The destination map
The security organization of the agentic era: one human head over every function, a fleet of agentic specialists underneath, and a gate that decides how far each function may run on its own.
L1
Manual
L2
Assisted
L3
Augmented
L4
Autonomous
You staffed your security functions with human experts because the work moved at human speed. Tickets, reviews, approvals: the whole control fabric assumes a person can read everything that needs reading.
Then agents arrived in volume. Industry analyst forecasts project 150,000 agents per enterprise by 2028. They hold identities, they call tools, and they act at machine speed. A human approval queue does not scale against that; the math of oversight breaks before the tooling does.
Industry analyst forecasts for 2028
So the question is not whether AI runs part of your security work. It is what the humans do when it does.
The hybrid model answers that as an engineering choice made before restructuring, not after. Decide the division of labor between the human head and the agent fleet for every function. Then let each function earn its autonomy at its own pace, through a gate, on evidence.
The ladder
AI-CMM is our model · calibrated to SAE J3016. Each rung names who holds the loop.
The functions
The functions on this page are a forward-looking reassignment of the same domains the site already stands on. Today, the diagnostic runs on the operating functions a board sees now; this page is the destination organization they are gating toward. One partition, two moments in time.
The human head
AI Generalist, GRC
GRC · A&A
The agent fleet
Compliance Mapping Agent
Audit Documentation Agent
Regulatory Monitor
Risk acceptance carries legal liability.
Never delegated
Risk acceptance, audit sign-off, regulator representation.
The human head
AI Generalist, Identity
IAM · CEK
The agent fleet
Access Intent Validator
Key Rotation Agent
Permission Graph Analyzer
Machine identities move at machine speed; approval queues are a vulnerability.
Never delegated
Identity policy, break-glass access, root-of-trust ceremonies.
Gate note: Intent validation gates separately at Augmented until misintercept rates prove stable.
The human head
AI Generalist, Data
DSP · HRS
The agent fleet
Data Classification Agent
PII Redaction Agent
Insider Threat Monitor
Redaction must be inline; post-hoc review is post-breach review.
Never delegated
Privacy policy, DPDP and GDPR obligations, insider escalation.
Gate note: The false-positive rate is a first-class gate metric; over-blocking demotes like leaking.
The human head
AI Generalist, ThreatOps
TVM · LOG · SEF
The agent fleet
Red-Team Agent
Forensic Reconstruction Agent
Anomaly Detection Engine
Containment at machine speed; high-impact remediation notifies a human.
Never delegated
Incident declaration, disclosure, the kill switch.
Gate note: SEF-04 is the live seed lever in our diagnostic; this card is where the model meets the Northwind demo.
The human head
AI Generalist, AppSec
AIS · IPY
The agent fleet
Code Scanning Agent
API Gateway Inspector
Protocol Validation Agent
Continuous scanning of AI-generated code; legacy integration needs human context.
Never delegated
Architecture exceptions, residual legacy risk.
Gate note: Agent-to-agent (MCP) validation gates on zero unverified task propagation across the window.
The human head
AI Generalist, Infrastructure
I&S · CCC · DCS · UEM
The agent fleet
Policy Optimization Agent
Drift Detection Agent
IaC Validator
Ticket-speed response to machine-speed drift is a standing exposure.
Never delegated
Change-freeze authority, physical security.
Gate note: Detection may run Autonomous while reversion holds at Augmented until rollback safety is proven.
The human head
AI Generalist, Supply Chain
STA
The agent fleet
Vendor Assessment Agent
Provenance Tracker
Shadow AI Discovery Engine
Vendor trust is a business judgment with contractual weight.
Never delegated
Trust tiers, contract risk, supply-chain attribution.
Gate note: Discovery runs hotter than decisioning; inventory completeness is the gate metric.
The human head
AI Generalist, Model Security
MDS · BCR
The agent fleet
Adversarial Defense Agent
Weight Integrity Scanner
Failover Coordinator
Adversarial defense is real-time math no human executes manually.
Never delegated
Model release authority, resilience posture, model retirement.
Gate note: MDS is new in AICM v1.1; the youngest domain starts with the shortest track record, so its evidence window runs longest.
A consolidation of the 18 domains and 247 control objectives of the CSA AI Controls Matrix v1.1.
The gate
Every function carries a measured autonomy score: task success, policy compliance, error frequency.
Every transition carries a gate-allowed threshold and a stability window.
The gate opens when the score clears the threshold and holds through the window.
It closes, and the function demotes automatically, the moment evidence breaches.
Promotion is earned in evidence. Demotion is automatic.
Read your own organization against that rule. A score that never rises is a capability problem. Adoption everywhere with an undesigned human-to-agent allocation is a mindset problem, and its gate stays shut no matter how many tools are bought. Only the engineered division of labor counts as transformation.
The gate is live today in the Function Diagnostic for Security Operations, seeded on the Northwind dataset; SEF-04 is the lever.
The full deployment pictured on this page is a destination, not a claim.
Workforce evolution
From
Legacy Security SME
Manual, siloed, tactical.
To
AI Generalist / AI Security Architect
Governance, intent, oversight, strategy.
The World Economic Forum projects that 59 percent of workers will need reskilling by 2030.
projectedWorld Economic Forum, Future of Jobs
Role-adjacent reskilling of the team you already have is the only plan that preserves institutional knowledge.
Reskill with Ace, your AI-security Center of Excellence.Then ace the new era.