Your AI Copilot Is a Bad Employee

I watched a friend who owns a small brewery last week. He was glowing. “I finally did it,” he said. “I got one of those AI things to handle all my customer service emails. I’m free.”

He pulled up his sent folder to show me. The third email down was a reply to a customer who’d had a bad experience with a bartender. The AI’s response began: “We sincerely apologize for the suboptimal service encounter you experienced.”

He hadn’t even noticed. He saw “sent” and thought “handled.”

This is the quiet disaster happening right now in small businesses everywhere. You’re told to get an AI copilot. You’re promised it will “take things off your plate.” So you point it at your inbox, give it a vague command like “answer customer emails,” and you walk away believing the work is done.

But you didn’t hire a new employee. You unleashed a bad one.

You wouldn’t let a new hire talk to customers unsupervised for a month. You’d train them. You’d shadow them. You’d review their first hundred interactions. Yet with AI, we do the opposite. We grant it immediate, unsupervised access to the most sensitive part of our business: our customer relationships. We treat it like a finished product, not a probationary worker that needs constant correction.

The problem isn’t that AI writes bad English. It writes technically perfect, deeply hollow English. It speaks in a voice that is no one’s. It defaults to corporate passive voice (“it is understood that,” “a decision was made”) when your business thrives on active, human directness (“I screwed up,” “Let’s fix this.”).

Before AI, an email from a pissed-off customer sat in your inbox until you had the emotional bandwidth to deal with it. Now, AI deals with it immediately… by pouring gasoline on the fire with a tone-deaf, boilerplate non-apology. The speed is the injury.

Let’s get specific. Where does your unsupervised AI go wrong?

It makes your small business sound like a Fortune 500 legal department. A simple question about your return policy gets a 200-word treatise on “terms and conditions” and “liability limitations.” You sell handmade pottery, not insurance. It lies with cheerful confidence. Asked if you can accommodate a gluten-free birthday cake, your AI, trained on generic data, might say “Yes, we offer a variety of gluten-free options!” when you have exactly one. Now you have a very disappointed, very sick customer arriving on Saturday. It has the emotional intelligence of a brick. “We are sorry for your loss,” it writes to a customer canceling a catering order because a family member died. Two sentences later: “Please note our 72-hour cancellation fee still applies.” It connected the data points but missed the humanity entirely. A human would have swallowed the fee and sent flowers.

The fix isn’t abandoning AI. The fix is managing it like you manage everything else in your business.

You must give it a persona. This is the single biggest step. Don’t let it write as “The Management.” Write a short bio for it. “You are Sarah, the friendly owner of a neighborhood bakery. You are direct, a little funny, and you always take responsibility. You never use words like ‘optimal,’ ‘utilize,’ or ‘please be advised.’” This takes five minutes and changes everything. You must audit its work, fiercely, at the start. For the first week, let it draft every reply, but send nothing without your eyes on it. Read each one out loud. Does it sound like you? Would you say this to a person standing in front of you? Correct it. Every time. You are training it on your voice. You must build guardrails, not just open gates. Most of these tools have settings. Use them. Tell it to never quote a price unless the number is in the original email. Tell it to never promise a specific menu item. Tell it to flag any email with the words “complaint,” “allergy,” or “bereavement” for your direct review. You are giving it a safety net.

The goal is not to make the AI perfect. The goal is to make it a useful draft horse, not a public-facing representative. Its job is to take the raw material of a customer email and shape it into a first draft that’s 80% there, saving you the time of starting from a blank page. Your job is to supply the final 20%: the tone, the nuance, the actual human judgment.

My brewery friend now has his AI set up differently. It drafts replies to simple stuff like “What are your hours?” and sends them. Anything with a hint of emotion or complexity gets drafted, then flagged for him. He spends less time typing, but more importantly, he spends zero time putting out fires started by a robotic, tone-deaf response.

So the action is simple. Today, before you close up, open your sent folder. Read the last five emails your AI “copilot” sent on your behalf. Out loud. If you wouldn’t have said it that way, you’re not managing a tool. You’re cleaning up after a bad hire. And it’s time for a performance review.