By accuracy standards, the first internal version of Veyro Agents was excellent. It updated close dates correctly 97% of the time. Within a week, our own sales team had switched it off. The 3% it got wrong were visible, embarrassing and impossible to trace — like a renewal pushed into next year because a champion said “let’s revisit in the new year” about a different project entirely. Accuracy turned out to be necessary and nowhere near sufficient. What people needed was to understand, and override, what the agent did.
Three principles behind every agent
Every action has a receipt. If an agent changes a field, it links to the exact email, call moment or billing event that justified it. No receipt, no action.
Confidence decides autonomy. Below a per-field threshold, the agent asks instead of acting. Thresholds are learned per workspace, not set globally.
Everything is reversible. Any agent action can be undone in one click for 30 days, and every undo teaches the agent something.
These read like product features, but they are really constraints on the model. An agent that can’t produce a receipt can’t act, which forces us to build systems that reason from evidence rather than pattern-match on vibes.
Receipts turned out to matter most. In our internal rollout, teams with receipts visible reverted 60% fewer actions than teams without them — not because the agent was more accurate, but because people could see it was right and stopped second-guessing it.
How the ask-first threshold works
Each agent action gets a calibrated confidence score, built from the same Signals that power Forecast. For low-risk fields like next-step dates, the default threshold to act autonomously is 0.85. For stage changes, it’s 0.93. For anything that touches a customer — an email, a meeting invite — there is no threshold at all. Agents draft. Humans send.
When a rep reverts an action, we log it as a labelled example for that workspace. After roughly 200 reverts, per-field thresholds adjust automatically. Most teams see their revert rate fall below 1% within six weeks.
Thresholds only move in one direction without human sign-off: towards more caution. If an agent’s accuracy drops on a field — say, after you redefine your stages — it falls back to ask-first automatically. Raising autonomy again requires an admin to approve it, with the evidence in front of them.
“I don’t need the agent to be perfect. I need to know why it did what it did, and to undo it in a second when it’s wrong.”
— Marcus Bell, Head of Sales Operations, Fernhill
Autonomy is earned, field by field
We don’t believe the future of revenue teams is fully autonomous. We think it’s a gradually expanding circle of work that people no longer need to touch, drawn by the people who own the outcome. Our job is to make drawing that circle safe, measurable and — ideally — boring. The best agent is the one you forget is running until you check the receipts.
That’s also why every agent’s accuracy is visible inside the product, by field, for every workspace. If an agent isn’t earning its autonomy, you should be the first to know — not the last.
