How to build an AI agent. The sheet first, then the tool.

You can build an AI agent in an afternoon with the tools you already use. What decides whether it works gets written first: what it has to do, where its data comes from, who approves and how it's measured.

This page is part of Itria's AI agents for business, where every agent has its own six-line sheet with its measurement alongside. Here you'll find the same sheet, blank, to design one yourself.

How to build an AI agent, in five steps

You build an AI agent in five steps, and only the last one is about the tool. The first four go on a sheet of paper: what it has to do, where its data comes from, where it leaves the result, who approves it. Then you decide the number it has to beat, before you switch it on, and only then do you choose what to build it with.

Step	What you decide	The mistake it prevents
1 · The taskjust one	One task, written the way you'd put it to a colleague: “reply to availability enquiries”.	The agent that “handles customers” and doesn't do anything specific.
2 · The data it takes inwith a real example	The sources: inbox, WhatsApp, a folder of PDFs, a spreadsheet, business software. For each one, a real example.	Instructions written for data that never arrives in that form.
3 · Where it writesone place per result	A draft in the inbox, a row in the spreadsheet, a contact in the CRM.	Results scattered across three places, which nobody double-checks.
4 · Who approvesbefore it goes out	What goes past a person: anything that touches a customer, money or an outgoing message.	A message to a customer that nobody has read.
5 · How it's measureddecided in advance	Today's number and the one to beat: minutes per task, errors per hundred, response time.	An agent that “seems to work” and nobody can judge.

Which tool: ChatGPT, Gemini, Claude, Copilot or n8n

For an agent that works on a single source and hands the result to a person, the assistants you already use are enough. With ChatGPT, Gemini, Claude and Copilot you write fixed instructions, attach the business's documents and test on real cases. On ChatGPT's price list, for example, custom GPTs come with the Plus plan, from 23 euros a month.

When the agent has to read from an inbox and write to a spreadsheet without anyone pasting the data across, you need an automation: n8n, Make or Zapier connect a model to the sources and systems. At that point the work shifts from writing instructions to building connections, and that's where you need skills many businesses don't have.

Istat confirms it: among businesses that considered artificial intelligence without adopting it, 58.6% cite a lack of skills and 43.0% high costs, in the survey Imprese e ICT, 2025. Skills come before price.

With what	What you need	When it's enough
ChatGPT, Gemini, Claude, Copilotan assistant	Written instructions, attached documents, a subscription.	A single source, and a person who reads the result before using it.
n8n, Make, Zapieran automation	A workflow that connects an inbox, spreadsheet or business software to a model.	Several sources, or a system to write to, with steps that are always the same.
Custom-builtsomeone builds it	Connections to your systems, approval, a log of what it does, measurement.	When it touches customers, money or outgoing messages, or several systems at once.

The sheet for designing an AI agent, to download

The sheet puts the five steps on one page: six lines for the agent, the cost of doing the task by hand and three questions to decide whether to do it yourself. It comes in Word and Excel, to fill in, and in PDF to print. The Excel file works out the cost per month and per year by itself.

Free on this page, in Italian: Word, Excel, PDF and the worked example. In English: Word, Excel, PDF. No form in the way and no email to leave.

A worked example: the agent that puts documents into Excel

The example is one of Itria's agents, with its measurement: across 30 documents it extracted 98.7% of fields correctly, 234 out of 237, in 6.264 seconds per document. It's a test bench measurement, meaning a test system, from 4 September 2026, and we can run it again in front of you. Its full sheet is on the page for the AI agent for invoice processing.

Line	Filled in
Task	Put the data from invoices, orders and delivery notes that arrive by email into Excel, without retyping it.
Takes in	PDFs attached to emails from suppliers and customers, including in other languages.
Does	Recognises the document, extracts the agreed fields and gives each one a confidence score; below the threshold it stops and shows it to a person.
Approved by	The person who runs the accounts, for fields below the threshold; every step leaves a line in the log.
Writes to	A shared Excel sheet, one row per document, with the columns chosen by the business.
Measure	98.7% of fields correct across 30 documents, 6.264 seconds per document. Test bench, 4 September 2026.

The cost of doing it by hand, with sample figures to replace with yours: 200 documents a month at 4 minutes each come to 13 hours and 20 minutes; at 25 euros an hour that's 333 euros a month, about 4,000 a year. That's the number to put next to the price of the agent, in how much an AI agent costs an SME.

Do it yourself or have it built: three questions

You can do it yourself when the agent takes data from one source, writes to one place and a person reads the result before using it. You need someone to build it when it touches several systems, when it acts towards a customer or on money, or when nobody can say what it does if a source doesn't respond. The sheet's three questions are there for this.

The question	If the answer is yes
One source, one place?	You can try it yourself with the assistant you already use, on a few real cases, measuring before and after.
Does it touch customers, money, outgoing messages?	Before switching it on, you need a written approval step and a log of what it does.
Does it stop when it can't find the answer?	Then it's been designed properly. If it answers anyway, it makes things up, and the damage lands where nobody's looking.

If the agent talks to a customer, it has to say it's a system. Article 50 of Regulation (EU) 2024/1689 on artificial intelligence, which applies from 2 August 2026, requires the information to be given “at the latest at the time of the first interaction or exposure”. On the sheet it goes in the Approved by line.

How much it costs to build an AI agent

Doing it yourself costs the subscription plus the hours of the person who builds and checks it; having it built costs 5,000 to 20,000 euros for a custom agent, according to the public price lists of five Italian providers checked again on 25 September 2026. The price bands, and how they compare with the cost of doing the task by hand, are in how much an AI agent costs an SME.

For comparison, the ChatGPT business subscription costs 21 euros per user per month billed annually, according to OpenAI's price list as checked on 13 October 2026. It's a tool for everyone; a custom agent does one specific task inside your systems, and the difference between the two is on the page about an AI agent and a chatbot.