AI agents: why listening to your critics saves your business

Every company has that person. In the meeting about the new AI project, they raise their hand and ask: "what if it goes wrong?". Often they leave the room labeled as difficult. Behind the times. Someone who "doesn't get technology". In this article I want to defend that person. When it comes to putting AI agents to work in your business, they may be the one protecting your money the most.

And the reason I'm writing this now is a piece of news that bothered me quite a bit.

What happened in the US and why it matters to you

Journalist Ken Klippenstein published a report showing that US federal agencies have started looking at AI critics as possible "foreign agents". In other words, questioning AI started being treated as suspicious. Almost as if doubting the technology meant you were working for the other side.

I won't get into the politics. That's not what this blog is about, and it's not why you're here.

What caught my attention is the pattern. When a technology becomes a strategic bet, whoever points out a problem starts being seen as an obstacle. It happens in government. And it happens, on a much smaller scale, inside your company too.

The owner invested. The manager sold the idea to the board. The vendor promised a 70% cut in support workload. Then someone from operations says the agent gave a customer the wrong answer. The natural reaction is to defend the project. The right reaction is to listen.

Why the critic is the best test for your AI agent

An AI agent is a system that doesn't just answer questions. It takes actions. It checks your inventory, books appointments, reissues invoices, replies to customers on WhatsApp. That's great. It's also exactly what makes mistakes more expensive.

A chatbot that gets an answer wrong leaves a customer confused. An agent that gets an action wrong leaves you with a canceled order, a discount that never existed or a meeting booked at the wrong time.

Who sees these mistakes first? Almost never the people who built the project. It's the people on the front line:

• The support rep who gets the complaint after the agent "solved" the case.
• The salesperson who notices the agent offered terms outside the price list.
• The finance person who spots three duplicate invoices in the same week.

These people aren't against AI. They're against the rework that landed on their desk. And they have information no metrics dashboard will show you.

Whoever complains about your AI agent is running, for free, the test you should be paying for.

An example where the skeptic saved the cash flow

Here's a typical case, with the details changed so no one gets exposed.

A clinic set up an agent to confirm appointments and reschedule them over WhatsApp. In the first two weeks, the numbers looked beautiful. More than 400 conversations resolved without anyone at the front desk touching their phone.

One receptionist, the longest-serving on the team, started complaining. She said "something was weird" with the schedule. Nobody paid much attention. After all, the numbers were great.

She was right. The agent moved the patient to the new slot, but in some cases it didn't free up the old one. The schedule looked full, but it had ghost gaps. By her count, about 6 appointments a week were lost. With an average ticket of R$ 250, that's R$ 1,500 a week. Around R$ 6,000 a month draining away with nobody noticing.

The dashboard said "success". The suspicious person said "something's wrong". The dashboard measured conversations. She measured reality.

The fix took one afternoon. The loss, if no one had listened, would have kept running for months.

How to turn criticism into a process, not a fight

The problem with loose criticism is that it turns into hallway gossip. "This bot is garbage" helps no one. What you want is to turn distrust into useful information. In practice, I recommend three simple things.

1. Create an official channel to report agent errors. It can be a WhatsApp group, a form, a spreadsheet. It doesn't matter. What matters is that people know where to write it down, and that it's seen as a contribution, not a complaint.

2. Ask for the concrete case, not the opinion. Instead of "the agent is bad", ask: which customer, what time, what they asked for, what the agent did. With three or four cases like that, whoever configures the system can find the pattern and fix it.

3. Review together, every week, for 20 minutes. In the first weeks after an agent goes live, sit down with the people on the front line and go over the week's errors. It costs little and saves a lot of headaches. Once things stabilize, you can space it out.

One detail that makes a difference: thank people in public when they find an error. It sounds silly. But if the first person who reports a problem gets treated as a nuisance, nobody else will report anything. And then you lose your best sensor.

The opposite mistake: listening only to fear

Now, to be fair, there's another side. Some people criticize AI because they're afraid of losing their job, because they don't like change or because they simply don't want to learn a new tool. That happens too.

My opinion, and it's a strong one: if you let fear decide, you stand still while your competitor automates whatever can be automated. Ignoring AI agents in 2026 is like ignoring WhatsApp Business in 2018. You can survive for a while. But it will cost you later.

The way out is to separate two things:

• Criticism with a concrete case: worth gold. Always investigate.
• Generic criticism, no example: worth a conversation. Understand what's behind it. It's often insecurity, and insecurity is solved with training and clarity about everyone's role.

A good sign you're on the right track is when the toughest critic on the team starts suggesting improvements for the agent. I've seen it happen a few times. The person who complained the most becomes the one who understands the system best. Because they paid attention when no one else was.

What to ask before putting an agent live

If you're thinking about using AI agents in your business, here are the questions I ask every client before turning anything on. They're the same ones your "resident skeptic" would ask.

• What happens if it makes a mistake? What's the realistic worst case? An annoyed customer or a financial loss?
• Who notices the mistake, and how fast? If the answer is "no one" or "only at the end of the month", there's a problem.
• Can it be undone? A sent message can't be taken back. An order can be canceled. Know which is which.
• Where does a human step in? Every agent needs a point where it stops and calls someone. A discount above X, a complaint with certain words, a customer who already asked to talk to a person.

If you can't answer these four with confidence, the agent isn't ready. No matter what the vendor said. And yes, I know I'm a vendor too. That's exactly why I ask these questions before you have to.

Distrust isn't the enemy, inattention is

The news from the US shows a dangerous reflex: treating people who question technology as a threat. At the scale of a government, that's a serious matter for others to debate. At the scale of your company, it's simpler. It's a management choice.

You can build an environment where pointing out an agent's error is seen as betraying the project. Or you can build one where it's seen as part of the job. In the first, the errors are still there, just hidden. In the second, they show up early, while they're still cheap.

AI will make mistakes. Every system does. The question is whether you'll find out in the first week or at the quarterly close.

This middle ground is exactly where I work. I help companies get automation and AI agents running, making it clear where the human steps in and how errors surface. If you want to understand how I approach these projects, take a look at who I am and how I work.

There is always someone in your company who asks "what if it goes wrong?". I stand up for that person.

A clinic put an AI agent on WhatsApp. The dashboard showed more than 400 resolved conversations. A success.

The longest-serving receptionist kept saying something was off with the schedule. She was right. The agent rescheduled patients but didn't free up the old slot. That was R$ 6,000 a month slipping away with nobody noticing.

The dashboard measured conversations. She measured reality.

Whoever complains about your AI agent is running, for free, the test you should be paying for. Create a channel for it, ask for the concrete case and thank them in public.

In my new article I explain how to turn that criticism into a process. The link is in the comments.

#AIAgents #ArtificialIntelligence #Automation #Management #SmallBusiness