The problem: the triangle nobody closes
Every Italian site foreman I’ve talked to in the last two years repeats the same sentence: “I don’t have time to open five different tools, I’m here to build.” Behind that sentence there’s a triangle that no software really closes:
- Vertex 1 — Sales. The CRM knows you signed a contract, but has no clue what’s happening on the site.
- Vertex 2 — Site. The site management software (if any) records attendance and progress, but doesn’t talk to sales.
- Vertex 3 — AI. AI agents stay outside both. When you use them, it’s to write emails, not to update your progress report.
Why generalist CRMs fail on the site
I’ve been implementing Salesforce and HubSpot for enterprise clients for over ten years. They’re powerful tools but designed around the concept of a deal: negotiation, conversion, forecast. A worksite is not a deal. It’s an autonomous micro-company with attendance, change orders, progress reports, non-conformities and weather. Trying to model it as a pipeline loses 70% of the relevant data — and the people on the ground stop updating it after two weeks.
Native AI skills, not bots stitched on top
The difference between “CRM with a chatbot” and “AI-native CRM” is radical. A chatbot reads your database in read-only mode and replies in natural language. A native AI skill does things instead: opens a non-conformity, updates a progress report, creates a task, sends a notification to the works director. Every skill is a function of the data model, not an external wrapper.
In WorkSiteCRM we’re particularly proud of three skills:
- Automatic progress report from Daily Log. The agent reads the site log lines, photos, attendance and task completion percentages, and compiles a signable report ready for the works director.
- NC from photos with AI vision. You snap a crack, the agent classifies it (severity, type), fills the NC record and assigns it to the right team.
- Auto-sent weekly report. Every Monday, a cron-skill produces a one-page report and emails it to the works director. Zero maintenance.
MCP: the protocol that unlocks the ecosystem
The Model Context Protocol is to AI what REST was to web APIs. It standardizes how an agent calls external tools. Instead of developing custom integrations for Gmail, Google Drive, n8n, your ERP and your BIM reader, you write (or download) an MCP connector and the agent uses it automatically.
For Italian construction this is a game-changer: you can connect the worksite to your custom billing software without paying for three months of integration development, because the connector is reusable and maintained separately from the main product.
Castaldo Solutions’ role
Castaldo Solutions doesn’t just build software: it builds AI skills trained on the Italian construction domain. Terminology, progress/NC formats, regulations, signing flows. We don’t give you a generic LLM that hallucinates item names: we give you an agent that knows your sector and your constraints. Every Enterprise customer also receives custom skills trained on their own specifications and workflows.
What to expect from upcoming articles
Over the next months we’ll cover concrete case studies (with real data where possible), technical deep dives on CPM with lag, weather overlay, GDPR-first multi-tenancy. If you want us to cover a specific topic, write to us — we build the editorial roadmap together with our customers.