Home / Case Study

Internal Case Study

We built AI to run
our own 5-company group.

Before building for clients, we built for ourselves. A 24/7 AI Chief of Staff managing operations across Italy, Netherlands, Switzerland, and Mexico. 8 specialized agents. 30+ daily automations. 18 months live without a single restart.

Running since January 2025
5 companies, 4 countries
8 specialized AI agents
Built by Señor Business AI
Build the same for my business → See all client projects
8 specialized AI agents running simultaneously
30+ automated tasks running every day
5 companies managed across 4 countries
18mo live in production, zero downtime

The problem: 5 companies, one person, not enough hours.

Running five companies across four countries generates an impossible amount of operational noise. There are emails to classify, stock levels to monitor, B2B leads to follow up, market intelligence to gather, weekly reports to produce, customer inquiries to handle in three languages, and a dozen other tasks that each take 15 minutes but together consume the entire day.

The options were: hire a team of assistants and coordinators, or build the system ourselves. We built the system. Not as a proof-of-concept. As the actual operational backbone we depend on every day.

The system: a multi-agent operating layer.

Santiago (the name we gave it) is not a chatbot. It is an orchestrated multi-agent system built on Claude AI, running 24/7 on a dedicated VPS, connected to email accounts, CRMs, e-commerce platforms, databases, and calendar APIs across all five businesses.

The core architecture has three layers:

The orchestrator. A central agent that receives commands via Telegram, routes them to the right specialist agent, monitors all background jobs, and delivers a structured morning briefing every day at 7 AM.

The specialist agents. Eight agents, each with a defined domain, memory, and set of tools. They operate autonomously on schedule or on demand.

The automation layer. 30+ cron-scheduled jobs that run without any human input: stock checks, email digests, market scans, pipeline reports, and customer service queues.

The 8 specialist agents

📧
Emiliano
B2B email agent for the food distribution company. Reads the inbox daily, classifies inquiries, drafts replies with conversation memory, and routes approvals to the owner.
💬
Sofia
Multilingual customer service agent embedded in the e-commerce storefront. Handles order inquiries and product questions in 18 languages, 24/7.
🏗️
Leonardo
Strategic consultant agent for a management consulting firm. Produces structured client analysis documents, proposals, and AI roadmaps in Italian.
📣
Andrea
Sales and outreach agent for an AI consulting service. Manages demo site generation, prospect research, and outreach sequences.
🔍
Researcher
Read-only web research agent. Runs competitor analysis, market scans, and due diligence tasks on demand or on schedule.
🛡️
Adriano
Inspector General. Audits the performance of all other agents weekly, generates improvement reports, and flags behavioral drift.
📊
Angelo
Senior strategy consultant agent. Produces McKinsey-style analysis for high-stakes decisions: market entry, pricing strategy, competitive positioning.
🎨
Isabella
Content and digital marketing agent. Manages social content, email newsletters, and brand asset production across companies.

What runs every day without human input.

The 30+ daily automations cover everything that used to require manual attention. A non-exhaustive list of what runs on schedule:

7:00 AM daily
Morning audio briefing
Structured voice brief delivered via Telegram: pipeline status, email summary, calendar, stock alerts.
9:30 AM daily
B2B inbox processing
Emiliano reads and classifies incoming business email, drafts replies, routes urgent items for approval.
Every 2 hours
Risk and anomaly scan
Monitors all systems for errors, unusual patterns, and security anomalies. Sends Telegram alert if anything is off.
Monday 7:00 AM
Market intelligence scan
Competitor activity, pricing signals, and market news delivered as a structured Telegram briefing to start the week.
4x daily
Deal and revenue radar
CRM pipeline updates, revenue tracking, and deal movement alerts across all active sales pipelines.
Monday 7:00 AM
Weekly self-score
Adriano audits all agent behavior from the past week, scores performance, and proposes improvements to the system prompt.

How it is built.

The entire system runs on a Hostinger VPS in Europe. The orchestrator is a Python process managed by systemd (not cron, not a cloud function). It restarts automatically on failure, logs every action, and sends Telegram alerts on any error.

AI layer: Claude AI (Sonnet) as the primary model for all agents. Gemini for text-to-speech audio briefings. All agents run via the claude-agent-sdk with persistent session IDs and context management.

Communication layer: All commands, briefings, and alerts flow through Telegram. A custom-built Telegram bot receives instructions, routes them to the right agent, and delivers outputs back in real time.

Data layer: SQLite for session persistence, conversation memory, audit logs, and cost tracking per user and per request. Google Drive for document storage and sync.

Integration layer: Gmail API (3 accounts), Google Calendar, Wix API, Shopify API, Apollo.io, Hostinger DNS MCP, PostgreSQL on AWS RDS (for Vigil SaaS).

Claude AI (Sonnet) Python 3.10 Telegram Bot Gmail API Gemini TTS SQLite systemd Hostinger VPS Google Drive API Apollo.io Wix API Shopify API

The key architectural decision: everything flows through Telegram.

We chose Telegram as the single interface for all agent interactions. Not a dashboard, not a web app. One Telegram bot that receives commands, reports results, sends alerts, and delivers briefings. This meant zero additional UI to maintain, instant mobile access from anywhere, and a natural conversation history that doubles as an audit log. Every agent action is logged, timestamped, and searchable.

What running this system actually looks like.

The morning starts with a 4-minute audio brief from the orchestrator. Stock levels at the food distribution company, reply-ready emails drafted in the B2B inbox, open deals in the pipeline with movement since yesterday, any anomalies detected overnight. All of this happens automatically before 7 AM.

During the day, commands go directly to Telegram. "Draft a follow-up to the German distributor who asked about wholesale pricing last week." The orchestrator retrieves the conversation thread, drafts the email with the right context and tone, and presents it for approval before sending. The whole interaction takes under 2 minutes.

The measurable outcome: tasks that used to require 3 to 4 hours of manual work per day now take under 30 minutes of review and approval. The cognitive overhead of tracking 5 businesses simultaneously dropped by roughly 80%. Nothing falls through the cracks because the system does not forget, does not get tired, and does not miss scheduled jobs.

How it evolved over 18 months.

January 2025
First deployment: single orchestrator + 2 agents
Basic Telegram bot with morning briefing and B2B email classification. Running on a simple VPS with no persistence.
March 2025
Multi-agent architecture deployed
Sofia (customer service) and Emiliano (B2B email) added as dedicated specialist agents with independent memory and tool access.
May 2025
Automation layer: 15 scheduled jobs
Market intelligence, pipeline tracking, and stock alerts added as cron-based autonomous runs.
August 2025
Expanded to 6 agents, 25+ automations
Leonardo (consulting documents), Andrea (outreach), and Researcher added. SQLite audit logging and cost tracking implemented.
November 2025
Adriano and Isabella added, self-improvement loop
Weekly self-scoring system deployed. Adriano audits all agents and generates improvement recommendations automatically.
June 2026
Current state: 8 agents, 30+ automations, 5 companies
Full production system. 18 months live. Zero unplanned downtime. We now build this same architecture for clients.

Why we built this before building for clients.

Most AI agencies sell the idea of AI. They show demos, describe use cases, and point to research. We chose a different path: build the system on ourselves first, run it in real conditions for a year and a half, and only then offer it to clients.

This matters for a few reasons. We know exactly where these systems break under real operational pressure. We know what 3 AM looks like when an agent has consumed the wrong context and needs to be corrected. We know how to design memory systems that do not degrade over time. We know how to handle the edge cases that no demo will ever surface.

When a client hires us to build their AI infrastructure, they are not hiring a team that read the documentation. They are hiring the team that has been running this in production for 18 months on their own business.

Every system we build for clients uses this same architecture.

The orchestrator, the agent memory model, the Telegram interface, the audit logging, the scheduled automations. We adapt it to the client's business, their tools, their team, and their language. But the foundation is tested and proven. We have been stressing it on ourselves for 18 months.

Want the same system running your business?

We have built this for our own 5 companies. We can build it for yours. Tell us what operations you want to automate and we will scope it with a fixed price.

Based in Vicenza, Italy. Active in Italy, Netherlands, Switzerland, Mexico, France, Spain. Remote worldwide.