Automate Your Output, Not Your Judgment

Almost every brand I coach opens with the same wish: can we just hand all of it to AI, the comments, the posts, the customer service, running automatically while we sleep? Before I answer, I usually ask them to picture something.

Picture this. A week from now, your Instagram inbox has three hundred replies sitting in it. Not one has been read by a human. The AI handled all of them, smooth tone, clean formatting, fast. And one of those replies says something your legal team would never, ever sign off on.

The AI didn't know. It just predicted whatever sounded most like customer service and shipped it.

And now that sentence is sitting under your brand's name. Screenshotted.

AI Makes You Faster, Including At Making Mistakes

Let me be clear up front: I'm not here to talk you out of using AI. I use it every day, and I teach companies how to use it.

The question was never whether to use it. Here's the part fewer people put on the sales page: the moment you hand the speed of your output to AI, you hand over the speed of your mistakes too.

A community manager who answers fifty comments a day needs a full day to make fifty mistakes. AI clears five hundred overnight. If it's wrong, it's wrong five hundred times, in your brand's voice, while you're asleep.

That's the side of full automation nobody markets. It's seductive because it kills the "slow" problem. It does nothing for the "wrong" problem. It just makes wrong faster.

When The Bot Talks, You're The One On The Hook

Air Canada found this out the hard way. Their website chatbot gave a grieving customer the wrong policy on bereavement fares. The customer followed the advice, couldn't get his refund, and took them to court.

Air Canada's defense was almost impressive: they argued the chatbot was a "separate legal entity," responsible for its own words.

The tribunal wasn't having it. The ruling was blunt. A company is responsible for everything on its website, whether the words come from a static page or a chatbot.

Translation: AI is not your liability shield. Legally, whatever it says, you said.

Then there's DPD. The UK courier's support bot had been running fine for years, until one system update sent it off the rails. It started swearing and writing poetry about how useless its own company was. A bored customer screenshotted the whole thing, and it pulled in 1.3 million views.

Two things worth noticing. It wasn't a bad bot. It had worked for years. And it didn't take a sophisticated attack to break it. Just one bored customer, and one update nobody had stress tested.

Could your brand absorb a stunt like that?

Even Klarna Walked It Back

If you think this is just small companies fumbling, look at Klarna.

The Swedish fintech was the poster child for "AI replaces customer service." They used an assistant trained with OpenAI to replace around 700 agents, and at one point it handled two thirds of all support queries. On paper, a clean win.

By 2025, they were hiring humans back. The CEO admitted they'd leaned too hard into efficiency and cost, and quality, along with customer trust, slipped with it.

But here's the detail that matters more than the mea culpa: they didn't rip the AI out. They rebalanced. High volume, repetitive stuff stayed with AI. The complex cases, the money cases, the ones where a wrong answer actually costs you: humans went back in.

This isn't an "AI failed" story. It's a "they finally figured out where to draw the loop" story.

Some Things AI Doesn't Know It Can't Say

Now the part where I watch Taiwanese brands get burned the most, especially in beauty, skincare, and supplements.

In Taiwan, you can't make false or exaggerated advertising claims for cosmetics, and you absolutely can't claim medical efficacy. That's not a guideline; it's the law. Exaggerated claims run roughly NT$40,000 to 200,000. Cross into medical efficacy territory and it jumps to NT$600,000 to 5 million. Tie a supplement to medical claims and the drug act can take it as high as NT$25 million.

The thing that trips people up most is the specific wording. Phrases that sound harmless, like the local equivalents of "activates hair follicles," "firms," "slims," "medical grade," or "repairs damaged skin," are explicitly flagged by regulators as exaggerated or medical claims.

Now bring AI into it. A haircare brand lets AI reply to comments automatically. Someone asks, "I've been shedding a lot lately, does this work?" Without missing a beat, the AI is very likely to say something like: "Yes! This helps activate your hair follicles and reduce hair loss!"

It sounds caring. It sounds exactly like what a good rep would say.

Except "activates hair follicles" is, almost word for word, on the prohibited list. And by Air Canada logic, that fine doesn't go to the AI. It goes to you.

The AI isn't malicious. It just doesn't know Taiwan has that rule, and it doesn't know regulators judge claims holistically, weighing your text, your image, and your video all together. That kind of judgment is hard to wrap up inside one prompt or one skill.

Which is exactly why some checkpoints keep a human.

The Efficiency You're Chasing Might Be Diluting Your Trust

Step back, and even when AI breaks no laws and makes no errors, full automation carries a quieter cost.

More than 60% of marketers already use AI to produce social posts. So "we use AI for content" isn't an edge anymore. It's table stakes. When everyone automates, your automation stops being special.

And customers can smell it. Surveys put it near half of people who feel content made by AI lacks authenticity, and around 90% who simply want brands to disclose when they've used AI. Researchers even have a name for it: the "trust penalty." The mere sense that a machine wrote it drags trust and engagement down.

Flip it around: AI content a human has actually reviewed and shaped outperforms the fully automated stuff several times over.

The human fingerprint isn't proof you're inefficient. In this market, it's proof someone actually cares. That slightly imperfect line, the inside joke only your community manager would make: that's the one thing separating you from the fifty other brands running the exact same automation playbook.

Instead Of Full Auto, Put A Human Right Before "Send"

So what do you actually do? My advice to companies is almost boringly simple.

First, don't let AI's output go straight out the door. Store it. SQL works. So does a plain Google Sheet or Excel: one column for the draft, one for status, one for who reviewed it. That's enough.

Second, the next day, have your community manager review them in a batch. Approve with one click, skip with one click. Or have the AI serve three versions and let the human pick one and paste it. AI does the exhausting part, zero to draft. The human owns the last mile: the one glance before "send."

I know what a lot of owners say. We're slammed, there are too many comments, we can't keep up. That's why we need full auto.

So I ask back: are you really so busy you can't review even five hundred comments?

The comments that actually need a reply are rarely five hundred. And even when they are, the bottleneck was never volume. A batch approval queue clears in half an hour. What's missing is the "a human saw this" step. The time you save by deleting that step is nothing next to the cleanup one wrong reply will cost you.

Final Thought

My take on AI has always been simple: automate what you can.

But "can be automated" and "should be automated" are two different things.

Let AI draft the reply, then let a human glance at it before it sends. Let AI generate ten versions of the post, then let a human decide which one is actually you. It barely slows you down, and it buys you one chance to pull the brake before things go sideways.

That's the part I spend the most time on in my corporate sessions and workshops these days: not how to automate the whole pipeline, but how to put the human back in the right spot.

AI can automate your output. Just don't let it automate your judgment.