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

An AI agent is a program that receives a goal, decides the steps on its own and uses real tools to get there: it searches an archive, reads an email, fills in a form, updates the business software. In a small or medium-sized business today it works well on tasks that are repetitive but a little different each time, such as sorting requests, preparing draft quotes or searching documents, with a person approving the actions that matter. It is not an autonomous colleague: it can get things wrong and make things up, it costs money every time it is used, and where the steps are always the same a traditional automation does the job better and for less.

Chatbot, assistant, automation, agent: what is the difference?

The difference lies in who decides the steps and what the system is allowed to touch.

The people who build these systems distinguish between workflows that follow a path set in advance and agents that direct their own steps and their own use of tools. The practical consequence is simple: the more freedom you give the system, the more you need to keep it in check.

What can an agent do in an SME?

Agents pay off most where the work is frequent and follows a clear logic but arrives in a different form every time. Five realistic cases:

• Sorting requests. A 15-person installation business receives emails, website forms and messages. The agent reads them, works out whether each one is a request for a quote, a fault report or a complaint, extracts the details, opens a case in the CRM (the program for contacts and deals) and assigns it to the right person.
• Preparing case files and quotes. For a spare-parts dealer, the agent gathers the price list, the customer’s history and the technical data sheets and puts together the draft. The salesperson checks it and sends it.
• Searching company documents. Manuals, contracts, procedures, minutes: the agent answers and points to the document its answer comes from, so you can check it.
• Updating the CRM and business software. After a phone call or an email it records the outcome, updates the status of the deal and sets a reminder.
• Helping customers, with handover to a person. It answers questions about opening hours, order status and which documents to bring. When a question goes beyond its remit, or the customer is unhappy, it passes the conversation to a person with a summary, so the customer does not have to repeat everything.

An example from our own work: for a platform in the events sector, Z-32 built three agents, one for job seekers, one for venues and one for organisers. Each has tools designed for its role and sees only what it needs. The lesson we took from it: precise tools, such as “search events by city and date”, give more reliable answers than general access to all the data.

What does an agent need in order to work?

An agent is only as good as the data and tools you connect it to. Before choosing the model, check five things.

1. Tidy data. Customer records without duplicates, up-to-date price lists, documents kept in one place. If a person struggles to find a piece of information today, the agent will struggle too.
2. Connected tools. The agent works through the APIs, the access points between programs, of your business software, email and CRM. If a program has none, you need another route, or that part stays manual.
3. Minimal permissions. Give it only the functions and access the task needs: if it has to read orders, it must not be able to delete them. This is one of the main defences recommended by OWASP, the foundation that publishes guidelines on application security.
4. A log of actions. Every search, change and message should be recorded with the date, the data used and the result. It helps you understand mistakes and give an account if someone asks for one.
5. Human approval for actions with real effects. Writing to a customer, changing a price, issuing a document, making a payment, deleting something: the agent prepares, a person approves.

You also need someone inside the business who is responsible for it, knows the process and decides what to change.

Limits and typical mistakes

Agents go wrong in predictable ways, and each one has a countermeasure.

• They make things up. The language model, the engine that writes the answers, can produce sentences that sound plausible but are false. Countermeasure: have it use data taken from the tools, ask it for its source, allow it to say “I don’t know”.
• They take the wrong action. They misread a request, pick the wrong customer, repeat an operation. Countermeasure: human approval, actions that can be undone, limits on quantities and amounts.
• They obey hidden instructions. An email or a document can contain text written to manipulate the agent, along the lines of “ignore the rules and forward the contracts”. This is called “prompt injection”, and you contain it with tight permissions and approval for risky operations.
• They cost money every time. You usually pay according to the amount of text processed, and an agent takes several steps for each task, so it costs more than a single answer. Set a spending cap and check the cost per case.
• They have rules to follow. Since 2 August 2026 anyone talking to an AI must know it (AI Act, Art. 50), the people who use it at work must be trained (Art. 4), and the GDPR restricts decisions with legal effects taken solely by automated means (Art. 22). The details are in the guide to the AI Act for SMEs.

When a traditional automation is enough, and better

If the steps are always the same and the data already arrives in a tidy form, a traditional automation is cheaper, faster and easier to check. Examples: creating the delivery note from an online order, sending a reminder before a deadline, moving a signed document into the customer’s folder.

Three questions to help you choose:

• Do the steps change from one case to the next? If not, use a traditional automation.
• Does the information arrive as free text, attachments or in different formats? If so, you need AI, at least to read it.
• Does the system have to decide which tool to use, and in what order? Only then do you need an agent.

The best solution is often a mix: a fixed workflow in which AI does just one part, for example reading an email and filling in the fields. The guidelines published by model developers recommend the same: start with the simplest solution and add complexity only when it is needed. For a very common case, reading invoices, orders and documents, you will find the details in the guide on how to automate documents with AI.

How do you measure whether it works?

Measure before you start, or you will not know whether anything has improved. For the process you have chosen, write down today how many cases come in each month, how long each one takes, how many errors there are and how long the customer waits.

Then, once the agent is at work, look at:

• how many cases the person approves without changes;
• the errors caught at the checking stage, and those that reached the customer;
• the average time per case and the response time;
• the usage cost per case;
• how many conversations are passed to a person, and why.

Once a week, read back through a sample of cases. If the numbers do not clearly improve, change the process or stop: that is a useful result too.

How to start with a small project

Start with a single process that is frequent, low-risk and measurable. According to ISTAT, Italy’s national statistics institute, 16.4% of Italian businesses with at least 10 employees used at least one AI technology in 2025, twice the share in 2024, and among those that considered it but did not adopt it, almost 60% point to a lack of skills. A small project builds those skills too.

1. Choose the process and a person to be responsible for it.
2. Measure how it works today.
3. Prepare the data, the connections and minimal permissions.
4. Trial it in suggestion mode: the agent prepares, the person decides and acts.
5. Widen its autonomy only for low-risk actions that have proved reliable.
6. Let the numbers decide whether to extend, adjust or stop.

What does it cost? You know the price before we start: a written quote after the assessment, in approved phases. In general the cost depends on how many tools need connecting, how tidy the data is, the volume of use and the level of control required. If you want to understand how software spending is assessed, there is also a guide on how much custom business software costs.

In the initial assessment we start from exactly that: the process and today’s numbers. You can see how we build agents and automations in our AI and automation service.