The hardest UX problem in autonomous AI is not making agents capable — it's making humans comfortable. Too much human-in-the-loop and you've just built an expensive assistant. Too little and you've built a liability. After 6 months of customer research, we rebuilt approval flows from scratch.
What we learned from the data
We analyzed approval patterns across 340 active Harnyss orgs. The median customer had configured approval thresholds so conservatively that their agents were generating 23 approval requests per hour during business hours. Approval response times averaged 4.2 hours — meaning most agents were blocked for longer than they were working. Effective agent utilization: 31%.
At the other extreme, 12% of customers had disabled approvals entirely. Of these, 3 had experienced incidents — an agent sending an unreviewed external email, an agent making an API call outside its expected scope. None were catastrophic, but all eroded trust.
The new model: risk-tiered batching
The new approval system classifies every agent action by its risk tier at the point of execution, using a combination of action type, target scope, and spend amount. Tier 1 (routine, internal, low-spend) auto-approves. Tier 2 (external-facing, moderate scope) batches into hourly digests. Tier 3 (high-spend, irreversible, key accounts) interrupts immediately.
In beta, customers on the new system saw agent utilization rise from an average of 34% to 79%, with zero increase in incident rate.