What can go live without ahuman? In our operation,nothing that speaks for theclient

There is a question practically no marketing operation has answered in writing, and it is very simple: what can be published in a client's name without a human looking at it?

It is not a rhetorical question. It is already being answered every day, in every agency and every in-house team that adopted AI. It just isn't being answered by decision — it is being answered by whatever fits in the week.

And what that looks like in practice is embarrassing. A generated caption scheduled straight away. A campaign email with automatic variations nobody read in full. Automatic replies to comments on a client's post. A video cut and published on the profile of someone who trusted your service.

None of those things is wrong in itself. Some are excellent productivity decisions. The problem is that they are decisions nobody made consciously.

Whoever works with someone else's brand is managing a borrowed asset. It is different from making a mistake on your own feed.

Here is the method behind this series: we mapped public AI cinema and creative culture blogs to understand how people who produce images on commission dealt with AI entering client work.

Declared bias: it is audiovisual, largely outside Brazil, and with formal contracts. It is not the reality of a four-person agency in the interior of Rio Grande do Sul.

But there is a finding worth gold for anyone who provides services. In that circuit, the conversation about AI in commercial work is rarely about quality. It is about disclosure and consent: does the client know there is generation in the process? Did they agree? Where is it written?

That is mature in a way the small-business marketing market still isn't. Their question isn't “can I use it,” it is “does the client know what they are buying.”

It is the question our market is avoiding — and avoiding it isn't neutral. Avoiding it means the client will find out on their own, at the worst possible moment.

Why the risk is different when the brand belongs to someone else.

On your own profile, a wrong piece costs you. You decide whether to own it, fix it, ignore it. It is your loss and your choice.

On a client's profile, the wrong piece creates three problems at once. Their customer sees the wrong information. The client questions you. And you have to explain why something was published without anyone having seen it.

That third part is the one that ends contracts. It isn't the mistake — it is the discovery that nobody was in the way. The client hired you precisely to have someone in the way.

And there is the aggravating factor of local knowledge. The machine doesn't know that this client doesn't open on Saturdays, that the product was discontinued, that their competitor is the godfather of the owner's son. That knowledge isn't in any prompt, and it is exactly what separates a publishable piece from one that makes the phone ring.

It is the direct application of the argument in human command in the age of AI, with the bill being paid by a third party.

I find it more useful to work with three tiers than with the question “do you use AI or not.”

The first tier is internal material. Drafts, variations, meeting summaries, idea gathering, first versions of text. Here AI can run freely and without ceremony. None of it reaches the public, mistakes are cheap and the gain is huge.

The second tier is low-risk published material. Behind-the-scenes posts, everyday scenes, generic tips, reuse of content already approved. It can be generated with AI, but it goes past a pair of eyes — a quick, two-minute look, in the final format.

The third tier is material that speaks for the client. Price, deadlines, warranty, results, availability, replies to complaints, anything that creates an obligation. That tier never goes live without human approval, and preferably with the client aware.

Three tiers. You can write them down in ten minutes and stick them on the wall, and that settles ninety percent of the decisions that are made on improvisation today.

Tier three is where all the money is, and it is the most ignored.

It is worth being specific about what goes in it, because in practice people underestimate its reach.

Price and payment terms. Delivery or service times. Warranty and return policy. Any claim of results — “lose weight,” “increase your sales,” “guaranteed approval.” Stock availability. Opening hours. Customer names or testimonials. Replies to public criticism.

Everything on that list creates an expectation the client's operation will have to honor. And the machine has no way of knowing what can be honored — it fills in with what is common in the market, which is almost always more aggressive than your client can deliver.

The typical result isn't a scandal. It is a counter with three people a week asking for something the business doesn't offer, and an irritated owner who can't understand where that came from.

That is the damage of when the generated image lies, with the difference that here it is born inside a paid service.

Now the part almost nobody does and that, in my experience, protects the most: telling the client.

I know it sounds risky. The feeling is that the client will think they are paying for machine work. I have heard people in the market defend silence as a commercial strategy.

In practice it is the opposite. Clients react well when the conversation is about method, not confession. The sentence that works is roughly this: AI does the first version and the volume, we do the direction, the review and the approval — and nothing that talks about price, deadlines or results goes live without you seeing it.

That does three things at once. It defines what you sell, which is judgment, not typing. It makes the client part of the control, which reduces your risk. And it defuses the bomb of the day they find out from someone else.

Transparency here isn't a moral virtue. It is risk management with a built-in commercial benefit — and it is practically free.

What happens when the policy doesn't exist.

It exists anyway, only implicit and unstable. In a quiet week, someone reviews everything. In a deadline week, nothing gets reviewed. The operation swings between excessive rigor and no control at all, and nobody knows which version applies today.

It also weakens the team. A junior person with no written rule is left with a bad choice: publish and take the risk, or always ask and look slow. They usually publish, because calendar pressure is more immediate than fear of mistakes.

And when something goes wrong, the conversation becomes personal — “you shouldn't have posted that” — when what was missing was a rule, not common sense.

A written policy protects whoever executes. It is the most underestimated point in this whole discussion.

And it is part of the same apparatus as the pipeline is bigger than the prompt: without a definition of done and without risk tiers, every publication is a new decision made at the worst moment, which is the moment of rush.

An honest objection: doesn't this limit the scale gains AI offers?

It limits a piece of them, and it is worth facing that head on. If your business model depends on publishing hundreds of pieces without anyone looking, human approval crushes the margin. That's true.

But then the question is a different one: are you selling a marketing service or access to a conveyor belt? Because the client can buy the conveyor belt directly, cheaper, in the same app you use.

What an agency or an in-house team sells, from now on, is exactly what doesn't scale on its own: knowledge of the client's business, judgment about what to publish and the willingness to own what went live.

And the savings still exist, just not where the sales pitch promises. They are in producing ten times faster what you already approved — not in approving ten times less. That distinction is the subject of the real cost of AI in content.

I close with our rule, not as a universal model, but as an example of something written down.

At Uzz.Ai, no piece that speaks for the client goes live without a human approving it. Zero. Price, deadlines, results, availability and replies to criticism are all in that tier.

Internal material runs with AI without ceremony. Low-risk published material gets a quick review in the final format. And the client knows where AI comes in, because it is in the proposal, not in a hallway conversation.

It costs us a few minutes per piece. It has already prevented more uncomfortable calls than I could list.

What I would recommend to the reader: don't copy our rule. Write your own. It can be more permissive than mine — there are operations where that makes sense. What it can't be is nonexistent, because a rule that doesn't exist gets decided by Friday night, and Friday night decides badly.

A client's brand is a borrowed asset. Getting it wrong isn't a production error, it is a breach of care.

I don't know any serious marketing operation that will stop using AI. Nor should it — the tool handles well what used to be expensive and repetitive. What sets apart the ones that will thrive is far less exciting: they will have written down where the human comes in.

That line is cheap to draw and expensive not to have. Ten minutes of conversation are worth more than any subscription upgrade this year.

And it is good to remember why the client hires you: not for access to the tool, which they already have. For someone who knows their business and owns what goes live. Automating precisely that part is selling exactly what wasn't bought.

AI accelerates, you lead. In services for third parties the phrase gains an important comma: you lead — and you answer for it.

This series continues in a new model is not a strategy and in when the generated image lies. But if you do only one thing after this text, write your tier 3. Today, on one sheet, with the names of who approves.