While X spent this week arguing about one thing, the cost of AI, the founder of an AI-automation company quietly told his own team the opposite of what he sells.

The week’s most-read thread was called “Token Budget Wars.” Inference cost, return per dollar of compute, marginal token utility. Everything orbited the invoice.

Then Max Brodeur-Urbas, founder of Gumloop, a platform whose entire product is running AI agents at scale, sent his team a memo telling them to stop using agents for everything. He called it slop. He wrote three rules to resist it and ended with a mantra: “Resist the slop, we win by staying human.”

When the person selling you the agents puts “stop” in writing for his own team, the why is worth reading.

Here’s the thesis. The expensive failure of agentic AI is not the token bill. It’s the slop you ship, which quietly erodes the customer and internal trust you spent years building. Knowing what not to automate is a judgment you make before you ship, not a problem your tooling solves. That decision is exactly what AI Maestro exists to make.

The tell: it came from the person selling the agents

Max’s first rule is the most useful one, and the most uncomfortable. If something takes less time to create than to read, it’s probably slop.

Picture the three-paragraph email an agent generated in two seconds that your customer spends a full minute reading, only to find it said nothing. The math doesn’t close. It cost the receiver more than it cost the sender. That’s slop in one sentence: volume that moves the work from the person who produced it to the person who has to consume it.

The craft, Max argues, lives in the opposite direction: in the brevity and clarity that take time to write and read fast. That’s the first thing an agent takes from you when you point it at everything without judgment.

And the detail almost nobody sat with: this is the owner of Gumloop saying it. A company that exists to run agents at volume. He isn’t an AI skeptic looking for a reason to be right. He sells the thing, and he just drew a line for his own team.

Where slop actually costs you

The other two rules point straight at the place where automation stops creating value and starts destroying it: the human relationship.

Customer support. Rule two is that support has to feel deeply human. A bot reply tells the customer “you weren’t worth a person’s time,” and that feeling is what later turns into churn. Automating support looks cheap on the quarterly spreadsheet. The real cost arrives six months out, when the customer who felt processed by a machine doesn’t renew and never tells you why.

Internal communication. Rule three is the counterintuitive one: having an agent reply to your coworkers is often worse than not replying at all. People want to talk to you, not to your automated stand-in. When you send an agent to answer in your place, the message that lands isn’t “I’m efficient.” It’s “you weren’t worth my time.”

There’s a third front Max doesn’t name, but every marketing team feels it: content. The blog post published because it was time to publish. The email that reads like a template. The piece that filled the calendar and moved nothing. Same slop, different surface.

The pattern is clean. Automation creates value when a machine consumes the output, or when volume genuinely helps. It destroys value when there’s a person on the other end deciding whether you care. Confusing those two cases is the costly mistake, and it never shows up on a token invoice.

Taste is the job now. That’s not enough.

There’s a fashionable answer to all of this, and it comes from the other side of the debate. Sam Altman said in February that having good taste might be the thing that gets you hired once AI can do the rest. Taste, judgment, the ability to tell good from barely-acceptable: that’s the scarce part.

I agree with the diagnosis. My problem is with what people conclude from it.

“Hire people with taste” sounds like an individual trait, and as an individual trait it doesn’t scale. In a thirty-person company you can’t trust that everyone, every day, makes the right call about what to publish and what to throw away. Slop doesn’t appear because the tool is bad. It appears because someone shipped without an explicit point in the process where a person decides this, not that.

So the operational question isn’t “do I have people with taste?” It’s “where in my process does someone deliberately decide what gets automated and what stays human?” If the answer is “nowhere, everyone improvises,” you already run a slop factory. You just haven’t measured it yet.

What IQ Source does about it

AI Maestro is, in one phrase, that gate. Two months of consulting where we map your operation’s real processes, not the org chart’s, and decide process by process which ones are worth automating and which aren’t, with an AI Opportunity Score and an explicit Go/No-Go gate at the end.

What sets us apart is that the gate says “No” often. For many processes, the correct answer is: don’t automate at all, or automate only the slice no customer ever touches. That sentence, “not this one,” is the one a vendor who bills you for the implementation will never say. We help clients resist slop, not mass-produce it.

Starbucks just paid the bill for skipping that gate. They rolled an AI inventory tool into 11,000 stores without testing it and retired it nine months later. It wasn’t a model problem. It was a judgment problem upstream: nobody decided, before scaling, whether the tool fit the actual task.

It’s the same bottleneck I covered a few days ago from the cybersecurity side: finding the work became trivial, deciding what to do with it is the scarce part. The judgment gate is the asset, not the agent.

Before the week ends, ask your team one question. Where, exactly, does someone decide today what we don’t automate? If the answer is silence, you don’t have an AI strategy. You have a slop factory waiting for a customer to notice it before you do.

Decide what not to automate with your team