Work / 07 projects

Seven projects, described without client names.

These are projects Mattia designed and led in 2026 for companies in Italy, working with engineers and trainers on delivery. Client names are left out for confidentiality; sizes, numbers and outcomes are as they happened.

01 / 07

Reading supplier invoices for customs declarations

Sector: Customs brokerage
Client: Italian customs broker

Fig. 01 / before and after

Invoices retyped by handSystem reads, declarant checks and files

The situation

Every declaration started from supplier invoices in dozens of layouts and languages. Declarants read them and retyped values, quantities, origins and commodity codes into the customs software, one line at a time.

What we did

01
Mapped the declaration process with the declarants and listed the fields that actually come from the invoice.
02
Designed, with an engineer, a system that reads each invoice, extracts those fields, normalizes them and sends them to the declaration software through its API.
03
Kept the declarant as the person who reviews and files, with the extracted data shown next to the original document.
04
Started with exports, then extended the same flow to imports and to EU trade statistics returns.

Where it stands

In production since 2026, with a monthly maintenance agreement.

02 / 07

Finding where AI fits across 41 office activities

Sector: Manufacturing
Client: Italian subsidiary of a German group, about 150 people

Fig. 02 / before and after

41 activities, no shared view10 interventions ranked

The situation

The company had a group AI assistant, Copilot licenses and an internal AI working group, but no shared view of which processes were worth changing first. Many processes had run the same way for 15 to 20 years.

What we did

01
Interviewed the heads of each function and ran a questionnaire answered by about 100 employees on how they spend their week.
02
Mapped 41 recurring activities, each with the time it takes, the tools involved and what AI could prepare or check.
03
Ranked ten interventions by hours freed, effort and risk, from quick changes in existing tools to two candidate agents.
04
Wrote the report for management with a sequence for the next twelve months.

Where it stands

Report completed in November 2026; the next phase is being scoped with the client.

03 / 07

A second round of Copilot training built around each role

Sector: Legal
Client: Law firm in Milan and Rome, about 130 business staff

Fig. 03 / before and after

Mixed classesTen classes by level and role

The situation

The firm had rolled out Microsoft 365 Copilot to almost all of its business staff and done a first round of training with another provider. Classes mixed people who had never used AI with heavy users, and the feedback reflected it.

What we did

01
Proposed an assessment of every employee before any class, plus a short survey on the tasks each person actually does.
02
Grouped ten classes by level and by role, so people in the same class work on the same kind of documents.
03
Fitted the schedule to shift coverage with two-hour sessions, two trainers in parallel, and a separate session for team leads.
04
Added short video lessons shared before the classes and a second assessment at the end, so the firm can see the change.

Where it stands

Program approved in November 2026; classes planned from January 2027.

04 / 07

A phone agent that answers the main line around the clock

Sector: Non-profit, health and social services
Client: Local committee of a national humanitarian organization

Fig. 04 / before and after

One line, office hours onlyAgent sorts and books, offices get tickets

The situation

The committee’s main number received very different requests: medical appointments, psychological support, social aid, patient transport, training courses and, sometimes, emergencies. Calls outside office hours went unanswered.

What we did

01
Mapped each type of request with the people who handle it and wrote down what the agent may do on its own and what it must hand over.
02
Designed the call flow with a voice-AI engineer: the agent books clinic appointments in the booking system, gives course information and sends emergencies to 112.
03
Turned every request it cannot close into a written ticket emailed to the right office, so nothing depends on a callback note.
04
Built a browser test page so staff could call the agent and check it before the switch.

Where it stands

Built in 2026; the rollout follows the client’s data-protection and phone-system steps.

05 / 07

An AI program for a product team, won through a public tender

Sector: Public sector technology
Client: Product department of about 30 PMs and engineers

Fig. 05 / before and after

Occasional chatbot useFive automations from the backlog

The situation

A state-owned digital payments company wanted its product managers and engineers to move from occasional chatbot use to AI built into their daily work, within the tools approved by its security team.

What we did

01
Wrote the tender proposal and won it, with a fixed scope and a public-sector contract.
02
Ran two days on site in Rome: one on how current AI systems work, one hands-on in small mixed groups.
03
Chose five automations with the team from their own backlog: test writing, activity reporting, documentation updates, the community newsletter and capacity planning.
04
Followed up with a calendar of short sessions, pairing and a demo day through December, booked by the team through a small scheduling app.

Where it stands

Running until December 2026.

06 / 07

Grouping 120 people by actual AI skill before training

Sector: Luxury manufacturing
Client: Yacht builder, about 120 office staff in scope

Fig. 06 / before and after

One group, very different levelsFour classes by score, one for executives

The situation

The company had Copilot licenses for part of its office staff, a new ERP rollout taking people’s time, an AI Act awareness obligation and two earlier AI initiatives that employees had found too technical.

What we did

01
Ran an objective AI skills assessment, completed by 92 people, with results by person and by level.
02
Built four classes of about 20 by score, plus a separate class for directors and executives.
03
Kept the content on Copilot inside the tools people already use, with exercises on their own documents.
04
Planned a second assessment at the end to measure the change.

Where it stands

Classes scheduled in October and November 2026.

07 / 07

Weekly AI lessons for 75 branches

Sector: Automotive retail
Client: Car dealership group with 75 branches

Fig. 07 / before and after

Headquarters onlyWeekly lessons for the branches

The situation

A multi-brand dealer group wanted AI use to spread beyond headquarters, but branch staff spend the day with customers and cannot sit in a classroom.

What we did

01
Ran an AI ambassador program at headquarters: a skills survey, four hands-on workshops and a monthly update.
02
Moved the branch program to short video lessons released every Monday, each on one task people do in the showroom or back office.
03
Set up two one-to-one sessions for each of the 19 executives, on their own work.

Where it stands

Headquarters program completed in August 2026; branch lessons continue weekly.

References are available on request once there is a concrete project to discuss.

Start with a 20-minute call.

Write to mattia@mullionstudio.com with a line on your firm and two or three times that work for you. If a review makes sense, we agree on scope and price in the same call.

Book a 20-minute call How a review runs