Atiba de Souza.Start a conversation
WritingMarketing

AI Can Write the Ad. It Cannot Know Who You're For.

By Atiba de Souza

An editorial cover image for a piece on AI Can Write the Ad. It Cannot Know Who You're For. The cover reads: MARKETING AI Can Write the Ad. It Cannot Know Who You're For.

In July, my own team didn't know what my company was

I own marketing for TrueMD, a three-founder venture. In July my team was building out its LinkedIn presence and the assets that go with it.

Work started getting dropped. Then assets started appearing that made no sense to me. Not bad work. Work aimed at something that wasn't the company.

I traced it back. Most of the team had never been told what TrueMD actually is. Only a subset knew. I had assumed everyone did.

One of them said it plainly: I didn't know what TrueMD was.

That was my blunder, not theirs. I was carrying the definition in my head and treating it as shared air.

Now run that same failure through a tool. The tool has no head to carry it in. Nobody hears it say I didn't know.

The tool didn't fail. The aim did.

The conversation about AI right now is the same everywhere: is the output good enough. It was never the scarce thing.

Three founders built TrueMD. Most of the team marketing it could not tell you what it was. That is the gap AI inherits, and it inherits it at speed.

Which means the skill people keep trying to automate is the one that gets more valuable every month, not less.

Knowing who you're for: the ability to name the person whose decision your marketing has to move, what they think they're choosing between, and who in their world says no on the way to yes. It is a judgment you hold, not a fact you can paste into a prompt.

You can hand a machine the task. You cannot hand it the judgment. Most people find that out after the assets ship.

The fastest tell is which brands you copied

When marketers learn the craft, they collect reference brands. The instinct is to pick the famous ones. The clean site. The ad they personally liked. The brand that gets written up.

Then they feed those references into their own work, or straight into the tool, and out comes content with the right polish and the wrong details.

Here's the move: stop picking brands you admire. Pick the exact brands your ideal customer actually chooses.

Include the ones you think are ugly. Include the one with the clunky checkout. Include the competitor that keeps winning the deal you cannot figure out how to win. Those are the ones your buyer has already said yes to. Those are the details your buyer already recognizes.

Because the details are the tell. The specific thing you choose to put into a piece of content signals, to the right reader, whether you know them. Get the detail wrong and everything around it stops being credible. That is not a copywriting problem. It is a recognition problem.

Now let the tool pick the details. It was trained on more marketing than you or I will read in a lifetime, weighted toward whoever shouted loudest. So it hands you references from brands your customer has never bought from and never will, dressed in sentences that read beautifully. Beautiful to you, because you don't buy like your customer either.

Let me rank this honestly. Everything in this section is the small move. Ten minutes, no meeting, this week: pull your last three pieces of content, name the reference brand behind each specific detail, then ask whether your ideal customer actually buys that brand. If the answer is no, the detail comes out.

That's the minor fix. The one that compounds costs you a meeting.

The meeting you will want to skip

You cannot prompt your way to knowing who you're for. It is not a task. It is thinking. And thinking does not transfer through instruction. It transfers through reflection.

CASE Method is the structure for that reflection, and it is not a brief. It is a scheduled, structured, post-task reflection meeting. Not a way to assign work, not a way to instruct it. A meeting you hold after the work is done.

The mechanics, in order. You delegate the task. You stay completely hands off while they do it. They complete it. Then you run the meeting: question-led, four questions, and the entire job of those questions is to excavate how they think so you can see the distance between their thinking and yours, and bridge it.

Now the part nobody warns you about.

That meeting's purpose is to show you the gap. Which means it will show you that you never transferred the thinking. You handed over the task, they did the task, and somewhere in the middle the thinking never made the trip. You will be sitting in a room looking at your own failure to transfer, on a Tuesday afternoon, with a calendar full of things that feel more productive.

And your instinct in that room will be to correct. Do not. The moment you start correcting, it stops being a reflection and becomes a performance review. What you get after that is not honesty, it is performance, and you will not get the truth out of that person for the rest of the quarter. That is the rewiring: you have to sit on your own expertise for the length of the meeting and let them be wrong out loud. That is the fee.

Four things it is not:

  • It is not a forward set of instructions, and it is not an acronym you staple to the task.
  • It is not a status check.
  • It is not a list of corrections you read out loud.
  • It is not E means easier. E is Easier Than Expected, a strengths-surfacing question: where did this turn out easier than you thought it would. That answer is not small talk. It tells you where to point that person next.

The distinction that matters: traditional delegation transfers output. Do it like this. CASE transfers input, which means how to process the information, how to recognize the pattern, how to make the call the way you would make it. One gives you a team that pings you all day for exceptions. The other gives you a team that thinks like an owner.

So here is the decision rule, and it is strict. Run it only after the task is complete. Run it only when there is judgment in the task. Never as a ritual. Never as a status check. Never before the work is done. On a task with no judgment in it, you are holding a meeting to hear someone describe a checklist.

Write down who you are for

Delegate the task
then stay completely hands off

They do the work
and complete it

CASE Method meeting
scheduled, post-task, question-led

Excavate how they think
bridge the gap to how you think

Name the gap.
Do not correct.

E: where was it
Easier Than Expected?

Point that person's next
assignment from the answer

Not a status check.
Not before the work is done.
Not a set of instructions.

The point the AI surfaced that I hadn't considered

We built the first ideal customer profile for ContentCoach.Social. Then I went and worked it: three weeks, five cities, several hundred people, and we signed early clients.

Somewhere in the building of that profile, the AI surfaced a point I had never considered. Internal resistance to change from other executives. The buyer wants the change. The buyer's colleagues can block it. I had never thought about those colleagues.

Here is the part I did not expect. That suggestion was not noise. It was exactly what happened with a couple of the new clients we signed. The person we sold to was not the person who could stop it.

The model surfaced the objection. Three weeks, five cities, and several hundred people confirmed it. Only one of those two things can be delegated.

You would expect the machine's contribution to be the thing you quietly drop once reality shows up. This time reality confirmed it. But look at what did the confirming. Not the model. Field time.

AI is good at surfacing a possibility. It cannot confirm one. Confirmation costs field time.

And confirmation is not the end of the profile. It is the input to the next version of it. After you meet real customers, you go back and update the ideal customer profile. That is the decision, and it is not optional housekeeping: the profile is not a document you wrote once, it is a document field time keeps rewriting. A profile written once and never refined would have produced messaging that ignored the colleague who says no. All that speed, aimed at a room the buyer has to walk back into.

What the meeting would have caught at TrueMD

Go back to July. The first off-target asset lands. Before anyone builds the batch, I hold the CASE meeting. Four questions, question-led, me not correcting.

It would not have taken a forensic trace to find it. Someone would have said some version of I didn't know what TrueMD was while it was still one asset, instead of a stack of them.

That is the demonstration, and the fix it produces is not a better brief. It is a written definition and a confirmation loop. Write down what TrueMD is. Then confirm every person working on it has read it. Not the subset you happen to speak with. Every person, before they produce anything.

The failure was never that the team could not think. The definition lived in my head and I never checked whether it landed. The meeting is how you check.

What it looks like when it's running

InconsistentConsistent
What you hand overThe task, and a hope that context came with itThe task, plus a written answer to who we are for
Reference brandsWhatever is famous, or whatever the model defaults toThe exact brands your ideal customer actually buys
When the meeting happensNever, or before the work starts as a kickoffAfter the task is complete, as structured reflection
What you do when their thinking differsYou correct it, and it becomes a performance reviewYou name the gap out loud and leave it with them
When you find out the thinking didn't transferAt launch, or afterIn the meeting, before the next task
What AI volume doesMultiplies the missMultiplies what is working

The meeting costs you an hour and your pride

If you only do the ten-minute reference check, do it. It's the minor move and it helps.

If you can stand the expensive one, here is what it actually asks of you this week. Pick the last task your team completed that had judgment in it. Schedule the reflection after the fact, not the kickoff. Sit in the room, ask, and do not correct, even when the answer tells you something about yourself you would rather not hear. Before the next task goes out, write down who you're for and confirm every person who touches it has read it.

You will hear something uncomfortable in that meeting. That is the meeting working. The alternative is finding out at launch, from a batch of assets that look beautiful and are aimed at nobody.

The machine will keep getting better at the work. It will not get better at knowing who you're for. That is still your job, and it is still transferable. Just not by instruction.

<!-- cluster-links:begin -->

Related reading

<!-- cluster-links:end -->