AI agents: what they are and when a small business needs one (explained by a 14th-century merchant from Prato)

For the past year it has seemed like every piece of software has its own "AI agent". Your ERP promises one, your CRM, even your inbox. But what is an agent really, what can it do today and, above all, do you need one? To answer, let's start with a man who had plenty of agents, none of them digital.

Francesco Datini and his factors

Francesco di Marco Datini, a merchant from Prato who died in 1410, built in the second half of the fourteenth century a network of companies and trading houses stretching from Avignon to Pisa, Florence and Genoa, all the way to Barcelona, Valencia and Majorca. He couldn't be everywhere. Each branch was run by his factors (fattori): trusted people who handled day-to-day business, buying, selling, shipping and collecting. Datini directed them from afar, by letter.

And what letters. His archive, kept in the palazzo where he lived and now home to the State Archives of Prato, holds more than 150,000 letters, along with account books, insurance contracts and exchange documents. It's one of the richest sources we have on medieval European trade, precisely because every factor had to write: what he'd done, at what price, with whom, what he planned next.

Look at the structure: an owner who sets goals, agents who decide the steps on the ground, constant correspondence that works as a log. The factor had autonomy, but within a perimeter, and he accounted for everything.

A well-designed AI agent is a digital factor. And the rules that worked in fourteenth-century Prato still work.

What an AI agent is, in one sentence

An AI agent is a system built on a language model that, given a goal, decides for itself which steps to take and which tools to use to reach it: looking something up, reading a file, querying the ERP, writing a draft, asking for confirmation.

The clearest definition comes from Anthropic's technical guide Building effective agents (December 2024): workflows are systems where models and tools are orchestrated through predefined code paths; agents are systems where the model dynamically directs its own process and tool use.

What an agent can do today, and what it can't

Today's agents are good at:

• using tools: querying databases, documents, websites and calendars, sending requests to programs through APIs;
• breaking a task into steps and adapting when one step doesn't go as planned;
• handling long, variable text: emails, contracts, tickets, notes;
• preparing drafts ready for human review.

They're still weak when:

• errors compound: in a ten-step chain, a small mistake at step three drags along everything after it. Anthropic's guide warns that autonomy means "higher costs, and the potential for compounding errors";
• judgment depends on unwritten context: your relationship with that customer, the history of that supplier;
• instructions come from untrusted sources: an agent that reads emails can be fooled by hidden text designed to give it orders. This risk, called prompt injection, tops the OWASP Top 10 for LLM applications, which also warns against giving systems "excessive agency", meaning too many permissions.

When a small business really needs an agent

In June 2025 Gartner, the technology research firm, predicted that over 40% of agentic AI projects will be canceled by the end of 2027, due to escalating costs, unclear business value or inadequate risk controls. The same release describes agent washing, existing products rebranded as "agents" without being agentic, and estimates that of the thousands of vendors claiming agentic products, only about 130 are real.

To stay out of that 40%, run four tests before you start:

1. Is the task variable? If the steps are always the same, classic automation is enough: it costs less and makes fewer mistakes. We explain why in our guide to business process automation.
2. Is the result easy to check? A draft quote can be reread in two minutes. An executed payment can't be undone with a click.
3. Are the data and tools reachable? An agent without access to your ERP is a factor without a warehouse.
4. Is there enough volume to justify the effort? Designing, testing and supervising an agent takes time. For three requests a month, it isn't worth it.

If you answer "yes" to all four, an agent may make sense.

Realistic examples for a small business

Typical scenarios, not promises:

• Quote requests: the agent reads the email, identifies the products, checks price list and availability, asks the customer for missing details and prepares the draft. A salesperson reviews and sends it.
• After-sales support: the agent collects the order number and photos of the problem, checks the warranty, proposes the standard fix and hands unusual cases to a person.
• Research and summaries: gathering information on a grant, a supplier or a regulation, with sources listed, to be verified before use.
• Logistics: reconstructing the status of a problem shipment by cross-checking emails, tracking and documents, as mentioned in AI for logistics and warehouses.

Common thread: in every example the agent prepares, the person decides. That's also why agents change jobs more than they eliminate them, as we discuss in AI and work: which tasks change.

The rules of a good (digital) factor

Back to Prato. What made a network of factors scattered across the Mediterranean manageable? The same things that make an agent manageable.

• Written instructions. Goal, limits, what to do when unsure. An agent with vague instructions improvises.
• A clear perimeter. Which tools it can use, on which data, up to what amounts. Minimal permissions, never admin credentials.
• A log of everything. Like the factors' letters: every agent action must leave a readable trail of who, what, when and why.
• Confirmation thresholds. Above a certain amount, or for certain actions (paying, deleting, sending externally), a human yes is required.
• Distrust of external content. What the agent reads in an email or on a website is data, not an order.
• Testing in a sandbox. Before touching real data, test in a separate environment, as Anthropic also recommends.
• Periodic review. Datini read his factors' accounts. You read your agent's logs.

Then there's security in the strict sense: an agent that reads email is as exposed to phishing as an employee, if not more. More in phishing, AI and cybersecurity for SMBs.

Want to know whether an agent makes sense for your company?

We help business owners and SMBs tell what's useful from what's just noise, starting from people's work. Explore our services or book a free intro call: in half an hour we'll figure out together whether you need a digital factor or just a good log.