AI Automation
Custom AI agents and agent teams that operate your business software—not chatbots. Built for your domain, with mandatory human approval and full audit trails on every action.

Action-taking AI agents embedded inside a Belgian B2B invoicing ERP
Action-taking AI agents embedded inside a Belgian B2B invoicing ERP
Sandelia is a Belgian B2B invoicing and ERP platform built around the PEPPOL network, which became mandatory for Belgian B2B invoicing in January 2026. The founder's vision goes beyond invoicing: to become the first European ERP with a full internal team of AI "employees" — agents that don't just answer questions, but actually operate the software on behalf of the business owner.




We designed and built a multi-agent AI system deeply integrated with the Sandelia platform — not a chatbot, but two specialized AI employees that take real actions inside the ERP, with enterprise-grade safeguards.


A coordinator dispatches requests across specialized agents (MIA, LEO, with HR, Marketing, Legal, and Operations on the roadmap). Inter-agent routing means MIA hands off accounting questions to LEO automatically and returns one clean answer.


Agents call real tools inside the ERP — `CreateInvoice`, `ListUnpaidInvoices`, `FetchKPIs`, `ExportQuarterlyDocs`, `GenerateReport` — with role-based permissions and structured validation so the LLM decides *when* to act, but the platform enforces *what* is allowed.


Mandatory confirmation popups for every write action, read/write permission separation, full audit logging, and explicit disclaimers across the product. Agents advise; humans approve.


Curated knowledge base of Belgian VAT, fiscal, and accounting regulations. The agent tries the internal knowledge base first, falls back to web search, and only then uses general model knowledge — always citing the source, with metadata-weighted recency so regulation changes (e.g., takeaway VAT moving from 6% to 12%) are applied correctly.
Python, LangGraph for multi-agent orchestration and tool routing, LangSmith for tracing and live prompt management, OpenAI API (GPT-4 / GPT-5 class) for reasoning.
FastAPI for the async API layer, Celery + Redis for background tasks, PostgreSQL for chat history, actions, and audit logs.
React chat UI embedded into the client's Next.js app via a self-contained module, light/dark mode, inline tool-confirmation popups, contextual and full-page interaction modes.
AWS - RDS, ECS, S3, ACM for TLS, auto-scaling alerts
Python, LangGraph for multi-agent orchestration and tool routing, LangSmith for tracing and live prompt management, OpenAI API (GPT-4 / GPT-5 class) for reasoning.
FastAPI for the async API layer, Celery + Redis for background tasks, PostgreSQL for chat history, actions, and audit logs.
React chat UI embedded into the client's Next.js app via a self-contained module, light/dark mode, inline tool-confirmation popups, contextual and full-page interaction modes.
AWS - RDS, ECS, S3, ACM for TLS, auto-scaling alerts
Tell us the process your team handles by hand every day. We’ll design the AI agent, web application, or dashboard that takes it off their plate — built around how your business actually operates.