The platform we run SEIRO on.
It started as an invoice template.
It began in 2025 as a faster way to make invoices. Today it runs our finance, our projects and our support, and it’s where every tool we use for clients gets tested first.
Order is real. Spacing is approximate.
Spreadsheets, templates, and apps that didn’t talk.
Our finances ran by hand: a spreadsheet and an invoice template. The rest of the work was spread across separate apps. A CRM, a kanban board, a few automations, time entries somewhere else.
Two things pushed us to change it. Paying for tools that didn’t fit how we worked. And knowing that if we didn’t learn to put this new technology to work on ourselves, we would fall behind.
We built what we needed. Then the next thing.
First, a small tool to make invoices from a template. Then expenses, so the reports for the people handling our taxes took minutes instead of an afternoon.
Then the team: projects, boards and time entries, all connected to invoicing. That gave us something we never had in spreadsheets: profit per project, without anyone adding it up.
AI came late. When there was too much to keep track of.
The first AI piece was an assistant that told us what to solve first. The first real job it took over was the one nobody wants: matching receipts to bank movements. One-off restaurants, receipts that look different in every place and every country. It reads them, suggests the match, or makes it.
Today, invoicing is one message. When a project reaches its goal, we tell the AI and it prepares the invoice from the project. An expense is a photo. The rest followed: analysis across the whole company, answers that pull from other departments and cite where they came from, and customer service that works alongside the team.
- Invoicing
Fill the template, check the numbersSay the goal is met. The invoice is ready. - Expenses
Match every receipt to the bank by handSend a photo. It reads and matches. - Tax reports
Export, clean, emailReady to send - Profit per project
Nobody knew for certainAlways up to date
A client site went down. We found out from the report.
A server setting let a broken update through. Our monitoring caught it. An agent went in through a guarded access layer we built, so it can only do what it’s allowed to, rolled the update back, then changed the setting so it can’t happen again.
What stayed, and what went.
Everything new runs here before it reaches a client. Most of it doesn’t make the cut.
| What we tried | Verdict | Why |
|---|---|---|
| Jev, from TypeSafe | Kept | Each AI request now carries only the tools it needs: 8 instead of 19 for a banking question. Two days later it was in a client’s build. |
| Agents with guarded server access | Kept | They fix what monitoring finds, inside strict limits. Every action ends in a report. |
| Automating our financial systems | Kept | Agents read receipts in any format, match them to the bank and book them only when every check passes. What needs a person goes to one review queue. |
| Adversarial multi-agent debate | Dropped | Agents tend to agree. When a strong model argues with a weaker one, the answer usually gets worse. The strongest model on its own did better. |
“What holds it back isn’t the AI. It’s how much time we have to build the next piece.”
Xavier Lopez Reynau · Founder, SEIRO