ServiceNow Super Intelligence Control Tower: A Runtime Control Review

ServiceNow's September 2026 update uses the name Super Intelligence Control Tower for the product previously presented as AI Control Tower. The renamed control plane brings inventory and governance into a shared ServiceNow operating model, with security and observability plus value measurement. Buyers still need to identify which routes receive runtime enforcement and which records support governance after an event.

The June 2026 welcome guide describes AI assets as configuration items in the ServiceNow CMDB and organizes the product around five pillars. The August and September 2026 update introduces the current Super Intelligence Control Tower name and a new interface. It also introduces runtime monitoring and a kill switch, along with expanded connectors and continuous evaluation. AI Gateway controls complete the update.

TL;DR

• Use the current product name, Super Intelligence Control Tower, while preserving AICT where ServiceNow uses the acronym.
• Test inventory completeness and owner accountability separately. Do the same for lifecycle approvals and governance evidence, then verify evaluation coverage.
• Map runtime controls to exact SDK and agent routes, plus MCP and provider routes. Check gateway routes before claiming enterprise-wide enforcement.
• Add an independent HTTP model-request control where route-level identity policy and evidence are required.

The five-layer review matrix

A control tower succeeds when each discovered asset has an accountable owner, governed state changes, useful telemetry, and an enforceable response path. Empty owner fields reduce a polished inventory to a directory, while a live trace without a policy decision remains telemetry.

ServiceNow's June guide names AI Steward and AI Product Owner. It also names AI Risk and Compliance Manager and CAIO or AI CoE Lead, plus security and privacy stakeholders. The September update adds asset-owner experiences and lifecycle controls. During acceptance, put a printed asset record in front of each role and ask who signs the next state change. The blank signature line exposes ambiguous ownership faster than a workshop deck.

Inventory: Super Intelligence Control Tower models AI systems, models, datasets, prompts, and MCP assets in the CMDB. Service Graph Connectors cover named providers, including AWS, Azure, GCP, OpenAI, Anthropic, and Moveworks. Microsoft Agent 365 was added in the September update. Test one sanctioned asset, one sub-production copy, one duplicate, one retired system, and one unknown asset. Record connector scope and discovery lag.

Lifecycle owner: Every asset needs a named product owner and steward. The update describes governed workflows for changes and retirement, plus required approvals and dependency reviews. It also describes alerts for stalled lifecycle work. Acceptance should transfer an asset between owners and withdraw an approver. Then attempt retirement with an active dependency. The workflow must preserve who approved each change and which dependency blocked it.

Governance control: The welcome guide describes risk assessments and lifecycle management. It also describes compliance frameworks and playbooks aligned with the EU AI Act and NIST AI RMF. The September release adds evaluator templates and prioritization, alongside guided workflows. Test a policy change against a fixed asset sample. Capture the old state and approval. Preserve the new control assignment and exception, then retain the evidence export rather than accepting a dashboard score alone.

Observability and evaluation: ServiceNow says external traces can arrive from AWS CloudWatch and Azure and Microsoft Foundry. Application Insights and GCP Vertex AI are also supported, as is Traceloop. The update describes near-real-time scoring of quality and safety using an LLM as judge. It covers step and tool traces, plus output traces, configurable sampling, and metric weights. Record evaluation prompts and model version. Preserve the threshold and sampling rule, along with failed-trace handling and human review. A score without those inputs cannot be reproduced.

Runtime enforcement gap: The release describes SDK-layer threat detection and blocking, a kill switch that coordinates credential revocation for selected environments, and an AI Gateway covering MCP discovery, policy, identity, short-lived OAuth credentials, pausing, and telemetry. Buyers must map each control to the routes it actually covers. Native SaaS AI, direct provider calls, local tools, unmanaged SDKs, and model endpoints outside the configured gateway may follow different enforcement paths.

Inventory completeness needs a denominator

Connector coverage should be written as a fraction. The numerator is assets discovered and normalized into CMDB records. The denominator is the set of cloud accounts and subscriptions, plus projects and ServiceNow instances. It also includes provider tenants and repositories, along with agent builders and MCP catalogs the enterprise expects to cover. "Connected to Azure" says little when the connector sees one subscription out of 40.

Reconcile the control tower with billing exports and cloud inventories. Check SSO application lists and API gateway logs, plus procurement records and developer-platform registrations. AI governance tools comparison can help classify inventory and workflow products separately from runtime products without treating them as interchangeable. Name an exception owner for every source that cannot be reconciled.

ServiceNow's duplicate detection and metadata-gap recommendations can improve record quality. The September article says steward review happens before recommended changes are applied. Preserve that approval evidence with the changed fields and reviewer's identity. Automated enrichment can suggest an owner or relationship, but the CMDB record becomes operational only when someone accepts responsibility for its accuracy.

Lifecycle and governance need adversarial tests

Run a lifecycle exercise with one agent that changes model provider and gains a new dataset. Have it add an MCP tool and move to production. The control tower should trigger the required review and retain the previous configuration. It should also identify dependent services and show the approver. Then attempt the same change through an unmonitored path. That second result reveals the actual governance boundary.

Connect each governance control statement to an evidence source. Use CMDB state and the approval record, plus the trace and evaluation result. Add the gateway decision and credential-revocation event, along with the incident case. AI audit trail requirements by regulation provides a useful structure for separating portfolio evidence from per-event records.

My opinion is that a control tower should be judged by its worst-owned asset, not its best dashboard. Pick the record with missing metadata and stale traces. Make it one with an overdue approval. If the system can drive that asset to a defensible state, the operating model is doing real work.

Observability and evaluations have different jobs

A trace reconstructs an execution with its steps and outputs. An evaluation interprets selected behavior against a metric. ServiceNow's update describes both, including configurable trace sampling and weighted quality and safety scoring. Sampling reduces collection cost, but it also changes the evidence available for incidents. Record which routes require full traces, which allow sampling, and how a high-risk signal raises the collection rate.

LLM-as-judge scoring introduces another model into the control process. Version its prompt and model, then preserve the threshold and exception workflow. Use fixed test cases with known expected outcomes and measure disagreement between the evaluator and human reviewers. Evaluation supports triage and trend analysis. A deterministic policy decision is still required when a prohibited request must be blocked every time.

Runtime enforcement requires a route map

ServiceNow's September update gives buyers several enforcement mechanisms to test. AI Gateway can block unapproved MCP servers at build time and scan tool metadata. It can also inspect selected runtime data and issue scoped OAuth credentials. Pausing servers and recording connection metrics complete the described controls. The kill switch coordinates containment for supported agents and identity providers. SDK-layer controls can detect and block threats where the SDK is present.

Create a route map with source identity, transport, intermediary, destination, control, fail posture, and evidence. Include one direct model API call, one ServiceNow agent, one third-party agent, one MCP tool call, and one native SaaS interaction. AI runtime security controls gives the architecture vocabulary for that test.

Ask which calls bypass the SDK and which agents can connect directly to an MCP server. Then ask which providers feed traces after execution and which kill-switch targets support verified revocation. Keep design-time approval and runtime blocking as separate findings. Do the same for post-runtime analysis and credential containment.

DeepInspect

DeepInspect can cover authenticated HTTP model traffic that an application deliberately routes through its stateless proxy. On that segment, it evaluates identity-aware policy for prompts and responses and writes a per-decision audit record independently of the calling application. Super Intelligence Control Tower can retain the portfolio view and owners, along with workflows and traces. It can also retain evaluations and broader response coordination. MCP tool execution and native SaaS activity remain with their actual controls, as do local processes and other routes.

Book a demo today.