Service-Anchored AI Governance

At Knowledge 2026, ServiceNow expanded AI Control Tower (AICT) into the enterprise control layer for every AI agent, model, dataset and workflow - across AWS, Google Cloud, Azure, SAP, Oracle and Workday, not just ServiceNow itself. AICT is no longer an add-on; it is built into Now Assist Pro+ and Enterprise+ applications by default. Functionally, every customer on those tiers now has an AI governance layer whether they have configured it or not.

AICT does five things well: Discover, Observe, Govern, Secure, Measure. It will find your agents, watch them at runtime, risk-assess them, govern their identities, and measure their value.

One thing AICT does not do is the first thing a regulator will ask about.

AICT produces an agent inventory. It does not, by itself, connect that inventory to the CSDM service model: the chain that says which Critical Operation or Important Business Service each agent actually acts on. That connection is not a ServiceNow gap; it is a modelling decision the customer has to make deliberately. Almost nobody is making it.

The consequence is concrete. When an agent misbehaves, you get an alert in a governance dashboard. You do not get a Business Service impact report your Chief Risk Officer can escalate, your operational-resilience self-assessment can cite, or your auditor can trace. Under FCA PS26/2 (in force 18 March 2027), DORA, the US §165(d) resolution-planning regime, and the EU AI Act, that distinction is the difference between an evidenced control and an unevidenced one.

This reference architecture shows the exact CSDM v5 chain that turns an AICT agent inventory into service-anchored, regulator-traceable AI risk. It is vendor-accurate, CSDM-validated, and built for regulated financial services.

A precise, current picture, because the architecture only makes sense if the product baseline is right.

1.1 The five pillars (verified)

1.2 What changed at Knowledge 2026

• 30 new enterprise connectors spanning AWS, Google Cloud and Azure plus SAP, Oracle and Workday. AICT now governs AI that runs outside ServiceNow, not just Now Assist.
• Built in by default for Now Assist customers. AICT capabilities are included with Now Assist Pro+ and Enterprise+ applications, not sold as a separate add-on. ServiceNow has expanded AICT into the corporate governance layer; for customers on the relevant Now Assist tiers, it is included rather than upsold.
• "Action Fabric": a major overhaul positioning AICT as the control layer for autonomous work.
• Agent kill-switches: runtime stop controls for misbehaving agents.
• MCP Server included in every Now Assist and AI Native SKU.
• NVIDIA Enterprise AI Factory validated-design inclusion: governance extends to large-scale model workloads in the data centre.
• Availability: features roll out from April 2026; AICT enhancements enter Innovation Lab in May 2026; general availability expected August 2026.

1.3 The ServiceNow "Customer Zero" benchmark

ServiceNow publicly reports running 1,600+ AI agents internally and tracking $500M+ in cumulative AI value as its own Customer Zero. These are ServiceNow's stated figures for its own estate, useful as a scale benchmark for what "many agents" looks like in practice, cited here as ServiceNow's own claim rather than an independently audited number.

The takeaway for this document: AICT is genuinely strong at Discover/Observe/Govern/Secure/Measure. None of those five pillars, on their own, produces a service-impact statement. That is the gap Section 2 describes.