Agentic ERP Intelligence Platform
A multi-agent AI layer inside an enterprise ERP: JARVIS, an orchestration layer that plans across business domains, domain agents with their own tools and permissions, governed execution against real services, human approval on anything that writes, and realtime voice on the same path.
The problem
An ERP asks people to know which module to open, which report to run and which department to chase. A chatbot bolted on top does not help — it has no permissions, no domain boundaries and no way to actually do anything.
The approach
JARVIS, an orchestration layer over domain agents. Each agent owns one business area and a fixed set of registered actions. Every action declares whether it reads or writes, and writes are sealed and held for human approval. The model can only ask for what the registry exposes, and each call runs through the same application service a human would.
The outcome
One request can now cross inventory, marketing, CRM and finance and come back as a grounded answer or a prepared action — with the permission checks, budgets and audit record that make an enterprise willing to switch it on.
- Role
- AI architecture & agent design
- Architecture
- JARVIS orchestrator · domain agents · tool registry
- Providers
- Anthropic · OpenAI · Gemini · DeepSeek
- Governance
- Permissions · approvals · budgets · audit
Most AI features in business software stop at prompt in, text out. This one executes. A request arrives by chat or by voice, the orchestration layer resolves what is being asked and who is asking, routes it to the agents whose domain it belongs to, and those agents call registered ERP actions rather than inventing SQL. Anything that writes stops for a person.
The interesting engineering is not the model call. It is everything around it: a registry of actions with declared effects and input schemas, an execution guard that enforces the caller's own permissions and company scope, delegation between agents with a depth limit, budgets on steps and tool calls and duration, versioned agent configuration, and an audit trail that can answer what an agent did and why.