How it works

AI decides what should happen.
Deterministic policy decides whether it is allowed to.

That principle shapes every layer of BizApps360 — from the tools the AI can call to the ledger that records why it called them.

Architecture

Seven layers, one execution foundation

1. Experience

Agent Employee dashboard, Action Center, approvals, explanations.

2. Agent orchestration

Goals, tasks, planning, state, memory and the execution loop.

3. AI / reasoning

Model router, LLMs, vision, classification and extraction.

4. Policy engine

Limits, permissions, deterministic business rules and approvals.

5. Action engine

Controlled tools for ERP, email, documents, CRM and other systems.

6. BizApps360 platform

AP, AR, workflows, exceptions, reporting and integrations — the proven execution layer.

7. Audit ledger

Immutable, traceable record of evidence, policy evaluation, action and approval.

Policy & autonomy model

Five levels of autonomy — set per action, per risk category, per amount

Level 0L0

Observe

AI reads and analyses. No action.

Level 1L1

Recommend

AI proposes an action with evidence.

Level 2L2

Assisted

AI executes after human approval.

Level 3L3

Controlled autonomous

AI executes defined low-risk actions within policy.

Level 4L4

Autonomous employee

AI runs the full defined workflow within strict financial and security boundaries.

Recommended initial financial controls

Low-risk matched invoicesAutonomous after validation
Medium-risk transactionsHuman approval
New vendorsHuman verification
Payment creationApproval required initially
Payment executionHuman-controlled initially
Memory or model output never overrides explicit policy.

Action engine

Controlled tools, not open credentials

Existing BizApps360 APIs become strictly-schematised agent tools. Every call is authenticated, tenant-isolated, permission-checked, idempotent, validated and logged. The LLM never receives unrestricted database or ERP access.

  • Strict schemas & authentication
  • Tenant isolation & permissions
  • Idempotency & transaction validation
  • Timeout, retry & rollback handling
  • Logging & auditability

AI decision ledger

Every meaningful decision answers one question: why did the AI do this?

Each entry captures the full chain from input to result, timestamped and versioned.

  1. 01Input document and transaction identifiers
  2. 02ERP facts retrieved
  3. 03Company policy applied
  4. 04Checks performed
  5. 05Decision or recommendation
  6. 06Tool / action invoked
  7. 07Human approval, if any
  8. 08Result or error
  9. 09Timestamp, agent and version
Agent Employee Deployment Program

Nine phases. Objective gates. No leap of faith.

#PhaseTypical durationOutcome
1Discovery1–2 daysWorkflow, policies, ERP and success criteria documented
2Secure connection2–5 daysSAP B1, email and document sources connected
3Historical learningSeveral daysPast invoices and exceptions used to validate behaviour
4Configure employee1–2 daysRole, authority, policies, tools and escalation rules
5Shadow mode1–2 weeksAI decisions compared with human decisions
6Assisted mode1–2 weeksHumans approve AI recommendations and actions
7Controlled autonomyProgressiveLow-risk work executed automatically
8Go-liveOngoingProduction Agent Employee with monitoring
9Monthly value reviewMonthlyROI, accuracy, interventions and expansion opportunities
Production readiness gates

Objective gates rather than subjective confidence

> 98%

Invoice extraction accuracy

> 97%

PO matching accuracy

> 95%

Exception classification

0

Critical financial errors

0

Policy violations

100%

Audit coverage

Illustrative targets. Thresholds are validated by risk category and customer workflow during the pilot; we never claim a universal accuracy guarantee.

Onboarding data checklist

What we ask for before shadow mode

  • SAP Business One version and deployment details
  • SAP Business One SDK / API access
  • Company and entity structure
  • Vendor master, PO / GRN examples, invoice samples
  • Historical approved and rejected invoices
  • Approval matrix, tolerance rules, tax rules, payment terms
  • Exception categories and AP inbox / document sources
  • User roles, escalation contacts, security & compliance requirements

See the governance model against your own policies.

The workflow assessment maps your approval matrix, tolerances and exception causes to the autonomy model — before any pilot begins.