How much do you actually trust an AI agent to touch your board without asking first?

Genuinely curious where this community lands on this.

We keep hearing two completely opposite takes from teams: "just give the agent access and let it work - that's the whole point," vs "the second it changes something without me seeing it first, I stop trusting it."

We're building in this space right now, and it's the single hardest product decision we keep coming back to - how much should an agent just do, versus stop and ask.

So, poll for founders/eng leads here: if an AI agent could handle your team's routine backlog grooming (triage, assignment, labeling) but paused before anything risky, would you actually turn it loose on the boring 80%? Or is it a little scary that it might drift and break something, or wipe data you actually needed?

Would love real examples - especially if you've tried something like this already and it went sideways or surprisingly well.