The dangers of AI in business: what is real and what is exaggerated

When we talk to business owners in Mallorca about automation, the conversation almost always reaches the same point: the dangers of AI in business. It is a reasonable concern. People read headlines about leaked data, about systems that make up answers and about jobs disappearing, and nobody wants to bring that into a family hotel in Alcúdia or a clinic in Palma. In this article we go through each danger honestly: how likely it is in an SME, what real harm it can do and the specific rule that keeps it under control. The conclusion, in advance: for a normal business, the risks exist but are manageable with a few common-sense rules.

Hallucinations: the model makes things up

This is the most real danger on the list. A language model can state with complete confidence that your hotel accepts pets when it does not, or quote a price that does not exist. It happens less than it did two years ago, but it happens.

In an SME the right question is where it can do harm. If the agent drafts a reply to a customer and a person reviews it before sending, a hallucination costs ten seconds of correction. If the agent replies on its own with no reference data, the error reaches the customer.

The rule that neutralises it:

• The agent answers from your documents (rates, terms, opening hours), not from its general memory.
• Everything that goes out is subject to a human approval step, at least during the first few months.
• When it cannot find the information, the agent is instructed to say "I will check and confirm", rather than filling the gap.

With those three rules, a hallucination goes from being a risk to being a typo that someone corrects.

Data leaks: where your customers' data ends up

Here it helps to separate two things. One is an employee pasting a customer list into a free consumer tool whose terms allow that text to be used to train models. That happens, and it happens without anyone having decided it. The other is a professional, sensibly configured system leaking data, which is considerably less likely.

What we do in our implementations:

• Data is processed and stored on servers within the European Union.
• Business accounts with a data processing agreement are used, never personal accounts.
• The agent only has access to what it needs for its task: if it handles bookings, it does not see payroll.
• A one-page document states which tools employees may use and with which data.

The risk of a leak in an SME almost always comes from informal use of tools rather than from professional automation. Putting that informal use in order is, in practice, the first benefit of adopting AI seriously.

Dependency: what happens if the provider disappears

This is a legitimate concern and a fairly exaggerated one at the same time. Legitimate because some businesses have built on a closed platform and then seen the price rise or the service shut down. Exaggerated because today there are open-source agent platforms (we work with Hermes Agent and OpenClaw) that run on several different models: Claude, GPT, Gemini or Grok. If a model provider changes its terms, you switch model and the rest of the system carries on unchanged.

The rule:

• An open-source platform that your business can keep using even if you change service provider.
• Documentation of every automation in plain language, kept at your business.
• No critical process depends solely on the agent: if the agent goes down, the team can do the task by hand as before, even if it takes longer.

Errors at scale: one mistake repeated a hundred times

A human error affects one customer. An error in a badly configured automation can affect every customer that morning. This danger is real, and it is the reason we never switch anything on at full scale on day one.

How it is managed:

• Draft mode first. The agent prepares, a person approves. It moves to automatic mode only for tasks where it has not failed for weeks.
• Volume limits. An agent that sends appointment reminders has a daily cap. If it exceeds it, it stops and raises an alert.
• A log of everything it does. Every action is recorded, so if something goes wrong you know exactly what was sent and to whom.
• Alerts when something is out of the ordinary: zero emails processed on a Monday, or fifty in an hour, are signs that something is wrong.

With these controls, the scenario of "a hundred customers received a wrong message" becomes "the agent stopped at five and raised the alarm".

The team's fear: "this is here to replace me"

This fear holds back more implementations than any technical failure. If staff believe the automation is the step before a redundancy, the implementation sabotages itself: nobody reports errors, nobody suggests improvements, nobody uses it.

What works in family businesses on the island:

• Explaining from the outset which tasks are being automated (the repetitive ones, the copying of data from one place to another) and which stay in people's hands (dealing with the customer, decisions, exceptions).
• Starting with the task the team hates most. When the agent takes ticket filing or quote follow-up off their plate, the attitude changes.
• Giving one member of the team the role of agent supervisor. Approving, correcting and proposing improvements is visible, valued work.

In high season, when hands are short, the team is usually the first to ask for the agent to take on more.

Frequently asked questions

Can an AI agent make decisions without anyone knowing?

Only if it is configured that way, and we do not do that. A well-implemented agent has limited permissions, a log of every action and approval steps for anything that affects customers or money. You decide what it can do alone and what needs your sign-off.

Is it dangerous to use free AI tools in the business?

The risk lies in the terms of use and in the data that gets pasted into them. For tests without customer data there is no problem. For real work with personal data you want a business account with a data processing agreement and servers in the EU, plus a clear internal rule about what may be uploaded.

What happens if the agent makes a mistake with a customer?

The same as when an employee makes one: it is corrected, an apology is made and the process is adjusted. The difference is that the agent leaves a complete record of what it did, so finding the cause takes minutes. In draft mode, moreover, the error is caught before it goes out.

If you want to know which risks genuinely apply to your business and which rules would be needed to automate with peace of mind, the assessment answers that with data from your own business. Request an assessment