketer was in meetings. Against that baseline, a system that produces a reviewed draft with its evidence attached is not a compromise.
Decision four: which lead is worth a person's time.
Scoring a lead has been an AI decision for years, and most companies have it half switched on. The system can decide the order in which a person should look at the inbound queue, from the form, the pages visited, the company, the search term that brought them and the history of what converted. What it must not do is decide the lead is qualified. In our model a lead becomes qualified when a person reviews it and sets a status in the CRM, and that status is the only event the system counts as a conversion. Scored by the system, qualified by a person is the line, and it is a line rather than a preference because the qualified count is what the ad platform is trained on.
Decision five: when to stop.
Every program in our plan has a kill rule written before it starts: what it must achieve, by when, and what happens if it does not. The system checks the rule daily against the table from decision one. When a campaign trips it, the system says so, with the numbers, the same morning, and proposes the stop. This is the decision companies are worst at making themselves, because stopping something is a confession and nobody volunteers one. A system does not mind confessing. It holds the rule that was agreed when everyone was calm and applies it when nobody is.
The system proposes the stop. It does not execute it, which brings us to the one.
The one decision that stays with a person: spending or publishing.
Every dollar and every public action stops at a person. Approve, hold or change, with a reason, in a thread that becomes the record. The caps are enforced in code below the gate, so an approval cannot exceed what the plan allowed, and the system cannot raise its own ceiling. Nothing runs on a credential a person can see.
We hold this line for three reasons, and none of them is that the model is not good enough.
The consequences are asymmetric. A wrong topic costs a draft. A wrong spend costs money and a wrong publish costs trust, and neither is fully reversible. The gate goes where the cost of error is.
The person is what makes the system accountable. An approval with a name on it is a decision someone can be asked about. A budget an agent spent on its own judgment is an outage report. We have written about why nobody should let an agent near a budget without a gate, and the argument is about accountability rather than capability.
The gate is how the system learns. Every hold and every change is feedback, recorded against the piece, and the weekly review turns it into rules. Remove the gate and the system stops being corrected by the one person whose judgment it is supposed to encode.
What this looks like from the person's chair.
The person in this model is not a marketer doing the work between meetings. They are the approver. In a normal week they read one consolidated explanation, reply "read", and then answer decisions one at a time, each in its own thread, with a 1, a 2 or a 3 and, if 3, what to change. Four decisions open at once, no more. The system does the rest and shows its evidence. The six-minute walkthrough shows a week of it.
The honest cost of that chair is attention, and the honest benefit is that the attention is spent only where money or reputation moves.
Where the line will move, and where it will not.
Decisions two through five will get better as the model and the data get better, and the person will disagree with them less often. That is expected and welcome. Decision one is already a read. The one decision will not move, not because AI cannot be trusted with money but because a system whose spend nobody signed is a system nobody is accountable for, and accountability is the product we are actually selling.
If you are evaluating AI for marketing, ask the vendor which of the six decisions their system makes, which it proposes, and where the gate is enforced in code rather than in policy. The vendors who can answer that in a sentence have built a system. The ones who answer with a demo of a chat window have built a feature.
See the AI Marketing OS run a marketing function before you commit to it.
Watch the 6-minute guided demo, then book a walkthrough: we read your CRM, ad accounts and content and show what a governed system would do differently. You pay only if you proceed. Or see how we approach it.
More on the same problem:
Comments