Deciding 7 min read

AI in business: where it really helps and where it doesn't

I'm not going to show you everything AI can do. I'll show you where its limits are, according to the providers' own documentation, and how to handle them so you can use it sensibly.

AI in business: where it really helps and where it doesn't
Author: Maksym Kamenchuk, Nexflow — Banská Bystrica Published Updated

A language model helps where someone checks its output or where its answer can be grounded in your documents: working with text, sorting enquiries, or answering questions over company materials. It is a poor fit where its mistake would go unnoticed or cause irreversible damage.

The reason is simple. Anthropic's documentation states that even the most advanced models can sometimes generate factually incorrect text. That's why I recommend having a person approve irreversible actions and being careful about what data you send to the model.

Why AI sometimes writes wrong things

Anthropic, the provider of the Claude model, admits this in its documentation: even the most advanced models can sometimes generate text that is factually incorrect or inconsistent with the given context. This phenomenon is called "hallucination" and, according to the documentation, it can undermine the reliability of AI-based solutions. Source: Anthropic documentation.

What follows from this: the correctness of the output is not always guaranteed, so you need a way to deal with that.

How the risk of incorrect answers is reduced

The same documentation describes several basic techniques:

In my view none of this guarantees anything, which is why I still recommend checking the outputs of important things. A similar principle, answering from your documents instead of from the model's memory, is used by an AI assistant over company data. I describe how it works in the article AI assistant over company data: how RAG works.

Where AI takes actions: the agent and human approval

A different situation arises when AI doesn't just write but also acts. In n8n, the AI Agent node connects a chat model with one or more tools, and the agent decides on its own which tool to call to complete the task. According to the documentation, tools use external services and APIs to perform actions and retrieve information.

Here the model's mistake turns from inaccurate text into a real action. In n8n you can require human approval before a specific tool runs. The workflow pauses and waits until a person approves or rejects the action. The documentation says this check is useful for irreversible actions: deleting data, sending external messages or making purchases. Source: n8n documentation.

Recommendation: anything that can't be taken back should be approved by a person. A message to a customer or the deletion of a record shouldn't go out without a check.

You can find more about agents on the page AI agent for business.

Data protection: what goes to the model

When a workflow sends text to a language model, the data goes to the model provider. So it pays to know how they will handle it. For the Claude API, Anthropic offers, among other things, a zero data retention (ZDR) agreement. It is a separate agreement with Anthropic, not an automatic setting.

The details are in the Anthropic documentation. In addition, the Office for Personal Data Protection of the Slovak Republic (Úrad na ochranu osobných údajov SR) points out that GDPR may apply to further handling of personal data, even from publicly available documents, and that the mere availability of a document is not a legal basis for further storing or providing it (dataprotection.gov.sk).

Nexflow does not provide tax or legal advice. Assessing GDPR compliance is a matter for your lawyer or data protection officer.

Overview: where it helps and where the risk lies

TaskRisk of errorRecommended approach
Drafting text, summariesIncorrect detail in the textA person checks it before use
Answers from company documentsInvented answerAnswer with a quote from the source, an "I don't know" option
Sorting enquiriesWrong categoryWhen unsure, hand over to a person
Sending a message to a customerIrreversible error outwardsHuman approval
Deleting data, purchasesIrreversible damageHuman approval, or don't leave it to AI

If customer communication is your first step, see AI chatbot for business. The differences between the formats are explained in AI agent vs chatbot vs automation.

When AI isn't worth it in a business

Five questions before deployment

  1. What happens if the model answers incorrectly? Who will notice?
  2. Is anything among the actions irreversible? If so, who approves it?
  3. What data will go to the model provider, and do you have a legal basis for it?
  4. Does the model answer from your materials, or from memory?
  5. What does the solution look like without AI, and would it be better?

You can find indicative prices in the article How much does AI automation cost for a small business?. My work starts at €390.

Frequently asked questions

Can AI make mistakes in a business?

Yes. Anthropic's documentation admits that even the most advanced models can generate factually incorrect text, known as hallucination. That's why I recommend human checks of important outputs and grounding answers in your documents.

How can I reduce the number of incorrect answers?

According to Anthropic's documentation, it helps to explicitly allow the model to say "I don't know", for long documents to first extract word-for-word quotes, and to require quotes for every claim. In my view this doesn't guarantee correctness, so I recommend checking important outputs.

Should AI send emails to customers on its own?

I recommend not leaving it unchecked. n8n lets you pause a workflow and request human approval before a tool runs, and the documentation lists sending external messages among the cases where the check is useful.

Does the model provider use my data for training?

Anthropic's zero data retention documentation states that if a feature requires retention, the data is not used for training without explicit permission. ZDR is a separate agreement, not an automatic setting. Verify the specific terms with the provider and, if needed, with a lawyer.

When is ordinary automation without AI enough?

When the task has fixed rules and can be written as "if A, then B". In my view, ordinary automation without a language model is enough in that case.

Sources

Related services

ServiceAI agent for business ServiceAI assistant over company data

Related articles

DecidingAI agent vs chatbot vs automation: what you need GuideAI Assistant for Company Data: How RAG Works GuideWhat to automate in a company first

Not sure whether AI is worth it in your business?

Describe the task you want to simplify. I'll tell you whether it needs AI or ordinary automation is enough, and where I would leave the decision to a person.

Describe the task →