Services · AI automation

AI automation for processes that already repeat

Agents for repetitive work with clear rules and enough volume to pay for the build. Northern Italy and Switzerland, on site when it helps.

Last updated 3 October 2026

In AI automation work I map repetitive processes and decide which steps need a fixed workflow and which need an AI agent to choose the next action. I'm Matteo Gazzurelli, an agentic AI consultant with MaGa Srl in Brescia, working with SMEs in Northern Italy and Switzerland. We set human checks where errors matter.

Who it's for

Five conditions. The size of the company matters less than the maturity of the process.

  • High volume, repetitive workThe same task many times a week, done by hand today.
  • Stable rulesYou can explain how each decision in the process is made.
  • Tools and data already digitalThe work happens in systems an agent can read and write.
  • A before and after you can countHours, errors or turnaround measured today, so the change is visible.
  • Enough volume to pay for the buildAn agent costs engineering time; the process has to give it back.

I turn down one-off tasks, processes that still change every month and work that lives only in someone's head.

Which task takes the most hours?

Three answers, and the agent estimates the hours you could get back in a year. The assumptions stay next to the number.

530hours back per yearAssumption: 85% of weekly hours on manual reporting, over 52 weeks. Measure current hours first; verify savings in a pilot.

The three answers are all the form asks for. How the answers are handled is in the privacy policy.

Use cases

Three scenarios. The hours apply the estimator's assumptions to each one; the real number comes from measuring your process.

Agent or workflow?

The first question in every engagement.

I call it an agent when the model chooses the next step during the task. In a workflow the steps are fixed in code before it runs.

Most repetitive work is better served by a workflow with fixed steps. An agent is worth building when the work needs judgement at several points and a person can check the calls that cost.

A workflow follows a fixed approval path.

An agent chooses which records to check next.

CriterionWorkflowPredefined stepsAI agentSteps chosen during the task
Next stepSet by codeChosen by the model
Best suited toPredictable sequencesTasks requiring adaptation
ToolsCan call APIs and databasesCan call APIs and databases
Human checksAt agreed checkpointsAt agreed checkpoints

The platform comes after the answer: Gemini Enterprise Agent Platform, Claude Managed Agents or a plain workflow engine, depending on the systems you already run.

Google vs Anthropic for agents, October 2026

What you get

A written go or no-go

For each process I say whether an agent is worth it, whether a workflow is enough or whether to wait, and why.

An architecture built for production

Checkpoints, logs and a way to roll back, designed before the first agent runs.

A team that can run it

Your people learn the system while we build it, so it keeps running after the engagement ends.

The technology chosen on your case

Google, Anthropic or something else, decided on the systems you already run, the data you have and who will maintain it.

A manufacturing SME, in detail

One engagement, anonymised: the context, the problem, what I did and what changed.

Manufacturing SME, Northern Italy

Context
A manufacturing company in Northern Italy with a small office team and a lot of repeated administrative work around orders and reporting.
Problem
The same data was typed into several systems by hand, and someone spent part of every day chasing the status of things. The owners wanted those hours back, inside the tools they already used.
What I did
I mapped the process as it actually ran, chose the technology on the case, and put agents on the repetitive steps with a person checking the ones that matter.
What changed
The person who used to copy the data by hand from one system to the next now checks the agent's work.
About Matteo

Related

Talk · 21 May 2026

Agentic AI: Stop prompting, start building agents

GDG Brescia, Blue Tech Hub, with Nicola Luna. In Italian.

Article · Aug 2026

Google ADK in 2026: Where It Fits in Gemini Enterprise Agent Platform

Where the Agent Development Kit sits in Google's stack, and when Cloud Run is enough.

Service · Remote

Fractional CTO

Architecture, hiring and releases, decided with the team and written down.

Have a process in mind?

Describe it in a few sentences: what repeats, who does it, in which systems. A reply within 2 business days.

Request an intro