ServiceNow has stopped selling AI assistance and started selling autonomy. On May 5 at its Knowledge 2026 keynote, the company launched what it calls an Autonomous Workforce: AI specialists pitched to run entire business processes across IT, HR, finance, and legal, end to end. Fortune framed the announcement as a move from AI assistance to autonomous execution. Docusign is targeting 90% autonomous IT ticket resolution, the City of Raleigh reports 98% deflection on employee requests, and ServiceNow says its own internal IT specialist resolves cases 99% faster than human agents. The vendor is selling autonomy at the ceiling.
Why It Matters
Anthropic published one of the clearest large-scale empirical studies so far of how teams actually use AI agents, based on millions of human-agent interactions across Claude Code and the public API (Anthropic, “Measuring Agent Autonomy in Practice,” 2026). The pattern is sharp. New users with under 50 sessions auto-approve about 20% of the time. By 750 sessions, that rises above 40%. They also interrupt the agent more often, 9% of turns versus 5%. Oversight does not vanish, it changes from approving each step to intervening when something goes wrong. Platforms with embedded governance may compress that curve, but they do not erase it. The harness underneath earns the ceiling.
The Decision
The question is not whether to deploy autonomous agents this quarter. The question is whether the autonomy your vendor is selling matches the oversight maturity your team has actually built. A 98% deflection rate in someone else’s deployment is data about their harness, not yours.
What To Do This Week
- Ask one team head: what is your auto-approve rate on AI workflows today, and what should it be next quarter?
- List your three highest-stakes AI workflows. For each, name the human reviewer and the signal that escalates.
- Before any new agent rollout, finish this sentence: “We would know we are operating beyond our oversight if X.” If nobody can finish it, the rollout waits.
What Not To Do
Don’t buy at the autonomy level the vendor demos. Buy at the level your harness earns. Don’t measure success by how often the human is out of the loop. Measure by how often the human is in the right loop. Anthropic’s users did not jump straight into mature autonomy. They learned their way there. Yours will need its own ramp.
Signal Boost
Anthropic, “Measuring Agent Autonomy in Practice” - the empirical paper behind the 20%-to-40% auto-approve shift and the rising-interruption finding. Read it before your next agent rollout meeting, the clearest evidence that the autonomy ceiling is earned by the harness, not granted by the vendor.