L
Constitutional Layer
Autonomy levels L0–L4
From every decision requiring human approval (L0), to a low-risk-auto / high-risk-approval split (L2), to self-evolution (L4) — stepwise, at the organization's pace.
TRUST & GOVERNANCE
To the question "what if the AI goes rogue?", we answer by design.
The first promise of an enterprise AI operating system is that it can be stopped.
Autonomy levels · approval gates · rollback·audit trail · budget guard · Kill Switch are part of the core skeleton, not features.
D
Audit Ledger
Decision Trail
"Who, when, on what grounds, and which alternative was chosen" is linked 1:1 to every execution. Available instantly on audit request.
R
Reversibility
1-Click rollback
Every executed decision carries a restore plan. Price, exposure, campaign and segment decisions revert to the previous state in one click.
K
Emergency Stop
Kill Switch & Budget Guard
When anomalies, budget overrun or user reports cross a threshold, only that Agent stops immediately. Other operational flows are unaffected.
AUTONOMY LEVELS
Self-evolution isn’t a switch you flip all at once.
It’s a level you raise gradually, matched to your organization’s pace and situation.
Start at L0 and widen autonomy only as much as KPI evidence accumulates. Policy, legal and high-risk decisions remain the human domain at every level.
L0ObservationAutomation ~8%
AI only proposes. Every execution is performed by a human.
L1AssistedAutomation ~30%
For low-risk, structured tasks, AI drafts and a human approves.
L2Mixed autonomyAutomation ~60%
Low-risk decisions run automatically; high-risk ones split off into the approval queue.
L3Risk-based autonomyAutomation ~85%
Mostly automatic execution. Only high-risk, policy and legal decisions are made by a human.
L4Self-evolutionAutomation ~95%
Execution patterns, validated by KPI evidence, self-evolve into the default for the next plan.
REVERSIBILITY
Unconditional automation that carries risk
doesn’t fit the enterprise environment.
Enterprises don't need fast automation — they need reversible automation. Here's how a generic AI and AXCommerce diverge when things fail.
Generic AI automation · hard to reverse
It's hard to trace what went wrong.
› AI executes
✕ Check the result — evidence is scattered
✕ "Where did it go wrong?"
With execution evidence scattered, tracing the cause is hard, and reversing means manually restoring data.
AXCommerce.AI · reversible from the start
Every execution carries a restore plan.
✓ Execution + restore plan enclosed
✓ Traced via Decision Trail
✓ 1-Click rollback
Every R2+ execution carries idempotency · audit · before/after snapshots · a rollback plan. If something's wrong, one click returns to the previous state.
COMPLIANCE
Audit and regulatory response are ‘built in’ from the start.
Every decision, execution, approval and rollback is logged automatically. AI Agent memory stores no raw PII, secrets or full logs — only summaries, owner, retention period, tenant scope and evidence references are allowed.
Regulatory response
GDPRPIPAEU AI Act
Security certifications (LG CNS)
ISO 27001SOC 2 Type IIK-ISMSCSAP