Monoes Workforce · Technical Model

The LLM isn't the brain
of the whole system.

It's one component of the execution layer. Splitting Workflow, Agent, Policy, and Connector apart is what lets us swap models, processes, or ERPs without breaking anything downstream.

4Architecture Layers

5Human-in-the-Loop Levels

6Starting Connectors

Architecture

Four Layers, Each With One Job

Keeping these separate is deliberate. A reasoning model should never be the thing deciding what's allowed.

Workflow

Deterministic process controller

Defines the steps a process moves through, in order, every time.

Agent

The reasoning brain of some steps

Reads, extracts, classifies, and decides, but only within a step the Workflow hands it.

Policy

Deterministic "what's allowed"

Rules like "invoice over $20,000 requires approval": versioned, auditable, testable, kept separate from the Agent's judgment.

Connector

Standard bridge to external systems

The bridge to your ERP, CRM, email, or spreadsheets (swappable without touching the process logic).

Human-in-the-Loop

The Goal Isn't to Remove Humans

It's to let people spend their time on judgment, not data entry. Every engagement starts at Level 2 and earns its way up as trust builds.

Manual

A human does everything.

↓

AI Copilot

The AI recommends; a human executes.

↓

AI Executes + Human Approval

The AI performs the work but waits for approval on important actions.

↓

Autonomous Execution

The AI executes authorized work without approval.

↓

Autonomous Exception Handling

The AI handles selected exceptions on its own.

Process Model

Every Process Is a Structured Model

Not a prompt, not a flowchart drawing: a versioned definition the Workflow layer actually executes. Example for invoice processing:

{
  "process": "invoice_processing",
  "trigger": "invoice_received",
  "steps": [
    { "type": "extract_document" },
    { "type": "validate_vendor" },
    { "type": "match_po" },
    { "type": "decision", "rule": "variance < 0.05" },
    { "type": "approval" },
    { "type": "erp_action" }
  ]
}

Evaluation

Every Worker Is Held to a Scorecard

Historical cases become the evaluation dataset for every new agent version, before it's promoted to run live. Illustrative example for an Accounts Payable worker:

AP Agent

Extraction Accuracy: [measured per engagement]
Vendor Matching:     [measured per engagement]
PO Matching:         [measured per engagement]
Auto-processing:     [measured per engagement]
Exception Rate:      [measured per engagement]
Human Override:      [measured per engagement]

Connectors

We Connect to What You Already Run

No rip-and-replace. The Connector layer is a standard bridge to your existing systems. New ones get added as customers need them.

ERPCRMEmailExcel / CSVAccountingTicketing

The LLM isn't the brainof the whole system.