IT

AI agent development for businesses

How to read the states

  • waiting
  • running
  • done
  • waiting for a person

The log

How does an AI agent work, step by step?

An AI agent takes a goal, makes a plan, uses the tools it has access to, checks the result and hands over to a person when it should. Follow an example task: scroll, and the log writes itself.

  1. As it arrives

    Goal

    The agent reads the email and sets what it has to achieve: work out who is writing, answer both questions and draft a reply a salesperson can approve.

    done

  2. Straight after

    Plan

    It breaks the work down: find the company in the CRM, look up how the integration works in the docs, take prices from the approved price list, suggest a call.

    done

  3. Over the next few seconds

    Tools

    It uses only the tools it has been given, one at a time, and notes what each one returned.

    • CRMlooks up the sender's company found, no active contract
    • Docssearches “back-office integration” two relevant guides
    • Price listreads the approved plans team pricing
    • Calendarlooks for free slots for a call three options

    done

  4. Before replying

    Check

    It reads the draft against the rules: every claim has a source, the prices are the approved ones, the tone is the company's. One detail isn't covered by the docs, so it flags it rather than making it up.

    done

  5. At the end

    Result

    It produces the draft reply, updates the contact's record in the CRM and attaches the suggested call times.

    done

  6. When a decision is needed

    Hand-off to a person

    The salesperson gets everything in one notification, fills in the flagged detail and sends it. The agent sends nothing on its own: a person always has the last word.

    waiting for a person

The kinds

What kinds of AI agent are there, and when should you use each?

Four shapes, from simplest to most complex. The rule of thumb: pick the simplest one that does the job, because every extra degree of autonomy costs you predictability and checks.

Kinds of AI agent: what they do, when to use them, when they fall short
KindWhat it doesUse it forFalls short when
AI assistantAnswers questions in a chat from the knowledge you give it.FAQs, internal support, searching your documents.It needs to act inside your systems, not just answer.
Tool-using agentDecides which tools to use (CRM, email, documents, calendar) to reach a goal.Tasks whose steps vary: qualifying leads, drafting replies, updating records.The steps never change: a workflow is cheaper and more predictable.
Automated workflow, for example in n8nRuns the same sequence every time, with AI only where text needs reading or writing.Stable, repetitive processes: routing, data extraction, notifications, syncing.Exceptions are the rule and every step needs judgement.
Multi-agent systemSeveral specialised agents pass work between them, with one coordinating.Long tasks with distinct roles, such as research, writing and review.Almost always at the start: it adds cost, latency and more places to go wrong.

Guardrails

What limits does an AI agent need?

An agent is only as useful as the limits you set: it should be able to do little, leave a trace of everything and stop when it isn't sure.

Oversight
Actions that matter (sending, paying, deleting, promising) go through a person. The agent prepares; a person decides.
Permissions
Access only to the tools and data the task needs, read-only wherever possible. Never admin keys.
Logs
Every step recorded: what it read, which tool it called, what it answered. Without a log you can't fix anything.
Costs
Every step spends model calls: spending caps, alerts and a ceiling on retries for each task.

Some limits remain: a model can be confidently wrong, it can be tricked by instructions hidden in the text it reads (prompt injection) and it handles data the GDPR protects. That's why we design every agent so that a mistake is visible and can be undone.

Frosted-glass sheets fanned out like the pages of a ledger, with small lilac marks

Experience

Who builds AI agents in the group?

AI agents are the daily work of Giallo Studio, the group's AI agency for Italian SMEs: AI audits, process automation, and agents connected to business tools. Across the group we use n8n for workflows and language models for agents.

Alongside Giallo sit the people who build custom web apps and AI features for startups: when an agent needs an interface, a database or an integration that doesn't exist yet, we build it within the same group. Everything we do is on our services page.

Questions about AI agents

What are AI agents and how do they work?

An AI agent is a program that uses a language model to decide the steps towards a goal and to call the tools it has access to, such as a CRM, email or documents. It works in a loop: plan, act, check the result and, when needed, pass the decision to a person.

How do you build an AI agent for your business?

Start from one precise, recurring task, write the process down with its exceptions, and decide which tools the agent may use and with what permissions. Then trial it on real cases with a person approving every action, and widen its autonomy only where the logs show it works.

How much does an AI agent cost?

It depends on the task, how many systems it connects, the model and the volume: on top of development there are model usage costs, which grow with requests. We don't publish a price list; once we've seen the process, we tell you what drives the cost in your case.

What's the difference between an AI agent and a chatbot?

A chatbot answers questions; an AI agent gets things done: it reads and writes in your tools, follows a plan and checks the result. That's why an agent needs permissions, logs and oversight that a chatbot doesn't.

Can an AI agent work without supervision?

For low-risk, well-measured tasks, yes: routing requests or updating fields, for example. For anything hard to undo, such as sending, paying or promising something to a customer, we recommend a person makes the call.

Further reading

Got a process that eats hours every week?

We reply within one business day.

Let's talk