What AI agents can really do in business, and what they can’t

An AI agent is software you give a goal to. It decides for itself which steps to take and uses the tools you connect, such as email, your calendar or your management software. In a business it can take over part of the repetitive work, like reading, replying, collecting information and preparing drafts.

What it can’t do is take responsibility for decisions. It doesn’t work without tidy data, and it isn’t one hundred percent reliable. Start from those limits and it becomes a useful tool rather than a promise.

The difference between a chatbot, an automation and an agent

People often use these three words as if they meant the same thing. They don’t, and the difference matters when you have to choose.

A fixed automation follows a sequence decided in advance. When a form comes in, the contact goes into the CRM and an email goes out. It always does the same thing, which is a strength, but it stalls when a case doesn’t fit the pattern. A chatbot answers in a chat, with prepared replies or with an AI model, and it usually talks without acting, so it won’t move an appointment or update a file.

An agent, by contrast, has a goal and some tools. It reads the situation, decides the next step, checks the calendar or the management software, writes, and stops to ask when it needs to.

Anthropic, the company behind Claude, draws the line like this. In automated workflows the steps are defined beforehand, while in agents the model directs its own process and its use of tools. The same source warns that this freedom has a price, which is more time, more cost and the risk that an early mistake carries through the following steps. That’s why the best solution is often a fixed automation with AI in a single spot, for example to read an email and work out what it’s about.

What they can do, sector by sector

These examples are realistic, but adapt them to your case. In each one the agent handles the repetitive part and a person stays responsible for the result.

Real estate agencies

The agent answers enquiries about a property using the listing details and asks what a colleague would ask, such as area, budget, timing and whether a mortgage is needed. It then offers free slots for a viewing and sends the estate agent a summary, so whoever goes to the viewing already knows who they’re meeting.

Professional firms

In a professional firm, the agent can chase missing documents and track what has arrived and what hasn’t. It can also answer clients’ recurring questions with the firm’s information and prepare drafts for the professional to review.

In Italy there’s a specific rule here. Under the national AI law (Law 132 of 2025), AI in the intellectual professions may only be used for support activities, with the professional’s own work prevailing. Clients must also be clearly informed about the systems used.

Shops and e-commerce

The agent answers product questions, such as sizes, materials and compatibility, starting from your catalogue. It tells customers where their order is by reading your management software, and it collects return requests with the details you need. The customer gets an answer right away, and your team deals with the exceptions.

Restaurants and hospitality

The agent takes bookings against real availability and answers questions about opening hours, parking and services, in the guest’s language too. On allergens and intolerances, though, it must use only verified information and hand over to staff, because that’s an answer you don’t leave to improvisation.

B2B companies

The agent prepares draft quotes from your price list and the terms already agreed with that client, and the salesperson checks them before they go out. It also searches procedures, technical sheets and internal documents, so a new hire finds answers without interrupting colleagues.

It can even make meetings usable, as MeetMind does. It’s one of my projects, an AI product for web and mobile that turns meetings into summaries, decisions, tasks and a searchable knowledge base.

What they can’t do, or shouldn’t

Where someone is responsible, a person has to decide. Approving a discount, closing a contract or giving tax or legal advice isn’t for the agent, which can only prepare the ground. The GDPR points the same way. It gives people the right not to be subject to decisions made solely by automated means when those decisions have legal or similarly significant effects on them, with specific exceptions.

Messy data makes the agent work badly. If your price list exists in three versions and your procedures live in people’s heads, it will repeat the same confusion, only faster.

Nor is an agent reliable every time. When unsure, language models sometimes guess and produce answers that sound plausible but are wrong, and a 2025 paper co-authored by OpenAI researchers acknowledges this. You reduce the risk by having the agent answer only from your sources, getting it used to saying “I don’t know”, and putting a human check wherever a mistake is costly.

Finally, an agent doesn’t replace your relationship with the customer. It takes the repetitive part so you have time for the conversations that matter, and anyone who wants to reach you still can.

What an agent needs to work

An agent needs five things to work well.

• Start with tidy data and processes, meaning one up-to-date source for prices, availability, terms and frequent answers. Often this is the first job, and it pays off even without AI.
• It needs written limits that say what it can do, what it must never do and when it hands over to a person.
• Behind it there should be a person, a “human in the loop”, who approves drafts, handles exceptions and now and then rereads a sample of conversations.
• Privacy and transparency matter too. You need to know where the data goes, which providers process it and under what contracts, as the GDPR requires. Anyone talking to the agent must also know there’s an AI on the other side, and there’s a separate article on what the European rules change.
• Finally, take one process at a time and measure it. Choose a task that is repetitive, frequent and easy to correct, and compare before and after with your own numbers, such as time spent, response times and requests handled. Then decide whether to expand.

A useful agent does one specific thing well, with your data and under your control.

The questions to ask whoever pitches you an AI agent

Before you sign, get a clear answer to these questions.

• Which data does it draw its answers from, and who keeps it up to date?
• What does it do when it can’t answer?
• Which actions does it take on its own, and which need approval?
• Where do the conversations end up, and who can read them?
• How will we measure whether it’s working?

If the answers are vague, the project probably is too.

How I work on AI integrations

I only use AI when it cuts repetitive work or unlocks a process, so I start from the process and not from the technology. I look at how you work today, where time gets lost and where a mistake is costly. Sometimes a simple automation is enough, and if so I’ll tell you.

When an agent makes sense, I connect it to the tools you already use, with clear limits and points where a person decides, and we measure together whether it works. If there’s a process eating hours of your week, let’s talk.