[object Object]

The procurement team approved the Enterprise Plus upgrade on the strength of three named Agentic AI use cases. Six months later, two of them are in production with measurable savings; the third was paused after an autonomous agent took an action nobody had authorized in the way the audit team meant. Agentic AI is real and the implementation discipline matters more than the licensing question.

What Agentic AI Does

AI agents diagnose, plan, and execute multi-step workflows with minimal human intervention within configured guardrails. Example flow: identify incident root cause from worklog and CI context, look up applicable runbook, apply the resolution by calling the right downstream system, notify users, close the incident — no human in the loop except for exception cases that the agent escalates explicitly. The unit of automation moves from “one task” to “one workflow.”

Agentic execution model:
  observe -> classify -> plan -> act -> verify -> log
  exception path: escalate to named human role
  budget: per-conversation token cap, per-flow action cap
  audit: full trace including model, version, prompts, decisions

The Tier Gate

Agentic AI is exclusive to the Enterprise Plus tier as of the 2026 packaging. Lower tiers see the capability in demos and not in their instance. Budget for the upgrade if agentic execution is strategic; or build equivalent narrower logic via flows on lower tiers with more human checkpoints. The Enterprise Plus delta only pays back if at least one Agentic AI workflow runs at meaningful volume; quantify the displacement before signing.

Setup Fundamentals

Define which workflows the agent owns end-to-end and which it merely augments. Set confidence thresholds for autonomous action versus human escalation; below the threshold, the agent escalates rather than acts. Establish audit trails (the AI Control Tower handles this) and explicit rollback paths for every autonomous action — what happens when the agent’s action turns out wrong. Do not turn autonomous agents loose on high-risk workflows (financial, security, compliance) without oversight; the risk-reward does not pencil.

// Agentic guardrail pattern
function executeWithGuardrails(action, context) {
  if (action.risk_score > context.autonomous_threshold) {
    return escalate(action, context, 'risk_above_threshold');
  }
  if (!action.is_reversible && action.risk_score > LOW_RISK) {
    return escalate(action, context, 'irreversible_action');
  }
  return execute(action, context);
}

Early Adoption Signals

Production deployments are running across multiple verticals as of mid-2026. Outcomes range from clear wins (password resets, common access requests, well-defined incident patterns) to edge-case failures that required rollback. The pattern of success: narrow scope, mechanical workflows, well-curated knowledge sources, explicit guardrails. The pattern of failure: broad scope, judgment-heavy workflows, stale knowledge, vague guardrails.

Common Failure Modes

Confidence thresholds set high without calibrating against observed accuracy — the model reports 85% confident but is right only 70% of the time, and autonomous actions accumulate errors. Calibrate before trusting confidence. Audit trail captured but never reviewed — the trail exists for the audit, not for ongoing operations; build a weekly anomaly review. Agents granted broader scope than needed “for flexibility” — least-privilege applies to non-human actors too.

Cost Considerations

Agentic workflows compound token cost compared with single-shot AI calls. Each step in an agentic flow may include planning prompts, tool-call prompts, observation prompts, and verification prompts. Cap per-execution token usage and surface daily cost in the AI Control Tower. A poorly-bounded agent that retries on transient errors can multiply daily cost overnight; the Control Tower’s anomaly alerts are the safety net.

What Changed in 2026

The Zurich release added the AI Control Tower as the governance plane for Agentic AI. Customers running agents without the Control Tower were doing it manually via custom logging and dashboards; the Control Tower replaces that with a supported native surface. Upgrade to Zurich before scaling agentic workloads, not after.

Implementation Sequence

Pick one workflow with high volume, contained risk, and a clean audit story. Implement as augmentation first (suggest, do not execute), measure agreement rate for 30 days. Promote to autonomous execution only when agreement rate exceeds 85% and the operations team trusts the agent. Expand to a second workflow only after the first is stable for 60 days. Aggressive multi-workflow rollouts produce the most expensive failures.

What to do this week: identify the one workflow your team would agentify first if it were free; design the guardrails for that workflow before evaluating the licensing change.

[object Object]
Share