Enterprises wrestling with AI ontologies hit their biggest snag not in building one but in spotting where a model’s “latent ontology” drifts from a firm’s curated definitions. Large language models map business concepts in embedding space, forming geometric links. Companies layer semantic rules and metrics on top. When those two maps don’t align, agents pull from the wrong “source of truth,” spawning errors in tasks from document classification to automated decision-making. The real work lies in detecting those gaps and patching just the misaligned bits. Shadow AI—unmanaged AI tools popping up in teams—has reignited old headaches: nobody owns them, security can’t spot them, and data flows unchecked. The tools plug into SaaS apps, ride on identity platforms and API keys, and operate outside IT’s radar. As organizations rush to deploy chatbots and copilots, they’re amplifying governance blind spots that already plagued cloud and access-control projects. At Zenith Live 2026, Zscaler rolled out a zero-trust stack for AI agents. It extends its Zero Trust Exchange with an AI Broker to handle machine-to-cloud and agent-to-agent traffic. There’s Endpoint AI Security to watch models running on laptops and servers. Behind the scenes, an AI Access Graph maps which data and services each agent can touch. Zscaler also beefed up AI Protect: asset inventories, code-scanning tools, prompt harvesters and an AI red-teaming service to stress-test agent behavior. On the billing front, Salesforce snapped up m3ter to bake usage-based charges into its Agentforce Revenue Management suite. As enterprises move from flat-fee subscriptions to meter-based, outcome-oriented pricing for AI services, Salesforce wants to own that meter. Instead of chaining together third-party platforms, firms will track compute, API calls and other consumption metrics right inside the Salesforce ecosystem.