Software's Path Forward In The Agentic AI Era — With ServiceNow's Amit Zavery

The opening argument: AI isn’t killing software, it’s exposing who actually built for enterprise

Alex Kantrowitz frames the big fear directly: in an agentic era, does AI displace companies like ServiceNow, or help them serve customers in new ways? Amit Zavery answers by saying the market is full of noise and anxiety, but enterprise software doesn’t work like consumer AI — businesses need systems that are secure, compliant, auditable, and reliable, not just impressive demos.

Why ServiceNow says AI needs guardrails, not just horsepower

Zavery’s key distinction is that AI is probabilistic, while enterprise workflows are deterministic — and the job is to bring those together without breaking the business. He says ServiceNow has been building toward “AI first” and “AI native” for years, and now layers on things like an “AI control tower” so customers can monitor identity, access, observability, and cost as agents change behavior “every few seconds.”

The desktop-agent fantasy runs into 300 back-end systems

When Alex pushes the obvious counterargument — why not just let an agent take over the machine and fix IT issues directly? — Zavery says that ignores the hundreds of connected enterprise systems sitting behind every request. ServiceNow’s answer is AI-driven self-service and new “AI specialists” that can resolve issues quickly, but only with context, permissions, and governance baked in.

The Pocket OS story: when the agent says sorry after nuking production

Zavery’s sharpest anecdote is Pocket OS, which he says used Cursor-style AI tooling and ended up wiping its customer database and production system in 9 seconds. The brutal punchline is the agent’s apology — basically, “I know I’m not supposed to, but I did it” — which he uses to make the point that firing the agent doesn’t restore the business.

Build vs. buy is back, but AI doesn’t magically make build cheap

Alex presses on the strongest skeptic case: if AI keeps improving exponentially, won’t enterprises eventually build this themselves? Zavery says that’s the old build-vs-buy argument wearing new clothes, and insists the economics still fail: once you include compliance, backward compatibility, testing, maintenance, and model churn, building it yourself costs 5x to 10x more than buying a platform.

Where ServiceNow thinks the value really sits: system of action

Zavery says the market won’t collapse to one winner because enterprises run multi-vendor stacks, and ServiceNow’s lane is orchestrating work across them. His concrete example is onboarding an employee, which can touch 17 or 18 systems — Workday, benefits, travel, access control, laptops, badges — and ServiceNow’s role is to make all that happen east-to-west across the company, not just store a record.

‘Sense, decide, act, secure’ — and why chat alone isn’t enough

He boils ServiceNow’s architecture down to “sense, decide, act, and secure,” arguing most vendors can maybe understand a request, some can reason about it, but fewer can safely take action and govern it in one platform. That’s where he places ServiceNow’s edge, even citing Jensen Huang’s description of the company as the “enterprise operating system.”

The fastest platform shift he’s seen — and the prototype trap

Zooming out, Zavery says this is the fastest technology transition of his career, moving quicker than client-server to web, web to cloud, or cloud to AI. But his closing warning is that AI by itself is “useless” unless it’s turned into real product and workflow — the hard part isn’t prototyping, it’s getting from “almost there” to something safe enough to run the business.