Changing modes per action from the UI isn’t available yet — it’s on the roadmap. Today the seeded defaults are what’s in force. This page explains how each mode behaves and which profile fits you, so you know what you’re opting into; if you want a grant changed before self-serve editing ships, email support@tryhoard.com.
Refresher: the five modes
Risk is computed per call against your actual pricing wizard answers. See Agent permissions and safety for the full mechanics.
”I just want it to read my data”
You want Claude or ChatGPT to answer questions about your store. Quick lookups, end-of-month summaries, “is this card priced right” gut checks. You don’t want it touching anything.
The assistant becomes a read-only analyst. It can still surface gainers, fee trends, repeat customers, anything in your data. It just can’t change a thing.
”I want suggestions but I’ll review each one”
This is the default profile. Sit here for two weeks. You’ll see what the assistant tends to suggest, you’ll click yes most of the time, you’ll catch the occasional thing that needs a second look.
Every change comes through Hoard with a confirmation link. You see the diff, you approve, it commits.
”I trust the agent on small price tweaks”
You’ve watched the audit log for two weeks. You always approve low-risk pricing edits. The confirmations have started feeling like noise.
Now a 1.10 → 1.15 multiplier bump commits silently. A 1.10 → 1.50 bump still asks. Floors still gate. Critical risk still escalates regardless of mode.
”Power user, I’ll catch problems in the audit log”
You’re running multiple stores. You’re talking to the assistant several times a day. You’ve internalized that the floors and risk scorer are the real safety net, the per-call confirmation is the optional second check.
Almost everything commits in one round trip. The assistant tells you what it just did. You review the activity log at the end of each session.
The thing this mode is good at: throughput. The thing it’s bad at: catching prompt-injection mischief immediately. The audit log catches it on review instead.
”Never let it touch my account settings”
Whatever profile you pick above, you can pinaccount.settings to deny regardless. This is worth doing.
Account settings is where your sync mode, your timezone, your shipping adder, and your pricing template live. Nothing the assistant should be reaching for casually. Lock it down:
You can always loosen it for the specific moment you need to change something, then re-lock. Or you can just keep it permanently denied and use the dashboard for those one-offs. Most sellers will never touch account settings via an assistant, so
deny costs you nothing.
The 3-week trust ladder
If you’re not sure where to start, walk the ladder. Each step is one week. Each step changes exactly one grant. If anything weird happens, hold or drop back.
After 3 weeks you’ll have a feel for it. Either you stay at
auto_low_risk for pricing forever, you push further to auto_unless_critical, or you decide you actually like the confirmations and roll back. All three are legitimate.
For the full month-by-month version, see Going from cautious to confident.
When to roll a setting back
You loosened something. Now you’re seeing weirdness. How do you know it’s time to retighten? The rough rule: two false positives in a week means roll back one step. A false positive is a commit you would not have approved if you’d been asked. One could be a fluke; two is a pattern. Rolling back is the same path as loosening: movepricing.edit_rule from auto_low_risk back to ask_for_mutations (email support until self-serve editing ships). You’re back to a confirmation prompt on every edit.
The activity log is your forensic tool here. See Reading your agent activity log for how to spot patterns.
Related reading
- Agent permissions and safety, the underlying three-layer model
- Your first agent-driven price change, what the default flow looks like
- Going from cautious to confident, the 4-week version of the ladder
- When the agent says ‘I can’t do that’, diagnostic tree for blocked actions